Citation
Development of an expert system for multichannel EEG signal analysis

Material Information

Title:
Development of an expert system for multichannel EEG signal analysis
Creator:
Chang, Tae Gyu, 1955-
Publisher:
[s.n.]
Publication Date:
Language:
English
Physical Description:
vi, 210 leaves : ill. ; 28 cm.

Subjects

Subjects / Keywords:
Amplitude ( jstor )
Electroencephalography ( jstor )
Expert systems ( jstor )
Heuristics ( jstor )
Knowledge bases ( jstor )
Rapid eye movement sleep ( jstor )
Signals ( jstor )
Sleep ( jstor )
Sleep stages ( jstor )
Waveforms ( jstor )
Dissertations, Academic -- Electrical Engineering -- UF
Electrical Engineering thesis Ph. D
Electroencephalography -- Data processing ( lcsh )
Expert systems (Computer science) ( lcsh )
Sleep -- Physiological aspects -- Data processing ( lcsh )
Genre:
bibliography ( marcgt )
non-fiction ( marcgt )

Notes

Thesis:
Thesis (Ph. D.)--University of Florida, 1987.
Bibliography:
Bibliography: leaves 200-209.
General Note:
Typescript.
General Note:
Vita.
Statement of Responsibility:
by Tae Gyu Chang.

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
The University of Florida George A. Smathers Libraries respect the intellectual property rights of others and do not claim any copyright interest in this item. This item may be protected by copyright but is made available here under a claim of fair use (17 U.S.C. §107) for non-profit research and educational purposes. Users of this work have responsibility for determining copyright status prior to reusing, publishing or reproducing this item for purposes other than what is allowed by fair use or other copyright exemptions. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder. The Smathers Libraries would like to learn more about this item and invite individuals or organizations to contact the RDS coordinator (ufdissertations@uflib.ufl.edu) with any additional information they can provide.
Resource Identifier:
021316241 ( ALEPH )
18007649 ( OCLC )

Downloads

This item has the following downloads:


Full Text











DEVELOPMENT OF AN EXPERT SYSTEM FOR MULTICHANNEL EEG SIGNAL ANALYSIS









BY

TAE GYU CHANG


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE
DEGREE OF DOCTOR OF PHILOSOPHY






UNIVERSITY OF FLORIDA


1987












ACKNOWLEDGEMENTS


The author would like to thank Dr. Jack R. Smith for all his support during the research. His responsible, kind

support, words of wisdom, and unconventionally creative research guidance all consolidated the author's rewarding and precious educational experience. The author also would

like to thank his other supervisory committee members, Dr. Donald G. Childers, Dr. Antonio A. Arroyo, Dr. S-Y Oh, and Dr. George Logethetis, for their aid, guidance, and criticism during the preparation of this dissertation.

Special thanks go to Dr. Jose C. Principe for his active participation in research guidance and discussions. The author also thanks to Miss Angela Ventura for her editorial assistance with the dissertation. The author is also thankful to fellow graduate students Fang S. Yu, Guezmir Khemais, Seung H. Park, Soon Y. Huh, Taihwan Yoon, and Chongtai Kim for their kind collaborations through years of laboratory work.

The author sincerely thanks his parents and his family for their support and love, especially his wife for her endless patience and love.

This research was supported by the National Institute

of Health Grant, No. R01-NS19996-02 and National Science Foundation Grant, No. ECE-8511857.


ii













TABLE OF CONTENTS


Page

ACKNOWLEDGEMENTS ...................................... ii

ABSTRACT .............................................. v

CHAPTERS

I INTRODUCTION ................................. 1

II WAVEFORM RECOGNITION SYSTEM .................. 13

Sleep EEG, EOG, and EMG ...................... 14
Waveform Detection Methodologies ................ 26
Analytic Approaches ....................... 28
Heuristic Analysis ........................ 32
Waveform Recognition System Design
& Implementation ............................. 37
Functional Description of the System ........ 37 Design Considerations ..................... 45
Spindles Detection ........................ 49
Rapid Eye Movement Detection ................. 59
Slow Eye Movement Detection .................. 67
K-complexes Detection ..................... 71
Delta Wave Detection and Measurement ........ 72 Muscle Artifact and EMG Analysis ............ 72

III SURVEY ON EXPERT SYSTEMS AND THEIR APPLICATIONS ................................. 74

Expert Systems Overview ...................... 74
Application Examples and Future Prospects .... 80

IV KNOWLEDGE-BASED TOKEN PROCESSING SYSTEM ........ 90

Sleep EEG Domain and Design Considerations ... 90
Sleep EEG Analyzing Domain ................ 90
Design Considerations ..................... 94
Design and Implementation .................... 107

V SYSTEM EVALUATION AND RESULTS ................ 134

Experimental Procedures - ....................... 135
Results - --------.................................. 137
Discussions .------.--............................. 151


iii








VI CONCLUSION ................................... 154


APPENDICES

A SYSTEM OPERATION SCREEN EXAMPLES ............. 161
B SYSTEM RULES ................................. 170
C MAN-MACHINE SCORING AGREEMENT TABLES FOR
EACH SUBJECT RECORD .......................... 183

REFERENCES ... ......................................... 200

BIOGRAPHICAL SKETCH ................................... 210


iv












Abstract of Dissertation Presented to the Graduate
School of the University of Florida in Partial
Fulfillment of the Requirements for the
Degree of Doctor of Philosophy


DEVELOPMENT OF AN EXPERT SYSTEM FOR
MULTICHANNEL EEG SIGNAL ANALYSIS BY

TAE GYU CHANG

May 1987

Chairman: Jack R. Smith
Major Department: Electrical Engineering


An automated computer analyzing system is designed for multichannel sleep data analysis. Sleep data are

normally analyzed by a human scorer's visual inspection of the record including perception of waveforms and segment-wise classification. A knowledge-based expert system, for data interpretation and classification, is designed on top of an early-processing system in which a heuristic signal processing approach is applied to design various waveform recognizers. This is a new approach to the

sleep data analysis, providing a different problem solving methodology from analytic signal processing techniques used in conventional approaches. This research also represents a

new application of a knowledge-based expert system to an intensive signal processing problem which requires a processing of a large amount of data with an on-line

v








monitoring feature. The whole idea of this approach is the simulation of the human expert's knowledge. The sleep data analyzing problem falls into the category of a knowledge-intensive heuristic problem domain where

well-defined algorithms or rules do not exist, but the gestalt perception and heuristic interpretation of a human expert are applied to solve the problem. Large variability of EEG characteristics and the lack of objective EEG models

add to the difficulties of analytic signal processing approaches in designing an automated computer analyzing system. The expert system technology proposes a different method for problem solving in heuristic domains such as sleep EEG analysis. It also provides a flexible and

transparent research environment allowing an easy access and

modification of the system knowledge in accordance with frequently varying requirements of the sleep data analysis and its clinical application areas. The developed system shows a man-machine agreement of average 83.6 % with a set of randomly selected 16 sample records for subjects 5 to 79 years old. The system performance is discussed with the test result. Problems and limitations for further improvements are also discussed based on the test result.


vi












CHAPTER I
INTRODUCTION



Research investigators have been working for more than 30 years on the development of automated sleep EEG analyzing computer systems [Ge75, Ba79, Is81, Ha83a, Ha86, Sm86]. Two main motivations are associated with this effort. First, an automated system would provide far more detailed and quantitative descriptions of EEG activities with better accuracy and consistency than a human scorer's visual inspection. The objectivity of the data, which could

be obtained from the automated computer analysis, may provide the standard measures to speed up development in the

fields of sleep research and clinical applications of EEG [Bi73, Is75b, Bi78, Bo8O, Cr85]. Second, an automated sleep

EEG analyzing system would be a great labor saving device. It would replace the human scorer in processing the huge amounts of EEG data involved in the research, thereby saving much of the total expense of this labor intensive activity.

However, a complete automated sleep EEG analyzing system has not yet been achieved either in terms of

satisfactory agreement with human scorers or in terms of practical usefulness of the automated system [Ba79, Is75a, Sm86]. Automated sytems developed so far are incomplete in

that they offer only partial solutions to the problem and

1








2

are impractical in that they require overly complicated implementation. This is mainly due to both the complex nature of the problem and the lack of knowledge about objective models for EEG waveforms [Mc75, Bo77, Pr77, Sa80, Is81, Nu8lb].

The main aim of this research is to develop an automated sleep EEG analyzing system which performs at a competitive level with a human scorer who visually inspects EEG records. The system design is based upon the idea of simulating the visual inspection and interpretation knowledge of the human experts. This is a new approach to

the sleep EEG analysis. A knowledge-based expert system, for the data interpretation and classification, is designed on top of the heuristic signal processing technique which is

applied to the early-processing part of the system for the recognition of waveforms. This research also represents a new application of knowledge-based expert system technology to an intensive signal processing problem.

The knowledge-based expert system technology results from the recent fundamental shift of interest in the fields of application oriented artificial intelligence [Da82, Ge83, Ha83b]. Most of the previous work has been focused on the construction of general-purpose intelligent systems. On the

other hand, current emphasis focuses on the construction of an expert system which aims at solving a domain-specific problem with specific and detailed knowledge of a human








3

expert. The expert system technology can also be regarded as the advent of a newer programming paradigm which is knowledge-based, has an inferencing role of machine intelligence, and can utilize the knowledge-base to act in a human-like intelligent manner. The intelligence implied in expert system technology specifies the idea that, the way it

solves a problem, the way knowledge is constructed and handled in the system, and the way it interacts with users all should be much the same as those of a human expert. Systems constructed in this way, which are regarded as results of different programming structure in the least sense, are higly flexible and permit a new active userinvolving environment. Users can access and modify the knowledge to any range since most of the domain knowledge constructed in this way is transparent to them. The extent

of flexibility of the active user involving environment, therefore, is distinguished from that of the limited flexibility of conventional programs where only a certain number of parameters can be interactively adjusted by the user.

The feasibility of the expert system technology depends heavily on the characteristics of a problem domain [Ha83b, Pr85]. The expert system technology is generally applied to the problem domain which involves extensive, heuristic knowledge, but which lacks well defined analytic models to solve the problem [Wa86]. The algorithm-based







4

conventional programming approach cannot provide a suitable way of solving these ill-structured problems of heuristic domains. The flexibility of system operation and the transparency to system's knowledge and its operation are crucial features of the knowledge-based expert system to meet the requirements of an active user involving environment. They allow easy access and modification of the

heuristic knowledge which is not inherently static, but changing frequently. Applications of the expert system technique include medical diagnosis [Sh79, Ku80, Bo84], equipment repair, computer configuration [Mc82], chemical data interpretation and structure identification [Ba82b], speech and image understanding [Er80a, Ni86], mineral exploration [Ba82b], military intelligence and planning [Ni86], and other decision-aids problems [Me8l]. The sleep

EEG/EOG analyzing problem also falls into the category of the problem domains where expert system technology should be

applied, since the visual interpretation of the EEG/EOG record stems mainly from the experience-based heuristic knowledge of the human scorer. Moreover, the knowledge involved in the sleep EEG/EOG analysis is not fixed but can change when new theories are proposed or more experience is acquired. On the other hand, the sleep EEG/EOG analyzing problem differs from those of other expert system domains, in the sense that it involves a large amount of data to be processed and consequently requires the system's processing








5

and reasoning efficiency to meet the requirement of an on-line monitoring capability.

Two major tasks involved in the automated sleep EEG analysis are the correct recognition of the specific waveforms in sleep EEG/EOG data and sleep stage scoring based upon the recognized information in the record. The wave activities information and the sleep stage score data can be used in related studies according to the purpose of the research or clinical application.

One of the fundamental difficulties involved in waveform recognition is that there are large inter-subject and intra-subject variabilities in sleep EEG characteristics [Wi74]. It is, therefore, difficult to design a robust detection system which works well for a wide range of subjects in the 5 to 79 year age group. Other inherent

difficulties include the fact that the straightforward and explicit definitions of the waveforms are not readily available for computer analysis. The waveforms, therefore, are only defined ostensively, and the ostensive definitions must reflect the large variabilities of the EEG characteristics. Significantly, the variability of the EEG

and the lack of good neurophysiological models for EEGs greatly limit the application of analytic signal processing techniques. The automated EEG waveform recognition system cannot rely on conventional signal processing techniques in

terms of completeness and practical usefulness of the








6

system. Thus, waveform recognition eventually comes down to the problem of simulating the human scorer's visual inspection by investigating various aspects of EEG characteristics. In this research, the time-domain

visual-simulation constitutes the basis for the signal processing methodology. The concept of this signal

processing methodology is referred to as the heuristic signal processing technique, in the sense that the EEG is heuristically analyzed applying criteria similar to those of

the human expert's visual inspection, but the analysis is not based on an analytic quantification of the EEG using conventional signal processing techniques such as spectral analysis, time-domain period-amplitude analysis, and other optimal filtering techniques, etc. [Hj70, Fe80, Ja8l, Kt8l, Cr85]. In chapter II, this heuristic approach is further discussed in comparison to other conventional techniques, and system design and implementation are described in detail.

A human scorer performs sleep stage scoring by applying a set of predefined sleep staging rules to the observed waveform information [Re68a, Ag72]. However, the sleep staging and data interpretation belong to ill-structured problems in the sense that there do not exist

solid algorithmic rules which can provide appropriate solutions to the problem. Rules, if any, are only used as marginal references when the human expert judges each








7

epoch's stage based on the gestalt observation of the EEG/EOG. The sleep staging problem should incorporate the human expert's process of gestalt observation and the heuristic rules for interpretaion. These rules are applied differently and change depending on various contexts such as

the subject's age, intra-subject EEG characteristics, and the nature of preceding and succeeding adjacent EEG/EOG epochs. The human scorer is adaptive and makes adjustments

in applying the scoring rules depending on the various contexts. The lack of well-defined algorithmic rules is the

fundamental difficulty in implementing an automated sleep staging system and is one of the significant reasons why the system is implemented in an expert system structure.

Several other important advantages result from implementing the system in an expert system structure. First, without any program modification, rules can be easily changed by the end user, who is most likely a clinician with

no programming background, using the knowledge-base editor, which is just a part of the integrated expert system. If the system is implemented in the conventional

algorithm-based structure, the end user must go into the program, edit, and recompile through the specific programming language support environment. This means that the sleep stage scoring system cannot be provided as an independent stand-alone running system, but will have to be provided as a part of a whole programming environment on a








8

specific machine. Second, the user-friendly environment and

the transparency of the expert system are very useful features required in a decision-aiding device for the clinical applications [Sh73, Ku80, Bo84]. For example, "why" and "how" explanation mechanisms and a user-friendly knowledge-base editor can allow easy modification, testing, and evaluation of the system. Third, the normative sleep staging system will be extended in the future for more advanced problems like sleep disorders or other diagnostic problems by incorporating other biological data, such as respiratory data, heart rate, etc. The expert system approach is appropriate for this kind of larger diagnostic problem [Sh79, Sz79].

Generally, there does not exist an acceptable

standardized expert system structure which works well for all problem domains [Ha83b, Wa86]. The appropriate system structure should be designed according to the specific nature of the problem domain. The scope of the system

development, therefore, includes the development of an expert system shell. The system design emphasizes reasoning

and processing efficiency which are important for the system's on-line monitoring capability, the architectural effectiveness of layered structures of data base and

knowledge base to handle the different types of knowledge associated with the different processing layers of the EEG analysis, and the effective handling of uncertainties in the








9


system.

This research is based on the previous achievements which have been obtained through the years of research in the EEG laboratory of the Department of Electrical Engineering at the University of Florida. The SAHC (Sleep Analyzing Hybrid Computer) provided the major basis for the development of a waveform recognition system in this research. The SAHC was developed by Dr. J.R. Smith through years of research with his colleagues [Sm71, Sm74, Sm75, Sm78a]. The SAHC was designed as an automated sleep EEG analyzing system for the processing of whole night, multi-channel EEG/EOG data. The SAHC could detect the

predefined waveforms such as alpha, beta, sigma, etc., and also perform the sleep stage scoring. The SAHC was a hybrid

system consisting of both an analog circuit part and a digital part. The waveform detectors were implemented mainly by analog circuits. The sleep stage scoring and data

summarization were performed by a microprocessor-based digital system. The SAHC system proved high performance level through evaluations by several sleep laboratories, especially in individual waveform detector performance levels.

However, it is necessary to implement an automated system, in a totally digital environment, to overcome certain performance limits and constraints of the hybrid system as a completely automated sleep analyzing machine,








10

and to provide it in a more compact environment. Especially, under the constraints of the hybrid machine, the

detectors of some waveforms, such as the Rapid Eye Movement (REM), the K-complexes, and the Slow Eye Movement (SEM) cannot be efficiently integrated into a single system, because the detection of those waveforms requires intensive use of memory to incorporate multi-channel EEG/EOG information and relatively long segment data information.

A REM wave detection system was developed by Lee [Le85] in a totally digital environment (T19900 microprocessor-based system). The SAHC waveform detection schemes were used in the prototype system implementation. A

computer language compiler mechanism was used to process token data. Character strings encoded with waveform information were referred to as tokens. The token data were processed by the host computer (Gould-8780) for the REM wave

detection and minute-summarized wave-activity descriptor generation. The implementation of an expert system for sleep stage scoring and other sleep diagnostic problems, incorporating the minute-summarized wave-activity descriptor information, was proposed speculatively. The compiler

generation tools, i.e., Lex [Le75] and Slrgen were utilized in the implementation of the token processing system. The Slrgen was a modified version of YACC [Jo75] by G. Logothetis [Lo84]. The amplitude and period information of all the slow-waves appearing in the EEG/EOG channels needed








11

to be sent to a host computer for REM wave detection by further processing. However, there exists a significant restriction in utilizing the computer compiler mechanism for the sleep data analysis. Most of the general parsing rules

of the computer language compiler mechanism are too much restrictive to incorporate all the widely varying potential sleep data patterns. The knowledge representation scheme of the system must be in a more flexible form to satisfactorily

reflect the visual inspection process for the diverse patterns of the sleep data.

The waveform recognition system, which is the earlyprocessing part of the overall system, is described in chapter II. The general characteristics of sleep EEG, EOG and EMG signals, the signal processing methodology, and the design and implementation of the waveform recognition system are described in detail. In chapter III, expert systems and

their application examples are reviewed with discussions on their limitaions and future prospects. Chapter IV describes

the knowledge-based token processing system, which is the second part of the overall system for the sleep EEG signal interpretation and classification. The characteristics of the sleep EEG scoring problem are described first, and then the.expert system design considerations are discussed along with the features of its domain. The detailed design and implementation of the knowledge-based token processing system is described according to the major structural








12

components of the system. The processing results of 16 all

night sleep EEG records are presented and discussed in chapter V. Conclusion is presented in chapter VI. Examples

of the expert system operation are illustrated with sample screens in appendix A. The system rules implemented in the knowledge base of the system are listed in a tabular form in appendix B. Man-machine agreement tables for each subject record are attatched in appendix C.












CHAPTER II
WAVEFORM RECOGNITION SYSTEM


This chapter describes the general characteristics of

sleep EEG/EOG/EMG, waveform recognition methodologies, and the design and implementation of the waveform recognition system. The waveform recognition system constitutes the early-processing part of the sleep EEG analyzing expert system. It performs measurement and detection of various waveforms in the EEG/EOG/EMG data. The advantages and the fundamental differences of the heuristic approach employed in the design of the waveform recognition system are discussed in comparison to other conventional analytic signal processing methodologies applied to the sleep EEG data analysis.

The waveform recognition system is designed and implemented based on a T19900 microprocomputer system. The

design considerations and the design and implementation details of the early-processing part are described along with all the waveform detectors. Recognized waveform

information is encoded in a character string and linked to a host computer for analysis, including sleep stage scoring.

Since the system is ultimately aimed at practical application, the following basic constraints must also be considered in developing the system. The system must be


13








14

designed such that it can be implemented on a small portable machine, e.g., a desktop personal computer interfaced with a microprocessor-based early-processing system. It must be

also capable of handling and providing all the EEG/EOG waveform information required by clinicians or researchers, thus resulting in a complete system.



Sleep EEG, EOG, and EMG



Sleep EEG has been used as one of the most important tools for sleep research and other related clinical applications [Ba79, Co80, Wi74]. It is generally known that

the scalp EEG originates from graded synaptic potentials generated by pyramid cells in the cerebral cortex, which are

triggered by rhythmic discharges from the thalamic nuclei. A pacemaker system situated in the thalamus and in the reticular formation probably regulates the synchrony of the cortical signals. The relatively slow time course of excitatory post-synaptic potentials (ESPSs) and inhibitory post-synaptic potentials (IPSPs) corresponds with EEG signal

and the summation is facilitated by the columnar structure of the neurons reaching from upper to lower layers of the cortex [Nu8la, Nu8lb, Wi74].

Several models have been devised to locate the signal source, to characterize the wave propagation, and to analyze the signal characteristics. Examples of these are: the








15

current dipole layer model [Nu8la] which assumes that the scalp potentials are due to current dipole layers occupying various surface areas of the cortex; a neuronal population

model [Ch73] which is based on the hypothesis that the surface potentials are a combination of ESPSs and IPSPs which occur both at different depths and different

latencies; and autoregressive or other parametric models [Fe7l, Ma76, Sa80, Ch8l, Is81] devised for signal processing purposes. However, there is not yet sufficient knowledge about the EEG signal origin and the significance of EEG activities such as the relationships between physiological and clinical states. So, in most cases the EEG signal is presented as a phenomenological model and the signal analysis or visual inspection relies on this EEG model.

The appearance of EEG signals depends on the location of the electrodes on the scalp and on the subject's state of alertness. Therefore, multichannel EEGs should be recorded according to the location dependency of the waveforms of interest. The locations of the electrodes may change and the total number of channels may be reduced according to laboratory practices, but three EEG channels, i.e., frontal (Fl-F7), central (C3-A2), and occipital (03-OzPz), one EOG channel (LE-A2), and one EMG channel according to the 10-20

electrode placement system are recommended for a complete analysis [Ja58, Re68a, Ag72]. The system is designed to have the capability for analyzing all these channels.








16

The frontal channel is recommended for inclusion in the system, since this channel shows clear beta activity much better than any of the other channels investigated. The beta wave activity will be used for a better classification of the REM sleep. The frontal channel also shows a time-coherent phasic relationship with the EOG channel better than any of the other EEG channels and this relationship will be used in REM, SEM, and k-complexes detection. The central channel shows clear sigma spindle activity better than other channels, and the amplitude level

of this channel is relatively higher than other channels. The occipital channel is used mainly for recording the alpha wave activity. This channel could be omitted in the system, since the central channel also shows fairly good alpha wave activity. At least one EOG channel is essential for REM, SEM, and k-complexes detection. The main purpose of the EMG

channel is to complement the capability of the system in separating wake and REM sleep. The EMG channel gives

information on muscle tonus level and this is used to confirm the awake stage, if the muscle tonus level is high or to confirm REM sleep, if the muscle tonus level is low. This complementary role of EMG, the REM wave detector, and the beta waveform detector together will give a very reliable REM sleep separation.

The sleep EEGs of normal subjects show changes of background EEG patterns and show appearances of different








17

waveforms depending on sleep stages. The human scorer

visually perceives the appearance of these waveforms and background patterns. He then assigns that segment of the record into one of the stages according to the perceived information. A clear definition of the waveforms in terms of measurable variables is impossible; the waveforms are not

analytically measured but directly perceived by the gestalt observation of the scorer. However, ostensive descriptions

of the waveforms in terms of various grapho-elements are obtained by an intensive observation process of many records for a wide range of subjects.

The waveforms which are used by the human scorer's visual analysis include alpha, beta, delta, sigma, theta, muscle artifact, REM, SEM, k-complexes, and EMG levels. Alpha, beta, sigma, and theta are conveniently classified as

spindles indicating that these waveforms are defined as a short burst of periodic waves. The delta is defined in terms of half-wave period and amplitude. The REM, SEM, and

k-complexes are defined most importantly in terms of wave shape, background patterns, and time-coherent information of

the multichannels, together with the amplitude and the half-wave period. Muscle artifact and EMG are high frequency signals. EMG is described as high, low, and medium according to the amplitude levels. Typical EEGs for

different sleep stages are shown in Fig. 2.1 to Fig. 2.7, and the corresponding EEG waveforms are labled on the







F1 - F7 EEG C3 A2 EEG 03 - OzPz EEG LE - A2 EOG


Muscle Artifact


Fig. 2.1 Awake Stage (Sleep Stage 0).








F1 - F7 EEG C3 - A2 EEG




03 - OzPz EEG



** *

LE - A2 EOG



* denotes Alpha waveforms.


Fig. 2.2 Awake Stage (Sleep Stage 0).









F1 - F7 EEG C3 - A2 EEG





03 - OzPz EEG EN)

LE - A2 EG


Fig. 2.3 Sleep Stage 1.













C3 - A2 EEG 03 - OzPz EEG









* denotes a Sigma spindle. K-complex


Fig. 2.4 Sleep Stage 2.








Fl - F7 EEG C3 -A2 EEG




03 OzPz EEG










K-complex


Fig. 2.5 SleeD Stage 3.



















L3 - zF EEG LE -A2 EOG


Fig. 2.6 Sleep Stage 4.








F1 - F7 EEG Beta Beta Beta
C3 - A2 EEG









LE - OzPz EMZ





REM waves


Fig. 2.7 Sleep Stage 5 (REM Sleep).








25

figures. Individual waveform definitions will be described further when the design of each waveform detector is presented.

Sleep stage zero represents the awake state, where the occipital EEG shows the dominance of the alpha activity together with occasional muscle artifacts. The EOG channel often shows the appearance of REM waves which is associated with rapid eye movement [As55, Ja7l]. Sleep stage one

mostly represents a transition period from the awake state to the sleep state. It is characterized by the disappearance of alpha waves, attenuation of EEG amplitude, the absence of distinctive sigma spindle waveforms and k-complexes, and sometimes by the slowly rolling appearance

of EOG, which is associated with eyes floating caused by drowsiness. Sleep stage two is characterized by the appearance of distinctive sigma spindle waveforms and/or the appearance of k-complexes. Sleep stages three and four represent deep sleep periods and are characterized by the dominant appearance of slow, large amplitude waves referred to as delta waves. The REM sleep period is referred to as stage five, which is associated with dreaming. The EOG wave shows the appearance of REM waves and/or rolling movements.

EEG channels show basically the same appearance as sleep stage one, but a prominent increase in beta activity on the frontal EEG channel is observed during REM periods.








26

Waveform Detection Methodologies



Various analytic signal processing techniques have been applied to the sleep EEG analysis. These are

conveniently divided into two groups, namely, frequency domain analysis and time domain analysis. In frequency

domain analysis, spectral parameters are first obtained using one of several spectrum estimation methods, and then various clustering and/or classification algorithms are applied to these spectral parameters for the

characterization of the EEG. In time domain analysis, the statistical characteristics of periodicity and amplitude distribution of the EEG are obtained using the time domain waveform descriptors, which are, in most cases, period and amplitude.

None of these analytic techniques are complete either

in terms of man-machine agreement or in terms of practical usefulness of the system for routine applications. These shortcomings are mainly due to the complex nature of EEG and

the lack of knowledge about objective models for EEG waveforms. These limitaions impose significant difficulties on the application of an analytic signal processing approach. Thus, satisfactory waveform detection cannot rely on a certain analytic signal processing technique. Moreover, the analytic quantification of the EEG in terms of

statistical variables cannot replace the human scorer's








27

heuristic quantification which is based on the visual perception of the waveforms in the record.

The heuristic signal processing approach employed in this research emphasizes the correct recognition of all the

waveforms of interest to the clinicians, instead of trying to characterize the EEG in terms of statistical variables, such as spectral parameters, periodicity, and amplitude distribution as are pursued in other analytic approaches. The recognition is performed by using various

grapho-elements and background patterns of the multichannel EEGs, EOG, and EMG, similar to the ones the human scorer also uses for the visual perception of the waveforms. The waveforms are ostensively defined in terms of these grapho-elements and background patterns for computer

analysis.

The discussions of the signal processing techniques are confined to the problem of automated analysis for routine sleep EEG applications. The discussions are mainly

focused on the analysis of the background activity and non-stationary activity (or paroxysmal event) aiming at the overall characterization of the sleep EEG. The other

problems of EEG analysis such as a spike detection, epilepsy

detection, evoked sensory response, etc., are excluded from the discussion. The advantages and disadvantages of each technique are discussed with respect to the practical usefulness and completeness of the technique for the routine








28

application of the automated system.



Analytic Approaches

Spectral analysis is one of the most popular techniques generally applied to EEG analysis. Spectral analysis is based on the underlying assumption that EEG activity may best be quantified by EEG spectral properties.

The spectral properties are described in statistical terms based on the assumption of stationary (a few seconds to a few tens of seconds) behavior of the EEG [Co77, Du73, Ja8l, La70, Ma72]. Spectral analysis has become a widely used approach along with the development of analytic signal processing techniques, especially with the advent of the FFT

and other fast computational algorithms [Co65, Co67, Ye72, We70, La80, Sm8l]. The general approach of these techniques

is to first estimate the power spectrum (or some other equivalent of it) of a certain length of the EEG segment. Each segment of the spectrum data is then further interpreted by applying various clustering and/or classification techniques [We78a, De77, Ge80]. An

appropriate power spectrum estimation is therefore the key to this approach. The spectrum estimation techniques have

evolved to show several diversities mainly in efforts to achieve better spectrum estimation with higher resolution and less computational complexity. The power spectrum is usually obtained by taking the Fourier transform of the








29

properly windowed data and squaring its absolute value. The

characteristics of the spectrum obtained using this direct method are equivalent to the ones obtained from the periodogram. Several schemes, such as averaging, windowing,

overlapping, etc., are developed on an empirical basis to improve the spectrum estimation, though there always exists a trade-off between the resolution of the spectrum

estimation and the bias and variance of the estimation [0p75]. The parametric modeling technique, which is

relatively recent in modern spectral analysis technology, gives a better spectrum estimation for a shorter data observation and also provides a mathematical model for the signal [Fe7l, To75, Ma76, Au8l, Ch8l, Is81]. The model coefficients can be used for the characterization of a segment of the record instead of using the power spectrum parameters.

Spectral analysis can provide an efficient

quantification of overall broad-band rhythms present in an EEG epoch since the rhythmic components are relatively enhanced at the corresponding frequency. On the other hand, spectral analysis is not suitable for describing short-term, transient events in the EEG since the transients are smeared in the power spectrum, which is an averaging process over a given time window of the data. To achieve a satisfactory spectral estimate for each sleep staging epoch, a relatively long observation, 30 seconds or more, is required. This








30

however, is obviously not suitable to describe the non-stationary behavior of EEG activities; i.e., short-term,

usually less than one second, existing phasic events or paroxysmal events, such as sigma spindles, K-complexes, or REM waves, cannot be described properly by the power spectrum estimate of a long EEG record.

The parametric modeling technique is used in several recent approaches to improve the resolution in the analysis and to detect transient activities. A quasi-stationary modeling technique with an adaptive segmentation scheme, where the transient behavior of EEG activity is included as a linear superposition on the segment-wise stationary model,

is introduced to solve this non-stationary behavior of the EEG [Bo77, Sa80]. In the study by A. Isaksson [Is81],

linear model parameters are allowed with time variances to include the transient behavior of EEG activity in the analysis. However, the postulate of linearity between EEG's

stationary behavior and non-stationary behavior is very doubtful, because of the observed evidence that the

background activity is wiped out by the appearance of a phasic event. It is also hard to verify the validity of the

linear model without including higher order correlation terms for the EEG.

Another disadvantage of the spectral analysis method is that the power spectrum is unsuitable for describing the

amplitude distribution property of the EEG, especially for








31

small amplitude EEG waveforms such as beta with a small S/N ratio. This is shown well in the study of A. Lubin, L.C. Johnson, and M. Austin [Lu69], where the sleep stage scoring

of human EEG is performed based on the spectral analysis approach. The study reports that the amount of beta activity is minimal in the REM period; however this result is opposite to the visual analysis; i.e., the amount of beta activity is far greater in the REM period than in the other NREM sleep periods.

Most of all, the spectral properties of the EEG obtained through the spectral analysis are not the direct EEG activity information of interest to the clinicians. So, the spectral property data should be interpreted in another domain by using various feature extraction techniques and/or pattern recognition techniques.

Time domain analysis has also been applied as a relatively simple method to quantify the rhythmic properties of the EEG. Basically, the intervals are measured between successive zero-crossing points or peaks where the first derivative of the signal is zero. The distribution of these intervals measures rhythmic activity. The relationships between the period and the amplitude provide various parameters which describe the characteristics of the EEG [Fe80]. A comparison of period-amplitude analysis, utilizing a period and an absolute amplitude histogram, with

spectral analysis shows that while the power spectrum








32

efficiently quantifies the overall power trend in the EEG data, period-amplitude analysis offers more resolution than the power spectrum in detecting details in amplitude and incidence within relatively narrow frequency bands [Kt8l]. Despite its simplicity and usefulness, this approach has been relatively less attractive to the researchers, mainly because it lacks an analytic tool for signal processing and the method is sensitive to noise and other artifacts [Is81]. The autocorrelation and cross-correlation techniques [Mi79]

have also been introduced to analyze the EEG, but with efficient computers available these techniques are now replaced by spectral analysis based on the FFT. Hjorth

parameters have been introduced to describe the EEG data quantitatively on the time domain, and these parameters are shown to be equivalent to zeroth, second and fourth moments of the normalized power spectral density [Hj70]. The three parameters used by Hjorth are as follows. The first is

called "activity" which is the variance of the amplitude or mean power in the epoch. The second is "mobility" which is the average power of the normalized derivative in the epoch.

The third is "complexity" which is the average power of the normalized second derivative in the epoch.



Heuristic Analysis

The recognition of the waveforms of direct interest to clinicians or sleep researchers is the goal emphasized in








33

the signal processing approach of this research. This,

however, differs from the other analytic signal processing approaches in which the EEGs are quantified in terms of statistical variables. The signal analysis approach is referred to here as "heuristic analysis" in the sense that the waveforms are recognized by mimicking the human scorer's

visual perception of the waveforms, i.e., by using various grapho-elements and background patterns of the EEG similar to the ones which the human scorer also uses. The signal analysis does not rely on a few limited signal parameters as is the case in period-amplitude analysis, spectral analysis, and parametric modeling approaches. Different sets of

various features, including grapho-elements, background

patterns, and temporal distributions of the multichannel record, should be utilized for recognizing each different waveform.

The sleep EEG analysis problem requires the heuristic approach for the following reasons: The present knowledge about the objective EEG waveform models is not sufficient, and the variabilities involved in the EEG are too large. Thus, a satisfactory analysis of the EEG cannot rely on one specific analytic signal processing technique. Wide

variations in the EEG characteristics are interpreted heuristically by the human scorer, and this capability of interpretation should also be reflected in the signal analysis. However, analytic signal processing techniques








34


are inherently unsuitable to reflect the heuristic factors in the human analysis, since they are based on analytic models of the signal. Because the present knowledge about

the EEG and its origin are not yet sufficient to provide objective models for EEG waveforms, it is realistic to rely on subjective models of sleep research and its clinical application areas, i.e., the methodology of EEG analysis based on the visual perception of specific waveforms and the background EEG patterns, rather than to try to provide a new analysis methodology for the clinicians. Most of the

analytic signal processing methods emphasize the objectivity

of the method and data, and suggest that these data should be utilized in the research and clinical applications instead of visually analyzed data. However, there are

significant difficulties in accepting the data provided by the analytic techniques, since reasonable validations of the

method and the data, in relation to the neurophysiological models of EEG, cannot be provided for a clinical use. Moreover, the data obtained by the analytic techniques are generally not the direct information the clinicians use. The best contribution of a computer application for EEG analysis at this stage should be in the replacement of the routine part of the labor intensive sleep analysis in a consistent and objective manner with a technique providing the same kind of information.








35

Explicit definitions of waveforms in terms of measurable variables are not readily available. The

waveforms should be defined ostensively in terms of the various grapho-elements, background patterns, and temporal distribution of the EEG for the heuristic recognition of the waveforms by a computer. Waveform definitions require a process of intensive observation of many records of a wide range of subjects. Since the waveform definitions differ widely from each other, the detailed description of the waveform definitions are given in the individual waveform detector design section and only brief summaries are shown in Table 2.1.

The waveform definition criteria include the individual full-cycle wave period window between

zero-crossings and peaks, amplitude thresholds, pattern specifications, waveform average period window, half-wave period, wave leading edge slope, background pattern screening, time-coherence of wave shapes in the multichannels, etc.. All the specifications related to the

waveform needs to be incorporated appropriately for correct waveform recognition. The weakness of sensitivity to noise or other artifacts can be avoided by carefully designing the

detection to incorporate several aspects of the waveform, although this weakness can be understood as a by-product paid for by the increase of resolution in describing waveforms on the time domain.







Table 2.1 Summary of Ostensive Waveform Definitions.


Criteria Pattern Frequency Amptitude
Window Threshold Others
Activities (Hz) (Gv)

S Alpha 6/6/3 7.5 - 12.0 7.0
P
I Beta 6/6/4 16.0 - 32.0 3.0
N
D Theta 6/6/3 2.0 - 7.5 10.0
L
E Sigma 6/6/4 12.0 - 16.0 5.0
S
REM - 0.5 - 3.0 30.0 Leading slope
background pattern screening
K-Complexes 0.5 - 2.0 70.0 Time synchrony
background pattern screening
Delta Half-Wave 0.5 ~ 2.0 16.7

SEM - 0.2 - 0.5 15.0 Leading slope
background pattern screening

Muscle Artifact 6/6/3 above 32.0 5.0

EMG 6/6/3 above 32.0 . Level descrimination


C-')







37

A simple implementation of the signal processing technique is an important aspect to be considered in the system design. The waveform recognition system should be designed to process all the waveforms on a real-time basis by incorporating several functional blocks, such as the A/D

conversion unit, the signal conditioning unit, detection unit, multi-channel information processing unit, and data-link unit, etc..



Waveform Recognition System Design and Implementation



The system, which includes multiple detectors with different detection schemes, is described in the following sequence: first, the functional description of the overall system, second, the design considerations, and third, the individual detector details.



Functional Description of the System

The waveform recognition system is implemented on a TI-9900 microprocessor-based microcomputer system, which includes a 12 bit A/D converter with two D/A channels, one 16 K-byte RAM board, two serial input-output (SIO) ports, and one parallel input-output (PIO) port.

The functional block diagram of the system is shown in Fig. 2.8. Data from three EEG channels and one EOG channel (or two EEG, one EOG, and one EMG, depending on the














TOKEN ENCODING





REM

DETECTOR I


chl ch2


BETA filter/detector SLOW filter/detector

fil ter/detector DELTA &
measurementsSIGMA filter/detector


THETA filter/detector


SLOW filter/detector ALPHA fil ter/detector REM filter/detector
( first layer SEM filter/detector
first layer

MUSCLE
ARTIFACTfi1ter/detector EMGlevel
discrimination


DETECTOR

- I


SEM


DETECTOR


first layer


second
layer


Fig. 2.8 Multilayered Parallel Processing Architecture
of Waveform Recognition System.


38


K-COMPLEX


ch3 ch4


A/ D
480 Hz


LPF
0 - 120Hz


I-


c h4


TOKEN
sendi nq through

S.I.0


host system


(jch3)-


- - -







39

montage selection) are sampled at 480 Hz per channel and passed through a digital low-pass filter with 120 Hz cutoff frequency. There are two processing layers in the waveform recognition system. The first layer includes the individual waveforms detection through the filter/detector unit. The data obtained at the first layer are used for the second layer processing. The second layer processing is performed

at a much wider interval of 0.25 second in order to detect certain waveforms, such as the REM, SEM, and K-complexes, which require the information of the other channels and/or the observation of relatively long periods of adjacent EEG/EOG background patterns. Also, the information of all the detected waveforms and the delta wave measurements are encoded in a character stream (tokens) and sent to a host computer for further analysis. The processing interval of

0.25 second for the second layer processing is selected by taking into account the following factors: the data

compression requirement such that one 5-1/4 inch floppy diskette should be sufficient to accommodate an entire night's four channel token data; the efficiency and

time-resolution requirement for the background pattern screening; and the period range of the delta wave.

Since, the single processor (TI-9900) system must be capable of processing the multi-tasks for real-time

detection of the various waveforms, an appropriate timemultiplexing scheme needs to be designed for the parallel







40

processing architecture of the waveform detection system. The time-multiplexing scheme of the system is shown in Fig. 2.9. Twelve time-slots consist of one cycle-frame where all the detectors and filters are multiplexed according to their sampling rates as is shown in the figure. Each time-slot, which is 1/480 second, contains several detectors and filters which are assigned according to the sampling rates of each filter and detector. Three of the time-slots, as marked in the figure, additionally allocate one of the second layer detectors once every 0.25 second.

The simplified main program flow is depicted in Fig. 2.10. After the initialization of all the buffers, A/D converter, interrupt control registers, and the other I/O ports, the program loops the cycle-frame for the processing. At the end of each time-slot, the interrupt handler routine is triggered by the internal clock. This routine performs the A/D conversion and checks the timer count to execute one

of the second layer detectors or token encoding routines. The A/D converted data of each channel are low-pass filtered

at 120 Hz to remove the high frequency components of the signal above 120 Hz which otherwise may cause aliasing to affect the filters and detectors of the later section. The flow chart of the interrupt handler routine is shown in Fig.

2.11.

The format of token data is shown in Fig. 2.12. Information of all the detected waveforms and delta







41


Time
slot 1 2 3 4 5 6 7 8 9 10 11 12

Processes
-pre- 0 0 0 0 0 0
Filter
ALPHA
dete- 0 0 0
ctor

BETA 0 0 0 0 0 0

pre- 0 0 0 0 0

DELTA filter
dete
ctor0

SIGMA 0 0 0 O 0 0

THETA 0 0 0

MUSCLE0000 0
ARTIFACT 0 0 0 0 0 0

EMG 0 0 0 0

SLOW(F1-F7) 0 0 0

SLOW(C3-A2) 0 0 0

SEM
SM0 0 O0
('st layer)
REM 0 0
(1'st layer)
Token
Second laye- encoprocessing SEM K-com REM ding

i ng
Fig. 2.9 Time-multiplexing Scheme of Waveform Recognition
System (one cycle-frame).








42


I__


Fig. 2.10 Major Flow Diagram for the Main Routine
of Waveform Recognition System.


Initialization Buffers I/0 ports
A/D converter Interrupt control
raiqtprs & clock


Multiplexing Cycle-frame
12 time-slots
Interrupt handler at the end of each time
slot


- - _J








43


A/D Conversion
4 ch 480 Hz


Token sending yes sednso



REM detection yes EM slot
no

K-comp detection yes K-com slo


<- SEM detection y es SEM slot ?
no

Token Encoding ye 0.25 sec ?

Sno


Interrupt Handler Routine Flow Diagram.


Fig. 2.11





TOKEN ENCODING:


- I 7j7 III '11




L - D - f REM SEM K





DF - Amplitude




- I - --------Period


Token Information









Delta Information


Fig. 2.12 Format of Tokens.








45

measurement data are encoded into a character string using a signature and bit mapping scheme. Token-encoding routine reads the flag of each waveform detector at every 0.25 second interval and uses two bytes for the encoding of all the flag information. It assigns one (high) to the

corresponding bit of the token if the flag of the waveform detector is high, or zero (low) if the flag is low. If a delta wave is detected, then additional two bytes are assigned for the amplitude and period information of the delta wave, respectively. The additional two bytes can be distinguished from the other pair of bytes by the signature bit assigned to the 3rd significant bit of the byte which contains the delta amplitude information. The sign of a delta wave is marked at the 3rd significant bit of the byte which contains the delta period information.



Design Considerations

Several important design considerations in terms of signal processing aspects, including the selection of A/D conversion unit, the determination of sampling rates for each filter and detector, and the filter characteristics, must be carefully reflected in the system design. The major

functional components of the system are the A/D conversion unit, the signal conditioning filters, the individual waveform detectors, and the token encoding and data linking unit.







46

It has been shown that a 12 bit A/D converter with 16 bit arithmetic processing is a suitable selection to obtain at least a 40 db S/N ratio [Pr86].

The A/D conversion sampling rate is one of the important design parameters. The rate must be appropriately

determined by considering the overall implementation constraints together with the characteristics of the signal. A higher sampling rate generally brings a higher measurement

resolution, but stiffens the constraints for the real-time processing of the filters and detectors. The sampling rate on the other hand must be high enough so that the aliasing,

possibly caused by the sampling, would minimize distortions in the signal of interest. The A/D conversion rate is 480 Hz in this system for each of the four channels. It is safely assumed that the aliasing contributed by the signal components above 240 Hz (1/2 * 480 Hz) is negligible. The sampling rate of 480 Hz also gives an appropriate

implementation combination with the required sampling rates of the next stage filters and detectors. The digitized data

at 480 Hz is first low-pass filtered with a -3 db cut-off frequency at 120 Hz. The digital low-pass filter is

implemented with the tranfer function H(z) = 0.5 * (11/z). The signal components under 120 Hz contain enough information for the waveform detection and analysis purposes

of this system, including the detection of the muscle artifact and the level discrimination of the EMG. The








47

muscle artifact and the EMG signal contain the highest frequency components compared to the other waveforms. The signal at the output of this low-pass filter can be further assumed to be band-limited at 120 Hz. Since the power

spectrum of the EEG itself shows 1/f characteristic, the frequency components above 120 Hz are drastically reduced and are consequently negligible at the filter output. This lowpass filtering allows a lower sampling rate

implementation of the later stage filters and detectors, reducing the overall implementation complexity.

Relatively broad-band linear phase FIR (Finite Impulse Response) filters are used for the signal conditioning filters in the system. The broad-band

frequency response characteristics are required for the signal conditioning filters, since narrow-band filters may distort the waveforms of interest. The flat pass-band and the sharp cut-off characteristics of the amplitude response are not important for the signal conditioning in this system. The broad-band characteristics and the marginal requirement in pass-band amplitude response and cut-off characteristics of the filter make it feasible to come up with a simplified filter implementation scheme [Pr86]. The filters designed in this way do not require any

multiplications or floating point arithmetic calculations, but can be implemented by shifts and additions. The

sampling frequency of an individual signal conditioning








48

filter is constrained by two factors. The first constraint

is the filter order, which increases as the sampling frequency increases when the bandwidth of the filter is fixed, thus increasing the processing time and the filter internal magnifications resulting in the possibility of an overflow. The sampling frequency therefore is limited by the highest feasible order of the FIR filter, which is reported as being around 10th order [Pr86]. The second

constraint is the lowest possible sampling frequency of the

filter limited in terms of the Nyquist rate such that the filter sampling frequency must be higher than twice the band-limit of the signal to avoid aliasing.

The period measure between zero-crossing points and/or peaks, which is performed by counting the samples between the two points, is one of the important parameters included in the detector unit. The sampling rate of the detector unit is therefore very important in terms of the measurement resolution. One sample interval is the maximum

possible error in determining the period of two points by counting the number of samples. The measurement resolution is readily represented by the following equation.



df = fo2/fs

1/fo is the period of the two points.

fs is the sampling frequency.

df is the error in frequency.








49

An example is illustrated with the case of beta. In case of

the beta, fo is approximately 25 Hz thus df is +-3 Hz with the sampling frequency fs = 240 Hz. Increase in the sampling frequency will improve the measurement resolution, but this at the same time decreases the available time-slot;

thus, the increase of the sampling frequency cannot be an appropriate solution to improve the resolution. The

measurement resolution can be increased also by the interpolation scheme. But the interpolation scheme also brings an increase in the processing time and involves an error of its own. One way, which is used in this system to improve the resolution, is to take the period as the average period of the several adjacent waves. In this way the

measurement error is reduced by the factor of the number of

the sequences included in the total period as is shown in the following equation.



df = fo2 / (N * fs)



Spindles Detection

The appearance of spindle bursts, such as the alpha spindle, beta spindle, theta spindle, and sigma spindle, is one of the well observed phenomena in human sleep EEG. These waveforms are conveniently grouped as spindles, and their detection schemes are basically similar, although the detailed nature of the waveforms is slightly different for








50


each.

The first important problem for the correct detection

of these waveforms is that these waveforms must be defined explicitly in terms of variables measurable by a computer. The waveforms are however heuristically perceived by the human scorer. Thus, straightfoward and explicit definitions

are not available. Ostensive definitions, which could be obtained by mimicking the human scorer's visual perception reflecting all the various aspects of the EEG, have to be used in the detector design.

A spindle is a short burst of waves in sequence that

form a waveform giving a distinctive appearance from the background. The spindles are roughly defined by the measure

of the periodicity of the individual waves and by the specification on the grouping (bursting) nature of the waveform. These can be further broken down into the following criteria.



o Individual wave period window: zero-crossings : peaks

o Individual wave amplitude threshold o Average period window for the whole sequence of waves in a spindle

o Pattern specification for the waveform








51

A typical functional block diagram of the spindle detector is shown in Fig. 2.13.

A spindle detector consists of a linear phase FIR filter followed by a full-cycle period discriminator, amplitude detector, positive-peak interval discriminator, and pattern recognizer.

Appropriate signal preconditioning is necessary to remove the effects of high frequency noise and of

large-amplitude, slow waves. Especially, if a spindle is superimposed on a large-amplitude, slow wave, it is

impossible to detect the spindle at the next detector unit by the zero-crossing and peak detection scheme. A

relatively broad-band linear phase filter is used for this signal pre-conditioning purpose as explained.

A sigma spindle is more subject to the effects of large-amplitude, slow waves, because a fair number of sigma spindles are superimposed with K-complexes and delta waves. On the other hand, the effect of the large-amplitude, slow waves are less serious for the other spindle waves, such as

beta, theta, and alpha, since these other spindles mostly appear on a flat EEG background while the EEG amplitude is relatively small. Thus, the lower edge cutoff

characteristics of the sigma spindle filter need to have a sharper cutoff characteristic such that the filter

successfully removes the effect of large-amplitude, slow wave of the EEG without any significant distortion of the







52


-- --> Fil teri ng


zero-crossing
&
peak detection


period window average period
zero-crossings - a
positive peaks window


H amplitude
threshold


pattern criterion


Fig. 2.13 Sigma Spindle Detector Functional Block Diagram.


_







53

spindle waveform. The filters for each spindle are summarized together with the filters for the other waveforms in Table 2.2.

The signal passed through the conditioning filter is processed at the next detection unit with various criteria.

The frequency of a wave is first defined as the inverse of the full-cycle period which is measured by counting the interval between adjacent two positive-peaks (or negative-peaks) and/or two positive-going (or negative-going) zero-crossing points. This definition of the frequency is different from that of the sinusoidal wave frequency mostly referred to as a spectral composition of signal in the engineering or scientific sense. However, the

definition given here is better justified as more closely mimicking (simulating) the human scorer's, since the human scorer measures the frequency of a waveform by counting the number of peaks in a specified time window. In the most general sense, the frequency is defined by a human as the number of events occurring per a unit period. The definition of frequency in terms of pure sinusoidal wave is one very specific case of the general definition.

The full-cycle zero-crossing points are first detected by checking the sign change of the samples. Then,

each wave's positive-peak is defined by the maximum sample value between the zero-crossing points. The positive-peak is described by the saved maximum positive amplitude and the





Filters Table.


SAMPLING
WAVES FREQUENCY PASS BAND FILTER TRANSFER FUNCTION
( Hz )



ALPHA 240 0 - 18 H1(z) = (z-2 + 1) 2(z-2 + z~1 + 1)(z~ + 1)
120 7 - 16 H2(z) = (z-5 - 1)(z-2 + 0.5z~ + 1)(z1 + 1)

BETA 240 9 - 40 H(z) = (z - 1)(z-2 + 1.5z~ + 1)(z-2 + 1)

DELTA 80 no filter H(z) = (zI + 1)

SIGMA240 12 - 40 H(z) = (Z-5_ - )z2 +.5~+1(z+1)-2SIGMA 1)(Z2 + 0.5z + 1)(z + 1)(z


THETA 120 0 - 9 H(z) = (z-2 + 1)2 (Z-2 + z-1 + 1)(Z-1 + 1)


ARTIFACT 240 60 - 120 H(z) = (z- - 1)

EMG 240 60 - 120 H(z) = (z - 1)

SLOWS 120 0 - 9 H(z) = (z-2 + 1)2(z-2 + z-1 + 1)(z~1 + 1)

REM 120 no filter H(z) = (z1 + 1)

SEM 120 no filter H(z) = (z1 + 1)

* filters are used only for the removal of high end frequency components in the signal.


-


Table 2.2








55

relative time displacement of the peak from the leading zero-crossing point. The peak has been continuously updated

whenever a new sample value is greater than the previous saved peak within the period. If the peak of the full-cycle wave is greater than a certain amplitude threshold, then the

periodicity of the individual wave is tested by using corresponding period windows for both intervals between positive-peaks and full-cycle zero-crossing points.

The reliability of the detection generally can be improved by including more measurements in the specification. For instance, the full-cycle zero-crossing

period is not sufficient for the spindle wave periodicity measure because the reliability of the detection could be affected by the period measurement error caused by large-amplitude, slow EEG waves, or other noise effects. On the other hand, a spindle's positive-peak interval

distribution gives a very good measure to mimic the human scorer's visual perception of the spindles. The

positive-peak interval is less subject to the large-amplitude, slow waves of the EEG. A positive-peak interval is a measure of the period between the previous wave's peak and the present wave's peak point; the interval

is obtained by adding the two intervals, one between the previous wave's peak and the previous wave's ending

zero-crossing point, and the other one between present wave's leading zero-crossing point and wave positive-peak.








56

Thus, if this positive-peak wave interval is out of specified range for a spindle, the previous wave is discarded even though it meets the zero-crossing period and amplitude criteria.

However, the way this system detects and utilizes the

information about peaks is distinguished from the generally understood peak detection technique. In general, the peak detection technique looks for the points where the first derivative value of the wave is zero, and applies appropriate period and amplitude criteria on these peaks to analyze waveform characteristics. This approach is taken in the sigma spindle study done by J.C. Principe and J.R. Smith [Pr82]. However, this peak detection technique is more subject to the local variabilities of waves than the zero-crossing technique. On the other hand, in this system,

the global wave peak is detected by taking the maximum sample value between zero-crossing points; thus the positive-peak is less subject to local variabilities of the signal.

The positive-peak interval criterion is used together

with the zero-crossing period specification for a reliable spindle detection, since the positive-peak interval alone can not impose any restriction on the wave's vertical variation in terms of the vertical asymmetry from the baseline. Also, the two-period-window scheme gives more flexibility in specifying the spindle wave periodicity. The








57

scheme allows a looser specification for each period window allowing more variations in terms of each period specification. But, on the other hand, the scheme reflects a tighter specification in terms of the global spindle shape specification by using the two period windows.

The human scorer is not very sensitive to the

individual wave period and amplitude variances in detecting the spindles. On the other hand, he relies more on the well-shaped global appearance of the whole waveform. The pattern criterion and the average period window are used to

specify the global spindle waveform grouping (bursting) nature. The average period window is applied to the total period of several adjacent waves sequence in the waveform. The average period window increases the measurement accuracy

by a factor equal to the number of waves averaged as is discussed in the previous section. A tighter average period

window is applied for the specification of the spindle waveform, on the other hand, the period windows for

individual waves can become looser to reflect individual wave variances in the spindle. The three numbers in the pattern criteria specify the total number of waves kept in a

window, the minimum required number of consecutive in-band waves in the waveform at onset, and the minimum number of the in-band waves to sustain the detection, respectively. The summarized specifications for the spindle waveforms are shown in Table 2.3. The flow chart of the spindle detector







58


Table 2.3. Summarized Specifications for Spindle Waveforms.


Criteria wave period window(Hz) average amplitude pattern
period threshold criteria
Spindles zero-cross positive window Spidespeaks (Hz) (uv)

Alpha 7.0 - 13.3 7.5 - 12.0 8.0 - 12.0 7.0 6/6/3


Beta 15.0- 34.0 15.0- 30.0 16.0 - 30.0 2.0 6/6/3



Sigma 11.0- 17.1 11.0- 17.1 11.43- 16.0 3.0 6/6/3


Theta 2.5 - 7.0 - 2.5 - 5.0 10.0 6/4/3








59

is shown in Fig. 2.14. This flow chart is applied for all

the spindle detectors with appropriate changes in the parameter values according to the criteria of each spindle detector.



Rapid Eye Movement Detection

The REM wave is detected based on ostensive definitions which include the rapidly rising leading edge, quiescent nature of EEG activity during REM sleep, and the REM wave amplitude and period window. The rapidly rising REM wave leading edge is used as a distinguishing feature from other EEG waves such as delta waves, and SEM waves occurring in the EOG channel. It is also necessary to

observe the nature of EEG channels' activity during REM sleep to eliminate false REM detections caused by the EEG activities appearing on the EOG channel. In particular, the appearance on the EOG channel of large-amplitude EEGs, such as k-complexes or delta waves in NREM sleep, may cause false

REM detections. This false positive detection has been a problem in accurate REM wave detection using automated computer analysis methods [Le84]. The EEG during REM sleep is typically represented by the appearance of

small-amplitude and high-frequency waves, no delta-like, large-amplitude, slow waves, no rolling movements, thus giving a quiet and flat appearance from the EEG baseline.









decrease
display counter

no zero-crossin

ime yes
overflow ? yes shift
I wave window
no set
overflow p>Ai?
flag

increase
timej




get new peak


verflo yes reset
flag ? overflow flan
no
yes
check ? no W t W ?
eriod '
yes no

increase Wcnt decrease Wcnt reset Wcnt



flag set ? no nt = W ma no
cnt -cmt

yes yes

Int> W mi ? no
yes
_______ n0A cn1


ec
vg. period ? 0
yes

set flag set display count~ r set disp-flag


yes



reset diso-flag reset fla


return

Fig. 2.14 Sigma Spindle Detector Flow Chart.


60


no


output disp-flag clear timers clear Peaks








61

The main idea behind the REM wave detection algorithm

of this system is to get a robust REM wave detection by using both the descriptive REM wave criteria and the EEG channels' background screening with the appropriately chosen time windows. The descriptive REM wave criteria include a slope threshold (350 uv/sec) on the leading edge, a period window (0.2 - 2.0 sec/half-wave), and a peak amplitude threshold (30 uv). Two quiescent testing time windows are applied, one each, to the central (C3-A2) and to the frontal (Fl-F7) channel EEGs as is shown in Fig. 2.15. The central channel time window is applied to the preceding and

succeeding two seconds from the terminating edge zerocrossing point of the REM wave. The frontal channel time window is applied to the preceding and succeeding one second

from the terminating edge zero-crossing point of the REM wave. The REM wave definition is therefore the following.



o Period: 0.2 - 2.0 sec/half-wave o Amplitude: 30 uv

o Slope: 350 uv/sec

o Quiescent Test

Frontal channel: +-l sec window, peak

amplitude ratio > 0.5 (between the biggest waves

in the window)

Central channel: +-2 sec window, no delta wave







62


F1-F7


1 sec 1 sec C3-A2


2 sec 2 sec


03-OzPz




LE-A2


Fig. 2.15 Time Windows for EEGs Background Pattern Screening.








63

A two-layered screening scheme implements the realtime REM detection algorithm. Each REM wave is detected at

the bottom layer through the filter/detector unit with the descriptive waveform criteria, and the EEG channels' background screening is performed at the upper layer with the 0.25 second sampling interval. The functional block diagram of the REM detection is shown in Fig. 2.16.

The leading edge slope is defined by the line connecting the leading zero-crossing point and the wave peak point of the signal as is shown in Fig. 2.17. In terms of the global description of the REM wave leading edge, this measurement is found to be a better representation of the edge slope than the measurement where the edge slope is defined as the locally averaged maximum slope value. In terms of accuracy in measuring the edge slope, this global slope measurement is more accurate because the effect of quantization error is greater for the shorter duration measurement. Also in terms of the robustness to the erroneous local variabilities of waves, the global slope measurement is better than locally averaged slope measurement.

To eliminate erroneous effects caused by the EOG baseline movements in measuring the leading slope, a "dead-zoning" is first applied to the filtered output signal with an appropriate selection of the dead-zone level. One third of the minimum REM wave peak threshold, i.e. 10 uv, is








64


filter period & amplitude
& threshold slope threshold
dead-zoning

first layer




Frontal quiescen Central quiescent
- 1 sec 2 sec
peaks comparison no delta

second layer


Fig. 2.16 REM Detector Functional Block Diagram.







65













dead-zone = 2 * 10 (uv)






(b) b


b
slope =----a


Illustrations of Dead-zoning and Leading Slope Measure REM Wave Detection. (a) EOG (LE-A2); (b) Output from the Filter/Dead-zoning Unit.


Fig. 2.17








66

selected as the dead-zone level, and is found sufficient for

the removal of the erroneous effects caused by the EOG baseline movements. This level is not too big for the purpose of an appropriate edge slope representation. To prevent the effect of filter smoothing in measuring the slope, a relatively wide lowpass filter (-3 db at 30 Hz) is used only for the removal of high-end frequency components in the signal [Kt75].

For the second layer, both the central channel (C3-A2) and the frontal channel (Fl-F7) are used to test the

EEG's quiescent nature during REM sleep resulting in a higher reliability than testing one EEG channel. A fixed amplitude threshold is used for the C3-A2 channel's quiescent test criterion. On the other hand, the relative

peak amplitude ratio between the EEG peak and the EOG peak is used for the Fl-F7 channel's quiescent test. It is

observed that the Fl-F7 channel often displays the appearance of delta-like waves, which actually is caused by eye movement activity. These delta-like waves may give an error in testing the Fl-F7 channel's quiescent

characteristics during REM sleep, although the magnitude level of the delta-like wave at the EEG channel is smaller than the REM waves.

It is observed that the frontal(Fl-F7) EEG channel shows more REM-related cross channel information than the central (C3-A2) EEG channel. The dominant appearance of








67

beta on the frontal channel during REM sleep is far more noticeable than it is on the C3-A2 channel. Beta is one of the important waveforms which can be utilized, together with the REM wave detection, for the purpose of accurate demarcation of REM sleep periods. It is also observed that

the opposite and synchronized phasic waveform relationship of k-complexes between EOG and EEG channels is more clearly shown on the frontal channel than on the other EEG channels.

The Fig. 2.18 shows the flow chart for the first layer processing of the REM detection. The Fig. 2.19 and 2.20 show the flow charts for the second layer processing of the REM detection.



Slow Eye Movement Detection

The slow eye movements (SEM) cause the EOG channel's slow, rolling fluctuations from the baseline. The SEM is physiologically related with the eyes floating, especially during the transition period from wake to sleep caused by drawsiness. The SEM detection scheme includes a half-wave period window, an amplitude threshold, the central EEG channel's background screening, and a wave leading slope threshold. The flow chart for the SEM detection is

basically similar to that of REM detector except for the absence of the peak comparison criteria with the other EEG channels.






68
1EOG
f ilter & dead-zoning
0 - 30 Hz, 10 uv
decrease delay counter

ay coun no clear REM1 flag

yes


no zer-rossing

imer yes yes
over f Dt o F


set overflwfg


EE no mp Amin

yes
ver flow yes
increase timer flag ?
0 clear overflow
no flag
get new peak period ?
Yes

slope 0 350 uv/
yes delay counter yes
>0 ?
0 eak comparis o "

clear REMI (I's t 1scyes
layer) flag pdt R4 ea clear REMI flag]
set REMI flag Jelay counter 0,25 sec

clear timer clear peak



E return : Fig. 2.18 REM Detector Flow Chart (first layer).








69


decrease counter! REM peak counted EEG peak counted



flagI = 0 ?

output Low0

ow-window
0 ?

ow-window yes
=0?

yes lag2 = 0 ? EM-window(8) no
=1?
yes
no EM-window(1)
- es set flag2

EM-window(8) s/ - I ? clear yes
REM-flag
output Low set REM-flag
clear REM-flag output High
\/ \J/


no clear REM-flag
set flag EM-window( clear REM-window(
0 ? output Low
yes


clear flagI clear flag2


to Part.B.


REM Detector Flow Chart (second layer - part A).


Fig. 2.19







70

from Part.A. shift left REM-window()



REM1-flag yes
=0 ?
:I~no y

-peak > no EG-peak(pr . sec) ' yes

REM-window(l)<-i



shift left s low-window()


yes slow-flag(fro al EEG) =0 ?
no

s/ EG-peak '>
Peak( 1sec)

delta
< I no slow-window(l)
(S --I

paks(prev. 1 sec' u pdate. (EEG & REM peaks)l


return


Fig. 2.20 REM Detector Flow Chart (second layer -part B).








71


K-complexes Detection

The main usage of the k-complexes detector in the present system has twofolds: to gain better confidence in scoring sleep stage two by complementing the spindle detector information, and to accurately demarcate the sleep stages two and three. The K-complexes detector uses an algorithm structure similar to that of the REM wave

detector, since the detection of the K-complexes also requires the cross-channel information such as the

synchronization of peaks, the opposite phasic relationship between EEGs and EOG, and a quiescent criterion.

Presently, synchronization of EEGs and EOG is

detected with a time window, i.e., 0.75 second, across the EEG and EOG channels (specifically Fl-F7, C3-A2, and EOG). The amplitude threshold (50 uv) and the relatively loose period threshold (0.3 - 1.5 sec/half-wave) are used as the k-complexes criteria. The current definition of K-complexes in the algorithm requires no delta-like slow and

large-amplitude wave within the interval of preceeding and succeeding two seconds from the K-complexes, since the present K-complexes detection algorithm is less reliable in sleep stages three and four. Therefore, the K-complexes information in these sleep stages is not utilized in the sleep scoring process.

The more descriptive K-complexes detection criteria, such as the opposite phasic relationship between the cross







72

EEG channels and the EOG channel, upward going characteristic, and slope criterion should be utilized for a better K-complexes detection.



Delta Wave Detection and Measurement

The delta detector is implemented with a period window (0.25 - 1.0 sec/half-wave) and a relatively low amplitude threshold of 16.7 uv. The amplitude and period values of each delta wave are sent to the host computer through a serial port for a post delta processing. The post

delta processing is necessary for an accurate delta summary because the variations of the EEG amplitude level provide poor delta wave detections if the same amplitude threshold is used for all subjects. Individual delta wave amplitude and period are necessary also for the purpose of more quantitative delta studies.

The delta amplitude is quantized into 16 levels from 16.7 uv to 100 uv with the 5.2 uv resolution and the period

is also quantized into 16 levels from 0.25 to 1.0 second with the 47 msec resolution.



Muscle Artifact and EMG Analysis

Muscle artifact is associated with the movement of body. The signal consists of high frequency components usually above 30 Hz. The detector consists of a highpass filter, a zero-crossing detector, an amplitude threshold (10








73

uv), a full-cycle period discriminator (34.3 - 120 Hz), an average period window, a 60 Hz notch filter, and a pattern specification (6/6/3). The flow chart of the muscle artifact detector is same as that of the spindle detector as is shown in Fig. 2.14, except for the addition of 60 Hz notch filter. The 60 Hz notch filter is operated in conjunction with the average period window. The purpose of

this filter is to reduce the effect of 60 Hz environmental noise in detecting muscle artifact.

The EMG signal amplitude is described into three levels, i.e., below 10 uv, between 10 uv and 20 uv, and above 20 uv. The EMG discriminator structure is same as the

muscle artifact detector but three different amplitude thresholds are applied for the EMG level discrimination.













CHAPTER III
SURVEY ON EXPERT SYSTEMS AND THEIR APPLICATIONS


In this chapter, general aspects of expert systems are briefly reviewed with explanations on their historical

*background, distinguishing features from algorithm-based programs, and architectural aspects. Several expert system

application examples are then briefly reviewed with an emphasis on a critical view of the expert system approach, discussing their limitations and future prospects in relation to their application domain characteristics.



Expert Systems Overview



Knowledge-based expert systems have been drawing considerable interest in the application-oriented artificial intelligence (AI) field since the mid-1960s. This area of AI has concentrated on the construction of a high-performance system which executes tasks at the level of

a human expert in a specialized professional domain. Historically, the knowledge-based expert system is a fundamental shift in the application-oriented AI area from the earlier effort to find general methods for

problem-solving and use them to create general purpose programs. This strategy to achieve a general problem 74








75

solver, despite some interesting progress, has produced no significant success because developing general-purpose programs has been too difficult and ultimately fruitless. The main emphasis has shifted to the knowledge imbedded in a problem domain. Researchers have pointed out that the problem-solving power of a program comes from the knowledge

it possesses, not just from the formalism and inference schemes it employs. This conceptual shift, thus, leads to the development of special-purpose computer programs, systems that are expert in some narrow problem area. These

programs are called expert systems. How to extract,

construct, and manipulate the domain knowledge constitutes the major concern in developing an expert system. A new set of principles, techniques, and development tools has emerged

and that forms the basis for knowledge engineering [Ba82b, Da82, Ha83b, Wa86].

In comparison with algorithm-based computer programs

of the conventional approach, the knowledge-based expert system is best distinguished by the characteristics of the problem domain it is applied to. It is not suitable to apply the algorithm-based conventional programming approach to expert system's problem domains. In these problem

domains, tractable algorithmic solutions usually do not exist, since many important tasks are related to the

intensive heuristic knowledge of a human expert, thus resisting precise description and rigorous analysis.








76

Planning, legal reasoning, medical diagnostics, geological exploration, and analysis of military situations exemplify these problems. Contemporary methods of symbolic and mathematical reasoning, which have limited applicability to the area of expert systems, do not provide a means for

representing knowledge, describing problems at multiple levels of abstraction, allocating problem-solving resources,

controlling cooperative processes, and integrating diverse sources of knowledge in inference. These functions depend primarily on a capacity to manipulate problem descriptions and to selectively apply relevant pieces of knowledge. The expert systems, depending on the application problem domain,

are categorized into several types such as interpretation, prediction, diagnosis, design, planning, monitoring,

debugging, repair, instruction, and control systems [Ha83b, Ge83, Wa86].

Expert systems are also distinguished by different program structures from the algorithm-based programs of conventional approach. Expert systems typically consist of

four major structural components, i.e., a knowledge base, a data base, a control mechanism, and a knowledge-base editor. In the expert system structure, emphasis is on the

separation of domain knowledge into a separate knowledge base in the form of modularized pieces of independent knowledge. A control mechanism provides a general strategy

in trying to match the relevant findings and data in the








77

data base to the knowledge base in order to derive a conclusion. Since, most of the domain knowledge is represented by a set of modularized pieces and is

constructed in a separate knowledge base, the domain knowledge can be handled in a flexible and manageable way by a knowledge-base editor. The control mechanism can also be

in a flexible and generalized form in the knowledge-based expert system structure. The knowledge-base editor, which is a user interface for the manipulation of the overall system, is also a distinguishing feature of an expert system, providing a flexible and transparent environment for the user. Without any modification of programs, a user can access and change the system knowledge through the knowledge-base editor. The expert system often posesses the capability to explain the line of reasoning for a conclusion through the knowledge-base editor.

Expert system architectures vary widely depending on their applications. The current techniques and principles of expert systems are based mostly on the relatively small number of early expert systems. Rule-based knowledge

representation constitute the major framework of the current expert system technologies [Ha85a]. Although the rule-based

knowledge system is appealing because of its general expressive power and well-defined syntaxes, more

sophisticated reasoning schemes and architectures are required for many intrinsically harder problem domains.








78

These domains usually include the problems of a large search space of possible solutions, diversity of knowledge, dimensionality of data or knowledge (e.g. time, space), lack of fixed inference process to solve the problem effectively,

handling of interactions among decisions of separate subproblems of the task, need for specialized representations, and incompleteness of data. A frame-based representation [Fi85] and a blackboard model approach [Ha85b] are among the several advanced techniques to cope with these difficulties. A frame provides a structured representation of an object or a class of objects. Each class of objects can be described as a specialization (subclass) of other more generic classes, thus the classes are represented into taxonomies in organizing frames. The frame provides a concise representation of useful relations, and supports a concise definition-by-specialization

technique that is easy for most domain experts to use. In addition, this representation structure provides automatic inferences as part of each assertion and retrieval operation. The taxonomic relationships among frames enable descriptive information to be shared among multiple frames via inheritance. The internal structure of frames enables an automatic maintenance of semantic integrity constraints.

Production rules by themselves are inadequate to define terms and to describe domain objects and static relationships among objects. Hybrid representations are








79

constructed by integrating frame and production rule representations, resulting in the consolidation of advantages of both representation techniques. The frames provide a rich structural representation for describing the objects referred to in the rules. The frames support a layer of generic deductive capability about those objects that do not need to be explicitly dealt with in the rules. Frame taxonomy can also be used to partition, index, and organize a system's production rules.

The blackboard model of problem solving is a highly structured special case of opportunistic problem solving [Ha79]. In addition to opportunistic reasoning as a knowledge-application strategy, the blackboard model

prescribes the organization of the domain knowledge, all inputs, and intermediate and partial solutions needed to solve the problem. The solution space is organized into one or more application-dependent hierarchies. The domain knowledge is partitioned into independent modules of knowledge that transform information on one level of the hierarchy into information on the same or other levels. The knowledge modules perform the transformation using

algorithmic procedures or heuristic rules that generate actual or hypothetical transformations. Reasoning is applied within this overall organization of the solution space and task-specific knowledge. In other words, the module of knowledge to apply is determined dynamically, one








80

step at a time, resulting in the incremental generation of partial solutions. At each step of the knowledge

application, either forward- or backward-reasoning methods can be applied.



Application Examples And Future Prospects



Application Examples

One of the applications of expert system technology is in the building of medical consulting systems designed as aids to medical decision making [Sz79, Ku8O, Bo84]. Mycin [Sh76], Casnet [We78b], and Internist [Po75 Po77] are among the examples of early expert systems mostly often referred to as typical expert system applications. The Mycin system

was developed to provide consultative advice on diagnosis and therapy for infectious diseases. The system is in a

form of a production rule based system, and the exactness and incompleteness of data and knowledge are handled by the

Certainty Factor scheme. Casnet is a computer system for medical diagnosis in the treatement of glaucoma. The system represents a disease not as a static state but as a dynamic process modeled by a network of causally linked pathophysiological states. The system diagnoses a patient

by determining the pattern of pathophysiological causal pathways present in the patient and identifying this pattern with a disease category. Internist is a consulting program








81

in the domain of internal medicine. Using the information presented during the consultation, the program tries to discriminate between competing disease hypotheses. Thus, this system is of the type that verifies a hypothesis formation using the system's knowledge represented in the form of a disease tree, or disease taxonomy.

The application of expert systems in medicine was motivated by two major reasons. First, the expert

computer-based system can provide great benefits by providing thorough and reliable diagnostic services.

Considering that most of the errors made by the clinicians are caused by omissions involved in the diagnostic process, the computer can provide reliable diagnostics by an

exhaustive consideration of all the possibilities and all the relevant patient's data. The second motivation is

related to current research interests in application oriented artificial intelligence. Clinical medicine has been a fertile area for the study of cognitive processes, and diagnosis as a cognitive process has been studied extensively.

Historically, statistical analysis and pattern

recognition, through a discriminant function based on the Bayesian decision theory, have been used for the development of computer analysis of medical diagnostic problems. The appeal of the statistical method is that the decisions based on such methods are optimal for given criteria. However,








82

the statistical approach is unsuitable in medical problems, because of various assumptions and simplifications such as independence and mutual exclusiveness of various disease states. These assumptions cannot be suitably validated and

the a priori and conditional probabilities required in the analysis are usually not available.

The medical decision aid is a typical example of a heuristic domain which includes a large amount of domain specific medical knowledge and physicians' heuristic knowledge obtained through years of practice and experience with many special cases. The physicians provide a diagnosis

based on the heuristic knowledge for handling incomplete data and uncertain information.

One of the difficulties in designing a medical consulting expert system with a human expert-like

performance is that the human expertise comes from his capability to consider and handle a large number of special cases. As the spectrum of a task becomes larger, more and more of special considerations are involved in solving the problem. Moreover, the heuristic knowledge of a physician also involves a kind of commonsense knowledge obtained through a long period of experience and education, and this

commonsense knowledge is immense in its nature and is usually very difficult to articulate and represent in a manageable form in a computer.








83

The development of computer-based consultation

systems also brings many formidable social, physical, and ethical problems that must be considered in expert systems' design. These problems include validating the systems, exporting them to hospitals or laboratories, getting physicians and patients to accept them, and determining the level of responsibility for the clinical decisions made with the help of these systems. It is reported that the biases

of medical personnel against computers are so strong that systems will eventually be rejected, regardless of performance [St72]. Thus, there exists another barrier in

achieving a significant contribution of computer diagnoses consulting systems for routine applications.

Several expert system applications have been shown also in the domain of chemistry and other engineering and scientific application areas. Dendral, Prospector, and Hearsay systems are among the well-known expert system examples in those application areas.

Dendral [Bu69, Ba82b] generates plausible structural representations of organic molecules with data obtained from mass-spectrographic analysis of unknown molecules and with a set of rules used by an expert chemist to infer constraints on molecular structures from such data. The elucidation of

molecular structures is fundamental to the application of chemical knowledge to important problems in biology and medicine. In many circumstances, the powerful analytic








84

techniques of X-ray crystallography and X-ray fine structure

analysis may not be applicable; consequently the analysis process is much too laborious and dependent on the intuitive expertise of chemists. Thus, there exists a legitimate reason for developing a computer aided expert system, and room for a significant contribution of the computer

analysis.

In this type of molecular structure elucidation or chemical compound synthesis domain, the major problem aspect is a heuristic search in limiting the combinatorial explosion of all the possible structural candidates.

However, the process of making up a set of rules about mass-spectrometry proves to be much too involved, since the theory of mass spectrometry is incomplete and the rules about it are inexact and difficult for experts to explicate.

The Dendral system's level of performance is generally far less than that of a human expert. This

indicates the difficulty involved in codifying the human expert's skilled process in the system. However, there exists a possible contribution of computer analysis in the structure elucidation domain, because a computer can provide a systematic search through the space of possible molecular

structures, a systematic use of what it does know to constrain the possible structures, and the calculating power

to handle fair amount of calculations involved in the process.








- 85

Prospector is a computer-based consultation system for mineral exploration [Du79, Ba82b]. The main function of Prospector is to match data from surface geological

observations against models of five different types of deposits. In the Prospector system, contextual information

and data are treated in the same plane of a probability propagation model, and the control scheme is heavily dependent upon the numerical comparisons. Thus, the system lacks flexibility for considering high level context information and situational considerations in an appropriate way. The model, which is based on the probability propagation, cannot provide a sufficient validation.

HEARSAY speech understanding system [Er80a, Er80b] was developed in the domain of voice chess. This system was

constructed with the idea of independent knowledge sources cooperatively solving a problem by posting hypotheses on a global blackboard data structure. Isolating the knowledge sources along functional lines provides efficient

modification of the problem-solving structure of the program, by allowing a free substitution of independent molecular knowledge sources. This molecular structure, where the knowledge sources do not address each other directly, allows great flexibilty as the system evolves and different combinations of knowledge sources and control strategies are tried. This blackboard structure has also been incorporated into several other systems solving diverse








86

tasks in crystallography, signal interpretation, vision, and psychological modeling.



Limitations and Future Prospects

Expert knowledge consists of two abstractively described parts. First, it consists of the symbolic descriptions that characterize the definitional,

taxonomical, and empirical relationships in a domain and secondly, the procedures for manipulating these descriptions. To achieve a high-level of performance, a human expert's skills also need to be well understood and included in the system's knowledge base. A skilled process

is usually related to fast response, efficiency, reduced error, reduced cognitive load, and increased adaptability and robustness. The term skill conveys the idea of appropriate knowledge and its effective use.

Most of the current expert systems are based on a limited number of knowledge-representaion and inferencing techniques. However, there exist several fundamental limitations in realizing human-like intelligence in a computer by the present A.I. approaches which are mostly based on the description of features and rule-like inferences. The essential problems related to the immense commonsense knowledge and to the lack of knowledge for human being's recognition processes, such as image-based

inference, similarity recognition, and relevance pruning







87

processes are crucial barriers to overcome the limitation of the current application-oriented A.I. technology in

achieving a machine intelligence competetive with a human expert's level of performance.

Based on this assertion, A.I and, in particular the expert system approaches are severly criticized in the book by H.L. Dreyfus and S.E. Dreyfus [Dr86]. It states that it

is impossible to achieve a computer system which can show human expert level performance incorporating human expertlike processes. This opinion is appealing considering that

the fundamental questions raised regarding the human cognitive processes are not answered and most of the current techniques are based on superficial simulations. More

importantly, expert level performance is related to the human expert's capability for considering large numbers of special cases including commonsense knowledge obtained through many years of experience and education. If the crucial key to obtaining the human expert-like performance is in the codification of this huge amount of heuristic knowledge, there exists a fundamental barrier to this approach. The heuristic knowledge is inherently difficult to codify into the system, and the amount of knowledge easily exceeds the manageable range of a computer. They also point out that most of the optimistic reports regarding

the early expert system examples are misleading, in the sense that, the status of the systems are in reality far








88

from routine applications. A system's high performance, if any, is not due to machine intelligence and A.I., but to the

computer's other superior aspects over human beings in handling data-intensive or calculation-intensive portions of the total task. This is exemplified in the Internist and the Dendral systems. It is also noticeable that no other systems ever attracted more attention than those early application examples mentioned above, implying a difficulty in expert system approach for routine applications.

However, there is unfortunately very little that can be offered as a constructive counter-proposal, for solving heuristic problems by a computer with a human-like level of performance and intelligence, that is not somewhat vague and

hand-waving. It is more realistic to accept the A.I. and expert system approach as a new way to solve different aspects of a problem in which conventional approaches cannot offer a suitable solution. Selection of an appropriate domain, then, will be a significant issue in successful applications of the expert system approach. Considering the

present status of A.I. and expert system technology, it is more reasonable to select a domain such that the expert system approach can give a complementary contribution to solving the heuristic portion of the whole task. The whole task still involves applications of conventional engineering or computer technologies and/or requires computer's superior calculating power to handle a fair amount of data-intensive








89

and calculation-intensive portions of the tasks. The

problem domain also needs to be selected by taking into consideration the amount of knowledge to be incorporated into the system such that it does not exceed the manageable

range of the computer. It is also better to avoid the problems if a validation from the domain professionals is a

significant issue, thus it is difficult to be accepted by the domain professionals because of the social, ethical, cultural, and other complicated problems, as shown in the medical diagnostic expert systems.












CHAPTER IV
KNOWLEDGE-BASED TOKEN PROCESSING SYSTEM


Sleep EEG Analyzing Domain And
Its Design Considerations


Sleep EEG Analyzing Domain

Human sleep EEG/EOG/EMG data are analyzed by a human

scorer through the visual perception and interpretation of the multi-channel data on a polygraph chart. This process is best interpreted as primarily a pattern matching and classification based on heuristically obtained knowledge in

the form of templates for the waveforms and sleep stages. The sleep data analysis, thus, falls into an application domain category of data interpretation by classification.

The expert classifies a segment of the record into one of the five sleep stages plus awake by interpreting the

data according to the perceived information, within the epoch, and other contextual information. In the process of visual scoring, the human expert perceives the occurences of

waveform activities, such as alpha, beta, delta, sigma, theta, muscle artifact, rapid eye movement, etc., by a gestalt observation of the multi-channel sleep EEG/EOG/EMG data on a polygraph chart.

Most of the patterns and templates do not readily exist in a fixed and definable form, since each segment of


90








91

the record must be interpreted in a different context according to the individual subject's signal

characteristics, several adjacent epoch patterns, and other higher order contextual situations which may affect the sleep stage scoring. Templates have to be articulated and extracted from the equivalent ones that are in the form of the expert's heuristic knowledge which comprises all the contextual data interpretations. The expert's knowledge is obtained by an intensive training process and is gestaltively applied to the visual inspection of the record.

In 1968, Rechtschaffen and Kales created a set of sleep stage scoring criteria that provide a standard reference for sleep stage scoring among researchers [Re68a]. A brief summary of these criteria is listed below.

Stage Wake (Stage 0): This stage corresponds to the waking state. It is characterized by alpha activity and/or

a low voltage, mixed frequency EEG. Certain subjects may have a virtually continuous alpha activity; other subjects may show little or no alpha activitiy in the record. This stage is usually accompanied by a relatively high tonic EMG,

and often REMs and eye blinks are present in the EOG tracing.

Stage 1: This stage is defined by a relatively low voltage, mixed frequency EEG with a noticeable theta activity. Stage 1 occurs most often in the transition from

wakefulness to the other sleep stages or following body








92

movements during sleep in normal subjects. During nocturnal

sleep, Stage 1 tends to be relatively short, ranging from 1 to 7 min. Scoring of Stage 1 requires an absence of clearly defined K-complexes and sigma spindles. Stage 1, especially

following wakefulness, is characterized by the presence of slow eye movements, each of several seconds duration, which are usually most prominent during the early portions of the stage. Rapid eye movements are absent. Tonic EMG levels are usually below those of relaxed wakefulness. The

transition from a high-alpha subject wake state to Stage 1 is characterized by a decrease in the amount of alpha activity. Fifty percent (50 %) of the epoch is used as the boundary for separation of Stage 0 and Stage 1.

Stage 2: This stage is defined by the presence of sigma spindles and/or K-complexes and an absence of sufficient delta activity to define the presence of Stage 3 and 4. If less than 3 minutes of the record which would ordinarily meet the requirements for Stage 1 intervene between two epochs of Stage 2, these intervening epochs are to be scored Stage 2, if there is no indication of movement

arousal or pronounced increase in muscle tone during the interval in question.

Stage 3: Stage 3 is defined by an EEG record in which

at least 20 % but not more than 50 % of the epoch consists of delta activity.








93

Stage 4: Stage 4 is defined by an EEG record in which

more than 50 % of the epoch consists of delta activity. Intervals between delta waves rarely persist for more than a few seconds in Stage 4, but are usually prominent in Stage 3 epochs. Sigma spindles may or may not be present in Stage

4.

Stage REM: Stage REM is defined by the concomitant appearance of relatively low voltage, mixed frequency EEG activity and episodic REMs. The EEG pattern resembles the one described for Stage 1, except the vertex sharp waves are not prominent in Stage REM. Alpha activity is usually somewhat more prominent in Stage REM than in Stage 1. As with the EEG of Stage 1, there is an absolute absence of sigma spindles and K-complexes. Stage REM should not be scored in the presence of a relatively elevated mental-submental EMG.

These criteria define the patterns of each sleep stage in terms of the number of occurrences of specific waveforms, total time of specific waveform activity in an epoch, and/or the amplitude level description of the record [Ag72, Re68b]. For some waveforms such as sigma spindle, K-complex, and REM, the number of occurrences is used in the criteria. For the other waveforms the total running time of

waveform activity within an epoch determines the criteria. The Rechtschaffen and Kales criteria, however, are a minimum set of quantitative specifications which can be used only as








94


a general guide for sleep stage scoring and training of human scorers. The actual sleep stage scoring is performed

by incorporating much more heuristic knowledge for the contextual interpretation. The actual EEG data processing is based on a gestalt perception of waveform activities together with a heuristic interpretation of the record and is not based on a precise and analytic measurement of the data.



Design Considerations

Several design considerations for developing an expert system for the automated sleep EEG/EOG/EMG signal analysis are discussed in the following. For sleep stage scoring, the on-line data processing aspect of an automated system is a distinguishing factor from other contemporary expert systems. Other related special design considerations

like reasoning and processing efficiency, architectural effectiveness, reasoning with uncertainty, and adaptiveness of the system are discussed.

To date, expert system design is limited to a few knowledge representations and inferencing methodologies which are exemplified through the earlier applications in the medical domain, chemical structure elucidation domain, and other engineering or scientific application domains [Ba82a, Ba82b, Co8l], which are reviewed in Chapter III. If-then production rules provide a very efficient way of




Full Text

PAGE 1

DEVELOPMENT OF AN EXPERT SYSTEM FOR MULTICHANNEL EEG SIGNAL ANALYSIS BY TAE GYU CHANG A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 1987

PAGE 2

ACKNOWL EDGEMENT S The author would like to thank Dr. Jack R. Smith for all his support during the research. His responsible, kind support, words of wisdom, and unconventionally creative research guidance all consolidated the author's rewarding and precious educational experience. The author also would like to thank his other supervisory committee members. Dr. Donald G. Childers, Dr. Antonio A. Arroyo, Dr. S-Y Oh, and Dr. George Logethetis, for their aid, guidance, and criticism during the preparation of this dissertation. Special thanks go to Dr. Jose C. Principe for his active participation in research guidance and discussions. The author also thanks to Miss Angela Ventura for her editorial assistance with the dissertation. The author is also thankful to fellow graduate students Fang S. Yu, Guezmir Khemais, Seung H. Park, Soon Y. Huh, Taihwan Yoon, and Chongtai Kim for their kind collaborations through years of laboratory work. The author sincerely thanks his parents and his family for their support and love, especially his wife for her endless patience and love. This research was supported by the National Institute of Health Grant, No. R01-NS19996-02 and National Science Foundation Grant, No. ECE-8511857.

PAGE 3

TABLE OF CONTENTS Page ACKNOWLEDGEMENTS ii ABSTRACT V CHAPTERS I INTRODUCTION 1 II WAVEFORM RECOGNITION SYSTEM 13 Sleep EEG, EOG, and EMG 14 Waveform Detection Methodologies 26 Analytic Approaches 28 Heuristic Analysis 32 Waveform Recognition System Design & Implementation 37 Functional Description of the System 37 Design Considerations 45 Spindles Detection 49 Rapid Eye Movement Detection 59 Slow Eye Movement Detection 67 K-complexes Detection 71 Delta Wave Detection and Measurement 72 Muscle Artifact and EMG Analysis 72 III SURVEY ON EXPERT SYSTEMS AND THEIR APPLICATIONS 74 Expert Systems Overview 74 Application Examples and Future Prospects .... 80 IV KNOWLEDGE-BASED TOKEN PROCESSING SYSTEM 90 Sleep EEG Domain and Design Considerations ... 90 Sleep EEG Analyzing Domain 90 Design Considerations 94 Design and Implementation 107 V SYSTEM EVALUATION AND RESULTS 134 Experimental Procedures 135 Results 237 Discussions 152 iii

PAGE 4

VI CONCLUSION 154 APPENDICES A SYSTEM OPERATION SCREEN EXAMPLES 161 B SYSTEM RULES 170 C MAN-MACHINE SCORING AGREEMENT TABLES FOR EACH SUBJECT RECORD 183 REFERENCES 200 BIOGRAPHICAL SKETCH 210 iv

PAGE 5

Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy DEVELOPMENT OF AN EXPERT SYSTEM FOR MULTICHANNEL EEC SIGNAL ANALYSIS BY TAE GYU CHANG May 1987 Chairman: Jack R. Smith Major Department: Electrical Engineering An automated computer analyzing system is designed for multichannel sleep data analysis. Sleep data are normally analyzed by a human scorer's visual inspection of the record including perception of waveforms and segment-wise classification. A knowledge-based expert system, for data interpretation and classification, is designed on top of an early-processing system in which a heuristic signal processing approach is applied to design various waveform recognizers. This is a new approach to the sleep data analysis, providing a different problem solving methodology from analytic signal processing techniques used in conventional approaches. This research also represents a new application of a knowledge-based expert system to an intensive signal processing problem which requires a processing of a large amount of data with an on-line V

PAGE 6

monitoring feature. The whole idea of this approach is the simulation of the human expert's knowledge. The sleep data analyzing problem falls into the category of a knowledge-intensive heuristic problem domain where well-defined algorithms or rules do not exist, but the gestalt perception and heuristic interpretation of a human expert are applied to solve the problem. Large variability of EEG characteristics and the lack of objective EEG models add to the difficulties of analytic signal processing approaches in designing an automated computer analyzing system. The expert system technology proposes a different method for problem solving in heuristic domains such as sleep EEG analysis. It also provides a flexible and transparent research environment allowing an easy access and modification of the system knowledge in accordance with frequently varying requirements of the sleep data analysis and its clinical application areas. The developed system shows a man-machine agreement of average 83.6 % with a set of randomly selected 16 sample records for subjects 5 to 79 years old. The system performance is discussed with the test result. Problems and limitations for further improvements are also discussed based on the test result. vi

PAGE 7

CHAPTER I INTRODUCTION Research investigators have been working for more than 30 years on the development of automated sleep EEC analyzing computer systems [Ge75, Ba79, Is81, Ha83a, Ha8 6, Sm86]. Two main motivations are associated with this effort. First, an automated system would provide far more detailed and quantitative descriptions of EEC activities with better accuracy and consistency than a human scorer's visual inspection. The objectivity of the data, which could be obtained from the automated computer analysis, may provide the standard measures to speed up development in the fields of sleep research and clinical applications of EEC [Bi73, Is75b, Bi78, B08O, Cr85] . Second, an automated sleep EEC analyzing system would be a great labor saving device. It would replace the human scorer in processing the huge amounts of EEC data involved in the research, thereby saving much of the total expense of this labor intensive activity. However, a complete automated sleep EEC analyzing system has not yet been achieved either in terms of satisfactory agreement with human scorers or in terms of practical usefulness of the automated system [Ba79, Is75a, Sm86]. Automated sytems developed so far are incomplete in that they offer only partial solutions to the problem and

PAGE 8

2 are impractical in that they require overly complicated implementation. This is mainly due to both the complex nature of the problem and the lack of knowledge about objective models for EEG waveforms [Mc75, Bo77, Pr77, Sa80, Is81, Nu81b] . The main aim of this research is to develop an automated sleep EEG analyzing system which performs at a competitive level with a human scorer who visually inspects EEG records. The system design is based upon the idea of simulating the visual inspection and interpretation knowledge of the human experts. This is a new approach to the sleep EEG analysis. A knowledge-based expert system, for the data interpretation and classification, is designed on top of the heuristic signal processing technique which is applied to the early-processing part of the system for the recognition of waveforms. This research also represents a new application of knowledge-based expert system technology to an intensive signal processing problem. The knowledge-based expert system technology results from the recent fundamental shift of interest in the fields of application oriented artificial intelligence [Da82, Ge83, Ha83b] . Most of the previous work has been focused on the construction of general-purpose intelligent systems. On the other hand, current emphasis focuses on the construction of an expert system which aims at solving a domain-specific problem with specific and detailed knowledge of a human

PAGE 9

3 expert. The expert system technology can also be regarded as the advent of a newer programming paradigm which is knowledge-based, has an inferencing role of machine intelligence, and can utilize the knowledge-base to act in a human-like intelligent manner. The intelligence implied in expert system technology specifies the idea that, the way it solves a problem, the way knowledge is constructed and handled in the system, and the way it interacts with users all should be much the same as those of a human expert. Systems constructed in this way, which are regarded as results of different programming structure in the least sense, are higly flexible and permit a new active userinvolving environmentUsers can access and modify the knowledge to any range since most of the domain knowledge constructed in this way is transparent to them. The extent of flexibility of the active user involving environment, therefore, is distinguished from that of the limited flexibility of conventional programs where only a certain number of parameters can be interactively adjusted by the user. The feasibility of the expert system technology depends heavily on the characteristics of a problem domain [Ha83b, Pr85]. The expert system technology is generally applied to the problem domain which involves extensive, heuristic knowledge, but which lacks well defined analytic models to solve the problem [Wa86]. The algorithm-based

PAGE 10

4 conventional programming approach cannot provide a suitable way of solving these ill-structured problems of heuristic domains. The flexibility of system operation and the transparency to system's knowledge and its operation are crucial features of the knowledge-based expert system to meet the requirements of an active user involving environment. They allow easy access and modification of the heuristic knowledge which is not inherently static, but changing frequently. Applications of the expert system technique include medical diagnosis [Sh79, Ku80, Bo84] , equipment repair, computer configuration [Mc82], chemical data interpretation and structure identification [Ba82b], speech and image understanding [Er80a, Ni86], mineral exploration [Ba82b] , military intelligence and planning [Ni86], and other decision-aids problems [Me81] , The sleep EEG/EOG analyzing problem also falls into the category of the problem domains where expert system technology should be applied, since the visual interpretation of the EEG/EOG record stems mainly from the experience-based heuristic knowledge of the human scorer. Moreover, the knowledge involved in the sleep EEG/EOG analysis is not fixed but can change when new theories are proposed or more experience is acquired. On the other hand, the sleep EEG/EOG analyzing problem differs from those of other expert system domains, in the sense that it involves a large amount of data to be processed and consequently requires the system's processing

PAGE 11

and reasoning efficiency to meet the requirement of an on-line monitoring capability. Two major tasks involved in the automated sleep EEG analysis are the correct recognition of the specific waveforms in sleep EEG/EOG data and sleep stage scoring based upon the recognized information in the record. The wave activities information and the sleep stage score data can be used in related studies according to the purpose of the research or clinical application. One of the fundamental difficulties involved in waveform recognition is that there are large inter-subject and intra-subject variabilities in sleep EEG characteristics [Wi74] . It is, therefore, difficult to design a robust detection system which works well for a wide range of subjects in the 5 to 79 year age group. Other inherent difficulties include the fact that the straightforward and explicit definitions of the waveforms are not readily available for computer analysis. The waveforms, therefore, are only defined ostensively, and the ostensive definitions must reflect the large variabilities of the EEG characteristics. Significantly, the variability of the EEG and the lack of good neurophysiological models for EEGs greatly limit the application of analytic signal processing techniques. The automated EEG waveform recognition system cannot rely on conventional signal processing techniques in terms of completeness and practical usefulness of the

PAGE 12

6 system. Thus, waveform recognition eventually comes down to the problem of simulating the human scorer's visual inspection by investigating various aspects of EEG characteristics. In this research, the time-domain visual-simulation constitutes the basis for the signal processing methodology. The concept of this signal processing methodology is referred to as the heuristic signal processing technique, in the sense that the EEG is heuristically analyzed applying criteria similar to those of the human expert's visual inspection, but the analysis is not based on an analytic quantification of the EEG using conventional signal processing techniques such as spectral analysis, time-domain period-amplitude analysis, and other optimal filtering techniques, etc. [Hj70, Fe80, Ja81, Kt81, Cr85]. In chapter II, this heuristic approach is further discussed in comparison to other conventional techniques, and system design and implementation are described in detail . A human scorer performs sleep stage scoring by applying a set of predefined sleep staging rules to the observed waveform information [Re68a, Ag72] . However, the sleep staging and data interpretation belong to ill-structured problems in the sense that there do not exist solid algorithmic rules which can provide appropriate solutions to the problem. Rules, if any, are only used as marginal references when the human expert judges each

PAGE 13

7 epoch's stage based on the gestalt observation of the EEG/EOG. The sleep staging problem should incorporate the human expert's process of gestalt observation and the heuristic rules for interpret aion. These rules are applied differently and change depending on various contexts such as the subject's age, intra-subject EEG characteristics, and the nature of preceding and succeeding adjacent EEG/EOG epochs. The human scorer is adaptive and makes adjustments in applying the scoring rules depending on the various contexts. The lack of well-defined algorithmic rules is the fundamental difficulty in implementing an automated sleep staging system and is one of the significant reasons why the system is implemented in an expert system structure. Several other important advantages result from implementing the system in an expert system structure. First, without any program modification, rules can be easily changed by the end user, who is most likely a clinician with no programming background, using the knowledge-base editor, which is just a part of the integrated expert system. If the system is implemented in the conventional algorithm-based structure, the end user must go into the program, edit, and recompile through the specific programming language support environment. This means that the sleep stage scoring system cannot be provided as an independent stand-alone running system, but will have to be provided as a part of a whole programming environment on a

PAGE 14

8 specific machine. Second, the user-friendly environment and the transparency of the expert system are very useful features required in a decision-aiding device for the clinical applications [Sh73, Ku80, Bo84] . For example, "why" and "how" explanation mechanisms and a user-friendly knowledge-base editor can allow easy modification, testing, and evaluation of the system. Third, the normative sleep staging system will be extended in the future for more advanced problems like sleep disorders or other diagnostic problems by incorporating other biological data, such as respiratory data, heart rate, etc. The expert system approach is appropriate for this kind of larger diagnostic problem [Sh79, Sz79] . Generally, there does not exist an acceptable standardized expert system structure which works well for all problem domains [Ha83b, Wa86] . The appropriate system structure should be designed according to the specific nature of the problem domain. The scope of the system development, therefore, includes the development of an expert system shell. The system design emphasizes reasoning and processing efficiency which are important for the system's on-line monitoring capability, the architectural effectiveness of layered structures of data base and knowledge base to handle the different types of knowledge associated with the different processing layers of the EEG analysis, and the effective handling of uncertainties in the

PAGE 15

system. This research is based on the previous achievements which have been obtained through the years of research in the EEG laboratory of the Department of Electrical Engineering at the University of Florida. The SAHC (Sleep Analyzing Hybrid Computer) provided the major basis for the development of a waveform recognition system in this research. The SAHC was developed by Dr. J.R. Smith through years of research with his colleagues [Sm71, Sm74, Sm75, Sm78a] . The SAHC was designed as an automated sleep EEG analyzing system for the processing of whole night, multi-channel EEG/EOG data. The SAHC could detect the predefined waveforms such as alpha, beta, sigma, etc., and also perform the sleep stage scoring. The SAHC was a hybrid system consisting of both an analog circuit part and a digital part. The waveform detectors were implemented mainly by analog circuits. The sleep stage scoring and data summarization were performed by a microprocessor-based digital system. The SAHC system proved high performance level through evaluations by several sleep laboratories, especially in individual waveform detector performance levels. However, it is necessary to implement an automated system, in a totally digital environment, to overcome certain performance limits and constraints of the hybrid system as a completely automated sleep analyzing machine.

PAGE 16

10 and to provide it in a more compact environment. Especially, under the constraints of the hybrid machine, the detectors of some waveforms, such as the Rapid Eye Movement (REM), the K-complexes, and the Slow Eye Movement (SEM) cannot be efficiently integrated into a single system, because the detection of those waveforms requires intensive use of memory to incorporate multi-channel EEG/EOG information and relatively long segment data information. A REM wave detection system was developed by Lee [Le85] in a totally digital environment (TI9900 microprocessor-based system). The SAHC waveform detection schemes were used in the prototype system implementation. A computer language compiler mechanism was used to process token data. Character strings encoded with waveform information were referred to as tokens. The token data were processed by the host computer (Gould-8780) for the REM wave detection and minute-summarized wave-activity descriptor generation. The implementation of an expert system for sleep stage scoring and other sleep diagnostic problems, incorporating the minute-summarized wave-activity descriptor information, was proposed speculatively. The compiler generation tools, i.e.. Lex [Le75] and Slrgen were utilized in the implementation of the token processing system. The Slrgen was a modified version of YACC [Jo75] by G. Logothetis [Lo84] . The amplitude and period information of all the slow-waves appearing in the EEG/EOG channels needed

PAGE 17

to be sent to a host computer for REM wave detection by further processing. However, there exists a significant restriction in utilizing the computer compiler mechanism for the sleep data analysis. Most of the general parsing rules of the computer language compiler mechanism are too much restrictive to incorporate all the widely varying potential sleep data patterns. The knowledge representation scheme of the system must be in a more flexible form to satisfactorily reflect the visual inspection process for the diverse patterns of the sleep data. The waveform recognition system, which is the earlyprocessing part of the overall system, is described in chapter II. The general characteristics of sleep EEG, EOG and EMG signals, the signal processing methodology, and the design and implementation of the waveform recognition system are described in detail. In chapter III, expert systems and their application examples are reviewed with discussions on their limitaions and future prospects. Chapter IV describes the knowledge-based token processing system, which is the second part of the overall system for the sleep EEG signal interpretation and classification. The characteristics of the sleep EEG scoring problem are described first, and then the. expert system design considerations are discussed along with the features of its domain. The detailed design and implementation of the knowledge-based token processing system is described according to the major structural

PAGE 18

12 components of the system. The processing results of 16 all night sleep EEG records are presented and discussed in chapter V. Conclusion is presented in chapter VI. Examples of the expert system operation are illustrated with sample screens in appendix A. The system rules implemented in the knowledge base of the system are listed in a tabular form in appendix B. Man-machine agreement tables for each subject record are attatched in appendix C.

PAGE 19

CHAPTER II WAVEFORM RECOGNITION SYSTEM This chapter describes the general characteristics of sleep EEG/EOG/EMG, waveform recognition methodologies, and the design and implementation of the waveform recognition system. The waveform recognition system constitutes the early-processing part of the sleep EEG analyzing expert system. It performs measurement and detection of various waveforms in the EEG/EOG/EMG data. The advantages and the fundamental differences of the heuristic approach employed in the design of the waveform recognition system are discussed in comparison to other conventional analytic signal processing methodologies applied to the sleep EEG data analysis. The waveform recognition system is designed and implemented based on a TI9900 microprocomputer system. The design considerations and the design and implementation details of the early-processing part are described along with all the waveform detectors. Recognized waveform information is encoded in a character string and linked to a host computer for analysis, including sleep stage scoring. Since the system is ultimately aimed at practical application, the following basic constraints must also be considered in developing the system. The system must be 13

PAGE 20

14 designed such that it can be implemented on a small portable machine, e.g., a desktop personal computer interfaced with a microprocessor-based early-processing system. It must be also capable of handling and providing all the EEG/EOG waveform information required by clinicians or researchers, thus resulting in a complete system. Sleep EEG, EOG, and EMG Sleep EEG has been used as one of the most important tools for sleep research and other related clinical applications [Ba79, Co80, Wi74] . It is generally known that the scalp EEG originates from graded synaptic potentials generated by pyramid cells in the cerebral cortex, which are triggered by rhythmic discharges from the thalamic nuclei. A pacemaker system situated in the thalamus and in the reticular formation probably regulates the synchrony of the cortical signals. The relatively slow time course of excitatory post-synaptic potentials (ESPSs) and inhibitory post-synaptic potentials (IPSPs) corresponds with EEG signal and the summation is facilitated by the columnar structure of the neurons reaching from upper to lower layers of the cortex [Nu81a, Nu81b, Wi74] . Several models have been devised to locate the signal source, to characterize the wave propagation, and to analyze the signal characteristics. Examples of these are: the

PAGE 21

current dipole layer model [Nu81a] which assumes that the scalp potentials are due to current dipole layers occupying various surface areas of the cortex; a neuronal population model [Ch73] which is based on the hypothesis that the surface potentials are a combination of ESPSs and IPSPs which occur both at different depths and different latencies; and autoregressive or other parametric models [Fe71, Ma76, Sa80, Ch81, IsSl] devised for signal processing purposes. However, there is not yet sufficient knowledge about the EEG signal origin and the significance of EEG activities such as the relationships between physiological and clinical states. So, in most cases the EEG signal is presented as a phenomenological model and the signal analysis or visual inspection relies on this EEG model. The appearance of EEG signals depends on the location of the electrodes on the scalp and on the subject's state of alertness. Therefore, multichannel EEGs should be recorded according to the location dependency of the waveforms of interest. The locations of the electrodes may change and the total number of channels may be reduced according to laboratory practices, but three EEG channels, i.e., frontal (F1-F7), central (C3-A2), and occipital (03-OzPz), one EOG channel (LE-A2), and one EMG channel according to the 10-20 electrode placement system are recommended for a complete analysis [Ja58, Re68a, Ag72] . The system is designed to have the capability for analyzing all these channels.

PAGE 22

The frontal channel is recommended for inclusion in the system, since this channel shows clear beta activity much better than any of the other channels investigated. The beta wave activity will be used for a better classification of the REM sleep. The frontal channel also shows a time-coherent phasic relationship with the EOG channel better than any of the other EEG channels and this relationship will be used in REM, SEM, and k-complexes detection. The central channel shows clear sigma spindle activity better than other channels, and the amplitude level of this channel is relatively higher than other channels. The occipital channel is used mainly for recording the alpha wave activity. This channel could be omitted in the system, since the central channel also shows fairly good alpha wave activity. At least one EOG channel is essential for REM, SEM, and k-complexes detection. The main purpose of the EMG channel is to complement the capability of the system in separating wake and REM sleep. The EMG channel gives information on muscle tonus level and this is used to confirm the awake stage, if the muscle tonus level is high or to confirm REM sleep, if the muscle tonus level is low. This complementary role of EMG, the REM wave detector, and the beta waveform detector together will give a very reliable REM sleep separation. The sleep EEGs of normal subjects show changes of background EEG patterns and show appearances of different

PAGE 23

17 waveforms depending on sleep stages. The human scorer visually perceives the appearance of these waveforms and background patterns. He then assigns that segment of the record into one of the stages according to the perceived information. A clear definition of the waveforms in terms of measurable variables is impossible; the waveforms are not analytically measured but directly perceived by the gestalt observation of the scorer. However, ostensive descriptions of the waveforms in terms of various grapho-elements are obtained by an intensive observation process of many records for a wide range of subjects. The waveforms which are used by the human scorer's visual analysis include alpha, beta, delta, sigma, theta, muscle artifact, REM, SEM, k-complexes, and EMG levels. Alpha, beta, sigma, and theta are conveniently classified as spindles indicating that these waveforms are defined as a short burst of periodic waves. The delta is defined in terms of half -wave period and amplitude. The REM, SEM, and k-complexes are defined most importantly in terms of wave shape, background patterns, and time-coherent information of the multichannels, together with the amplitude and the half-wave period. Muscle artifact and EMG are high frequency signals. EMG is described as high, low, and medium according to the amplitude levels. Typical EEGs for different sleep stages are shown in Fig. 2.1 to Fig. 2.7, and the corresponding EEG waveforms are labled on the

PAGE 24

1

PAGE 28

22

PAGE 29

23
PAGE 31

25 figures. Individual waveform definitions will be described further when the design of each waveform detector is presented. Sleep stage zero represents the awake state, where the occipital EEG shows the dominance of the alpha activity together with occasional muscle artifacts. The EOG channel often shows the appearance of REM waves which is associated with rapid eye movement [As55, Ja71]. Sleep stage one mostly represents a transition period from the awake state to the sleep state. It is characterized by the disappearance of alpha waves, attenuation of EEG amplitude, the absence of distinctive sigma spindle waveforms and k-complexes, and sometimes by the slowly rolling appearance of EOG, which is associated with eyes floating caused by drowsiness. Sleep stage two is characterized by the appearance of distinctive sigma spindle waveforms and/or the appearance of k-complexes. Sleep stages three and four represent deep sleep periods and are characterized by the dominant appearance of slow, large amplitude waves referred to as delta waves. The REM sleep period is referred to as stage five, which is associated with dreaming. The EOG wave shows the appearance of REM waves and/or rolling movements. EEG channels show basically the same appearance as sleep stage one, but a prominent increase in beta activity on the frontal EEG channel is observed during REM periods.

PAGE 32

26 Waveform Detection Methodologies Various analytic signal processing techniques have been applied to the sleep EEG analysis. These are conveniently divided into two groups, namely, frequency domain analysis and time domain analysis. In frequency domain analysis, spectral parameters are first obtained using one of several spectrum estimation methods, and then various clustering and/or classification algorithms are applied to these spectral parameters for the characterization of the EEG. In time domain analysis, the statistical characteristics of periodicity and amplitude distribution of the EEG are obtained using the time domain waveform descriptors, which are, in most cases, period and amplitude. None of these analytic techniques are complete either in terms of man-machine agreement or in terms of practical usefulness of the system for routine applications. These shortcomings are mainly due to the complex nature of EEG and the lack of knowledge about objective models for EEG waveforms. These limitaions impose significant difficulties on the application of an analytic signal processing approach. Thus, satisfactory waveform detection cannot rely on a certain analytic signal processing technique. Moreover, the analytic quantification of the EEG in terms of statistical variables cannot replace the human scorer's

PAGE 33

27 heuristic quantification which is based on the visual perception of the waveforms in the record. The heuristic signal processing approach employed in this research emphasizes the correct recognition of all the waveforms of interest to the clinicians, instead of trying to characterize the EEG in terms of statistical variables, such as spectral parameters, periodicity, and amplitude distribution as are pursued in other analytic approaches. The recognition is performed by using various grapho-elements and background patterns of the multichannel EEGs, EOG, and EMG, similar to the ones the human scorer also uses for the visual perception of the waveforms. The waveforms are ostensively defined in terms of these grapho-elements and background patterns for computer analysis . The discussions of the signal processing techniques are confined to the problem of automated analysis for routine sleep EEG applications. The discussions are mainly focused on the analysis of the background activity and non-stationary activity (or paroxysmal event) aiming at the overall characterization of the sleep EEG. The other problems of EEG analysis such as a spike detection, epilepsy detection, evoked sensory response, etc., are excluded from the discussion. The advantages and disadvantages of each technique are discussed with respect to the practical usefulness and completeness of the technique for the routine

PAGE 34

28 application of the automated system. Analytic Approaches Spectral analysis is one of the most popular techniques generally applied to EEG analysis. Spectral analysis is based on the underlying assumption that EEG activity may best be quantified by EEG spectral properties. The spectral properties are described in statistical terms based on the assumption of stationary (a few seconds to a few tens of seconds) behavior of the EEG [Co77, Du73, JaSl, La70, Ma72]. Spectral analysis has become a widely used approach along with the development of analytic signal processing techniques, especially with the advent of the FFT and other fast computational algorithms [Co65, Co67, Ye72, We70, LaSO, Sm81] . The general approach of these techniques is to first estimate the power spectrum (or some other equivalent of it) of a certain length of the EEG segment. Each segment of the spectrum data is then further interpreted by applying various clustering and/or classification techniques [We78a, De77, GeSO]. An appropriate power spectrum estimation is therefore the key to this approach. The spectrum estimation techniques have evolved to show several diversities mainly in efforts to achieve better spectrum estimation with higher resolution and less computational complexity. The power spectrum is usually obtained by taking the Fourier transform of the

PAGE 35

29 properly windowed data and squaring its absolute value. The characteristics of the spectrum obtained using this direct method are equivalent to the ones obtained from the periodogram. Several schemes, such as averaging, windowing, overlapping, etc., are developed on an empirical basis to improve the spectrum estimation, though there always exists a trade-off between the resolution of the spectrum estimation and the bias and variance of the estimation [Op75] . The parametric modeling technique, which is relatively recent in modern spectral analysis technology, gives a better spectrum estimation for a shorter data observation and also provides a mathematical model for the signal [Fe71, To75, Ma76, Au81, Ch81, Is81]. The model coefficients can be used for the characterization of a segment of the record instead of using the power spectrum parameters . Spectral analysis can provide an efficient quantification of overall broad-band rhythms present in an EEG epoch since the rhythmic components are relatively enhanced at the corresponding frequency. On the other hand, spectral analysis is not suitable for describing short-term, transient events in the EEG since the transients are smeared in the power spectrum, which is an averaging process over a given time window of the data. To achieve a satisfactory spectral estimate for each sleep staging epoch, a relatively long observation, 30 seconds or more, is required. This

PAGE 36

30 however, is obviously not suitable to describe the non-stationary behavior of EEG activities; i.e., short-term, usually less than one second, existing phasic events or paroxysmal events, such as sigma spindles, K-complexes, or REM waves, cannot be described properly by the power spectrum estimate of a long EEG record. The parametric modeling technique is used in several recent approaches to improve the resolution in the analysis and to detect transient activities. A quasi-stationary modeling technique with an adaptive segmentation scheme, where the transient behavior of EEG activity is included as a linear superposition on the segment-wise stationary model, is introduced to solve this non-stationary behavior of the EEG [Bo77, Sa80] . In the study by A. Isaksson [Is81], linear model parameters are allowed with time variances to include the transient behavior of EEG activity in the analysis. However, the postulate of linearity between EEG's stationary behavior and non-stationary behavior is very doubtful, because of the observed evidence that the background activity is wiped out by the appearance of a phasic event. It is also hard to verify the validity of the linear model without including higher order correlation terms for the EEG. Another disadvantage of the spectral analysis method is that the power spectrum is unsuitable for describing the amplitude distribution property of the EEG, especially for

PAGE 37

31 small amplitude EEG waveforms such as beta with a small S/N ratio. This is shown well in the study of A. Lubin, L.C. Johnson, and M. Austin [Lu69], where the sleep stage scoring of human EEG is performed based on the spectral analysis approach. The study reports that the amount of beta activity is minimal in the REM period; however this result is opposite to the visual analysis; i.e., the amount of beta activity is far greater in the REM period than in the other NREM sleep periods. Most of all, the spectral properties of the EEG obtained through the spectral analysis are not the direct EEG activity information of interest to the clinicians. So, the spectral property data should be interpreted in another domain by using various feature extraction techniques and/or pattern recognition techniques. Time domain analysis has also been applied as a relatively simple method to quantify the rhythmic properties of the EEG. Basically, the intervals are measured between successive zero-crossing points or peaks where the first derivative of the signal is zero. The distribution of these intervals measures rhythmic activity. The relationships between the period and the amplitude provide various parameters which describe the characteristics of the EEG [Fe80]. A comparison of periodampl itude analysis, utilizing a period and an absolute amplitude histogram, with spectral analysis shows that while the power spectrum

PAGE 38

32 efficiently quantifies the overall power trend in the EEG data, period-amplitude analysis offers more resolution than the power spectrum in detecting details in amplitude and incidence within relatively narrow frequency bands [Kt81]. Despite its simplicity and usefulness, this approach has been relatively less attractive to the researchers, mainly because it lacks an analytic tool for signal processing and the method is sensitive to noise and other artifacts [Is81]. The autocorrelation and cross-correlation techniques [Mi79] have also been introduced to analyze the EEG, but with efficient computers available these techniques are now replaced by spectral analysis based on the FFT. Hjorth parameters have been introduced to describe the EEG data quantitatively on the time domain, and these parameters are shown to be equivalent to zeroth, second and fourth moments of the normalized power spectral density [Hj70] . The three parameters used by Hjorth are as follows. The first is called "activity" which is the variance of the amplitude or mean power in the epoch. The second is "mobility" which is the average power of the normalized derivative in the epoch. The third is "complexity" which is the average power of the normalized second derivative in the epoch. Heuristic Analysis The recognition of the waveforms of direct interest to clinicians or sleep researchers is the goal emphasized in

PAGE 39

33 the signal processing approach of this research. This, however, differs from the other analytic signal processing approaches in which the EEGs are quantified in terms of statistical variables. The signal analysis approach is referred to here as "heuristic analysis" in the sense that the waveforms are recognized by mimicking the human scorer's visual perception of the waveforms, i.e., by using various grapho-elements and background patterns of the EEG similar to the ones which the human scorer also uses. The signal analysis does not rely on a few limited signal parameters as is the case in period-amplitude analysis, spectral analysis, and parametric modeling approaches. Different sets of various features, including grapho-elements, background patterns, and temporal distributions of the multichannel record, should be utilized for recognizing each different waveform. The sleep EEG analysis problem requires the heuristic approach for the following reasons: The present knowledge about the objective EEG waveform models is not sufficient, and the variabilities involved in the EEG are too large. Thus, a satisfactory analysis of the EEG cannot rely on one specific analytic signal processing technique. Wide variations in the EEG characteristics are interpreted heuristically by the human scorer, and this capability of interpretation should also be reflected in the signal analysis. However, analytic signal processing techniques

PAGE 40

34 are inherently unsuitable to reflect the heuristic factors in the human analysis, since they are based on analytic models of the signal. Because the present knowledge about the EEG and its origin are not yet sufficient to provide objective models for EEG waveforms, it is realistic to rely on subjective models of sleep research and its clinical application areas, i.e., the methodology of EEG analysis based on the visual perception of specific waveforms and the background EEG patterns, rather than to try to provide a new analysis methodology for the clinicians. Most of the analytic signal processing methods emphasize the objectivity of the method and data, and suggest that these data should be utilized in the research and clinical applications instead of visually analyzed data. However, there are significant difficulties in accepting the data provided by the analytic techniques, since reasonable validations of the method and the data, in relation to the neurophysiological models of EEG, cannot be provided for a clinical use. Moreover, the data obtained by the analytic techniques are generally not the direct information the clinicians use. The best contribution of a computer application for EEG analysis at this stage should be in the replacement of the routine part of the labor intensive sleep analysis in a consistent and objective manner with a technique providing the same kind of information.

PAGE 41

mea: 35 Explicit definitions of waveforms in terms of isurable variables are not readily available. The waveforms should be defined ostensively in terms of the various grapho-elements, background patterns, and temporal distribution of the EEG for the heuristic recognition of the waveforms by a computer. Waveform definitions require a process of intensive observation of many records of a wide range of subjects. Since the waveform definitions differ widely from each other, the detailed description of the waveform definitions are given in the individual waveform detector design section and only brief summaries are shown in Table 2.1. The waveform definition criteria include the individual full-cycle wave period window between zero-crossings and peaks, amplitude thresholds, pattern specifications, waveform average period window, half-wave period, wave leading edge slope, background pattern screening, time-coherence of wave shapes in the multichannels, etc.. All the specifications related to the waveform needs to be incorporated appropriately for correct waveform recognition. The weakness of sensitivity to noise or other artifacts can be avoided by carefully designing the detection to incorporate several aspects of the waveform, although this weakness can be understood as a by-product paid for by the increase of resolution in describing waveforms on the time domain.

PAGE 42

36

PAGE 43

37 A simple implementation of the signal processing technique is an important aspect to be considered in the system design. The waveform recognition system should be designed to process all the waveforms on a real-time basis by incorporating several functional blocks, such as the A/D conversion unit, the signal conditioning unit, detection unit, multi-channel information processing unit, and data-link unit, etc.. Waveform Recognition System Design and Implementation The system, which includes multiple detectors with different detection schemes, is described in the following sequence: first, the functional description of the overall system, second, the design considerations, and third, the individual detector details. Functional Description of the System The waveform recognition system is implemented on a TI-9900 microprocessor-based microcomputer system, which includes a 12 bit A/D converter with two D/A channels, one 16 K-byte RAM board, two serial input-output (SIO) ports, and one parallel input-output (PIO) port. The functional block diagram of the system is shown in Fig. 2.8. Data from three EEG channels and one EOG channel (or two EEG, one EOG, and one EMG, depending on the

PAGE 44

38 ch] ch2 chS ch4 llcli3). BETA filter/detector SLOW f i 1 ter/detector DELTA fi 1 ter/detector measurements _ _ SIGMA f i 1 ter/detector THETA filter/detector SLOW filter/detector ALPHA f i 1 ter/detector REM f i 1 ter/detector _ ( first layer ) " SEM filter/detector _ ( first layer ) " MUSCLE ARTIFACT filter/detector EMG level discrimination TOKEN 1 1 1 ENCODING 1 1 1 • 1 1 1 DCM 1 1 . 1 DETECTOR | TOKEN 1 1 sending through S. I.O hos SEM ' 1 1 sys 1 DETECTOR | 1 1 1 1 J K-COMPLEX 1 1 DETECTOR f 1 1 first layer ^ layer ^ Fig. 2.8 Multilayered Parallel Processing Architecture of Waveform Recognition System.

PAGE 45

39 montage selection) are sampled at 480 Hz per channel and passed through a digital low-pass filter with 120 Hz cutoff frequency. There are two processing layers in the waveform recognition system. The first layer includes the individual waveforms detection through the filter/detector unit. The data obtained at the first layer are used for the second layer processing. The second layer processing is performed at a much wider interval of 0.25 second in order to detect certain waveforms, such as the REM, SEM, and K-complexes, which require the information of the other channels and/or the observation of relatively long periods of adjacent EEG/EOG background patterns. Also, the information of all the detected waveforms and the delta wave measurements are encoded in a character stream (tokens) and sent to a host computer for further analysis. The processing interval of 0.25 second for the second layer processing is selected by taking into account the following factors: the data compression requirement such that one 5-1/4 inch floppy diskette should be sufficient to accommodate an entire night's four channel token data; the efficiency and time-resolution requirement for the background pattern screening; and the period range of the delta wave. Since, the single processor (TI-9900) system must be capable of processing the multi-tasks for real-time detection of the various waveforms, an appropriate timemultiplexing scheme needs to be designed for the parallel

PAGE 46

40 processing architecture of the waveform detection system. The time-multiplexing scheme of the system is shown in Fig. 2.9. Twelve time-slots consist of one cycle-frame where all the detectors and filters are multiplexed according to their sampling rates as is shown in the figure. Each time-slot, which is 1/480 second, contains several detectors and filters which are assigned according to the sampling rates of each filter and detector. Three of the time-slots, as marked in the figure, additionally allocate one of the second layer detectors once every 0.25 second. The simplified main program flow is depicted in Fig. 2.10. After the initialization of all the buffers, A/D converter, interrupt control registers, and the other I/O ports, the program loops the cycle-frame for the processing. At the end of each time-slot, the interrupt handler routine is triggered by the internal clock. This routine performs the A/D conversion and checks the timer count to execute one of the second layer detectors or token encoding routines. The A/D converted data of each channel are low-pass filtered at 120 Hz to remove the high frequency components of the signal above 120 Hz which otherwise may cause aliasing to affect the filters and detectors of the later section. The flow chart of the interrupt handler routine is shown in Fig. 2.11. The format of token data is shown in Fig. 2.12. Information of all the detected waveforms and delta

PAGE 47

41 Proce me .slot sses\ 1 2 3 4 5 6 7 8 9 10 11 12 ALPHA "preMlter o o o o o o detector o o o BETA O o o o o o DELTA prefilter o o o o o o detector o o SIGMA o o o o o o THETA o o o MUSCLE ARTIFACT o o o o o 0 EMG o o I o o o o SL0W(F1-F7) o j o o SL0W(C3-A2) o 1 o o SEM (I'st layer o o ° i REM (I'st layerl o u o 1 Second laye' processing SEM K-COIHD REM 1_ I 1 Token encoding Fig. 2.9 Time-multiplexing Scheme of Waveform Recognition System (one cycle-frame).

PAGE 48

42 Initialization : Buffers I/O ports A/D converter Interrupt control rPffi<;tPr<; % ^^r^r\, 5 Multiplexing Cycle-frame 12 time-slots Interrupt handler at the end of each time slot Major Flow Diagram for the Main Routi of Waveform Recognition System.

PAGE 49

43 Token sending yes REM detection yes K-comp detection yes SEM detection ^ Zes. Token Encoding yes i A/D Conversion 4 ch 480 Hz 2.11 Interrupt Handler Routine Flow Di aqram.

PAGE 50

44

PAGE 51

45 measurement data are encoded into a character string using a signature and bit mapping scheme. Token-encoding routine reads the flag of each waveform detector at every 0.25 second interval and uses two bytes for the encoding of all the flag information. It assigns one (high) to the corresponding bit of the token if the flag of the waveform detector is high, or zero (low) if the flag is low. If a delta wave is detected, then additional two bytes are assigned for the amplitude and period information of the delta wave, respectively. The additional two bytes can be distinguished from the other pair of bytes by the signature bit assigned to the 3rd significant bit of the byte which contains the delta amplitude information. The sign of a delta wave is marked at the 3rd significant bit of the byte which contains the delta period information. Design Considerations Several important design considerations in terms of signal processing aspects, including the selection of A/D conversion unit, the determination of sampling rates for each filter and detector, and the filter characteristics, must be carefully reflected in the system design. The major functional components of the system are the A/D conversion unit, the signal conditioning filters, the individual waveform detectors, and the token encoding and data linking unit.

PAGE 52

46 It has been shown that a 12 bit A/D converter with 16 bit arithmetic processing is a suitable selection to obtain at least a 40 db S/N ratio [Pr86] . The A/D conversion sampling rate is one of the important design parameters. The rate must be appropriately determined by considering the overall implementation constraints together with the characteristics of the signal. A higher sampling rate generally brings a higher measurement resolution, but stiffens the constraints for the real-time processing of the filters and detectors. The sampling rate on the other hand must be high enough so that the aliasing, possibly caused by the sampling, would minimize distortions in the signal of interest. The A/D conversion rate is 480 Hz in this system for each of the four channels. It is safely assumed that the aliasing contributed by the signal components above 240 Hz (1/2 * 480 Hz) is negligible. The sampling rate of 480 Hz also gives an appropriate implementation combination with the required sampling rates of the next stage filters and detectors. The digitized data at 480 Hz is first low-pass filtered with a -3 db cut-off frequency at 120 Hz. The digital low-pass filter is implemented with the tranfer function H(z) = 0.5 * (l1/z). The signal components under 120 Hz contain enough information for the waveform detection and analysis purposes of this system, including the detection of the muscle artifact and the level discrimination of the EMG. The

PAGE 53

47 muscle artifact and the EMG signal contain the highest frequency components compared to the other waveforms. The signal at the output of this low-pass filter can be further assumed to be band-limited at 120 Hz. Since the power spectrum of the EEG itself shows 1/f characteristic, the frequency components above 120 Hz are drastically reduced and are consequently negligible at the filter output. This lowpass filtering allows a lower sampling rate implementation of the later stage filters and detectors, reducing the overall implementation complexity. Relatively broad-band linear phase FIR (Finite Impulse Response) filters are used for the signal conditioning filters in the system. The broad-band frequency response characteristics are required for the signal conditioning filters, since narrow-band filters may distort the waveforms of interest. The flat pass-band and the sharp cut-off characteristics of the amplitude response are not important for the signal conditioning in this system. The broad-band characteristics and the marginal requirement in pass-band amplitude response and cut-off characteristics of the filter make it feasible to come up with a simplified filter implementation scheme [Pr86]. The filters designed in this way do not require any multiplications or floating point arithmetic calculations, but can be implemented by shifts and additions. The sampling frequency of an individual signal conditioning

PAGE 54

48 filter is constrained by two factors. The first constraint is the filter order, which increases as the sampling frequency increases when the bandwidth of the filter is fixed, thus increasing the processing time and the filter internal magnifications resulting in the possibility of an overflow. The sampling frequency therefore is limited by the highest feasible order of the FIR filter, which is reported as being around 10th order [Pr86] . The second constraint is the lowest possible sampling frequency of the filter limited in terms of the Nyquist rate such that the filter sampling frequency must be higher than twice the band-limit of the signal to avoid aliasing. The period measure between zero-crossing points and/or peaks, which is performed by counting the samples between the two points, is one of the important parameters included in the detector unit. The sampling rate of the detector unit is therefore very important in terms of the measurement resolution. One sample interval is the maximum possible error in determining the period of two points by counting the number of samples. The measurement resolution is readily represented by the following equation. df = fo^/fs 1/fo is the period of the two points, fg is the sampling frequency, df is the error in frequency.

PAGE 55

49 An example is illustrated with the case of beta. In case of the beta, fo is approximately 25 Hz thus df is +-3 Hz with the sampling frequency fg = 240 Hz. Increase in the sampling frequency will improve the measurement resolution, but this at the same time decreases the available time-slot; thus, the increase of the sampling frequency cannot be an appropriate solution to improve the resolution. The measurement resolution can be increased also by the interpolation scheme. But the interpolation scheme also brings an increase in the processing time and involves an error of its own. One way, which is used in this system to improve the resolution, is to take the period as the average period of the several adjacent waves. In this way the measurement error is reduced by the factor of the number of the sequences included in the total period as is shown in the following equation. df = fo2 / (N * fs) Spindles Detection The appearance of spindle bursts, such as the alpha spindle, beta spindle, theta spindle, and sigma spindle, is one of the well observed phenomena in human sleep EEC. These waveforms are conveniently grouped as spindles, and their detection schemes are basically similar, although the detailed nature of the waveforms is slightly different for

PAGE 56

50 each. The first important problem for the correct detection of these waveforms is that these waveforms must be defined explicitly in terms of variables measurable by a computer. The waveforms are however heuristically perceived by the human scorer. Thus, straight foward and explicit definitions are not available. Ostensive definitions, which could be obtained by mimicking the human scorer's visual perception reflecting all the various aspects of the EEG, have to be used in the detector design. A spindle is a short burst of waves in sequence that form a waveform giving a distinctive appearance from the background. The spindles are roughly defined by the measure of the periodicity of the individual waves and by the specification on the grouping (bursting) nature of the waveform. These can be further broken down into the following criteria. o Individual wave period wi o Individual wave amplitude o Average period window for in a spindle o Pattern specification for idow: zero-crossings : peaks threshold the whole sequence of waves the waveform

PAGE 57

51 A typical functional block diagram of the spindle detector is shown in Fig. 2.13. A spindle detector consists of a linear phase FIR filter followed by a full-cycle period discriminator, amplitude detector, positive-peak interval discriminator, and pattern recognizer. Appropriate signal preconditioning is necessary to remove the effects of high frequency noise and of large-amplitude, slow waves. Especially, if a spindle is superimposed on a large-amplitude, slow wave, it is impossible to detect the spindle at the next detector unit by the zero-crossing and peak detection scheme. A relatively broad-band linear phase filter is used for this signal pre-conditioning purpose as explained. A Sigma spindle is more subject to the effects of large-amplitude, slow waves, because a fair number of sigma spindles are superimposed with K-complexes and delta waves. On the other hand, the effect of the large-amplitude, slow waves are less serious for the other spindle waves, such as beta, theta, and alpha, since these other spindles mostly appear on a flat EEG background while the EEG amplitude is relatively small. Thus, the lower edge cutoff characteristics of the sigma spindle filter need to have a sharper cutoff characteristic such that the filter successfully removes the effect of large-amplitude, slow wave of the EEG without any significant distortion of the

PAGE 58

52 zero-crossing Filtering & amp! i tude peak detection threshold period window zero-crossings positive peaks average period window pattern criterion Fig. 2.13 Sigma Spindle Detector Functional Block Diagram.

PAGE 59

53 spindle waveform. The filters for each spindle are summarized together with the filters for the other waveforms In Table 2.2. The signal passed through the conditioning filter is processed at the next detection unit with various criteria. The frequency of a wave is first defined as the inverse of the full-cycle period which is measured by counting the Interval between adjacent two positive-peaks (or negative-peaks) and/or two p o s 1 t 1 v e g o 1 n g (or negative-going) zero-crossing points. This definition of the frequency is different from that of the sinusoidal wave frequency mostly referred to as a spectral composition of signal in the engineering or scientific sense. However, the definition given here is better justified as more closely mimicking (simulating) the human scorer's, since the human scorer measures the frequency of a waveform by counting the number of peaks in a specified time window. in the most general sense, the frequency is defined by a human as the number of events occurring per a unit period. The definition of frequency in terms of pure sinusoidal wave is one very specific case of the general definition. The full-cycle zero-crossing points are first detected by checking the sign change of the samples. Then, each wave's positive-peak is defined by the maximum sample value between the zero-crossing points. The positive-peak is described by the saved maximum positive amplitude and the

PAGE 60

54

PAGE 61

55 relative time displacement of the peak from the leading zero-crossing point. The peak has been continuously updated whenever a new sample value is greater than the previous saved peak within the period. If the peak of the full-cycle wave is greater than a certain amplitude threshold, then the periodicity of the individual wave is tested by using corresponding period windows for both intervals between positive-peaks and full-cycle zero-crossing points. The reliability of the detection generally can be improved by including more measurements in the specification. For instance, the full-cycle zero-crossing period is not sufficient for the spindle wave periodicity measure because the reliability of the detection could be affected by the period measurement error caused by large-amplitude, slow EEG waves, or other noise effects. On the other hand, a spindle's positive-peak interval distribution gives a very good measure to mimic the human scorer's visual perception of the spindles. The positive-peak interval is less subject to the large-amplitude, slow waves of the EEG. A positive-peak interval is a measure of the period between the previous wave's peak and the present wave's peak point; the interval is obtained by adding the two intervals, one betwee.n the previous wave's peak and the previous wave's ending zero-crossing point, and the other one between present wave's leading zero-crossing point and wave positive-peak.

PAGE 62

56 Thus, if this positive-peak wave interval is out of specified range for a spindle, the previous wave is discarded even though it meets the zero-crossing period and amplitude criteria. However, the way this system detects and utilizes the information about peaks is distinguished from the generally understood peak detection technique. In general, the peak detection technique looks for the points where the first derivative value of the wave is zero, and applies appropriate period and amplitude criteria on these peaks to analyze waveform characteristics. This approach is taken in the Sigma spindle study done by J.C. Principe and J.R. Smith [Pr82] . However, this peak detection technique is more subject to the local variabilities of waves than the zero-crossing technique. On the other hand, in this system, the global wave peak is detected by taking the maximum sample value between zero-crossing points; thus the positive-peak is less subject to local variabilities of the signal . The positive-peak interval criterion is used together with the zero-crossing period specification for a reliable spindle detection, since the positive-peak interval alone can not impose any restriction on the wave's vertical variation in terms of the vertical asymmetry from the baseline. Also, the two-period-window scheme gives more flexibility in specifying the spindle wave periodicity. The

PAGE 63

57 scheme allows a looser specification for each period window allowing more variations in terms of each period specification. But, on the other hand, the scheme reflects a tighter specification in terms of the global spindle shape specification by using the two period windows. The human scorer is not very sensitive to the individual wave period and amplitude variances in detecting the spindles. On the other hand, he relies more on the well-shaped global appearance of the whole waveform. The pattern criterion and the average period window are used to specify the global spindle waveform grouping (bursting) nature. The average period window is applied to the total period of several adjacent waves sequence in the waveform. The average period window increases the measurement accuracy by a factor equal to the number of waves averaged as is discussed in the previous section. A tighter average period window is applied for the specification of the spindle waveform, on the other hand, the period windows for individual waves can become looser to reflect individual wave variances in the spindle. The three numbers in the pattern criteria specify the total number of waves kept in a window, the minimum required number of consecutive in-band waves in the waveform at onset, and the minimum number of the in-band waves to sustain the detection, respectively. The summarized specifications for the spindle waveforms are shown in Table 2.3. The flow chart of the spindle detector

PAGE 64

59 is shown in Fig. 2.14. This flow chart is applied for all the spindle detectors with appropriate changes in the parameter values according to the criteria of each spindle detector. Rapid Eye Movement Detection The REM wave is detected based on ostensive definitions which include the rapidly rising leading edge, quiescent nature of EEG activity during REM sleep, and the REM wave amplitude and period window. The rapidly rising REM wave leading edge is used as a distinguishing feature from other EEG waves such as delta waves, and SEM waves occurring in the EOG channel. It is also necessary to observe the nature of EEG channels' activity during REM sleep to eliminate false REM detections caused by the EEG activities appearing on the EOG channel. In particular, the appearance on the EOG channel of large-amplitude EEGs, such as k-complexes or delta waves in NREM sleep, may cause false REM detections. This false positive detection has been a problem in accurate REM wave detection using automated computer analysis methods [Le84] . The EEG during REM sleep is typically represented by the appearance of small-amplitude and high-frequency waves, no delta-like, large-amplitude, slow waves, no rolling movements, thus giving a quiet and flat appearance from the EEG baseline.

PAGE 65

60 decrease display counter yes set overflow flag 1 ncr tirr ease er 1 — ^ get new peak shift wave window reset overflow flaa yes increase W cnt decrease W cnt reset W cnt set flaq set display countfer set disp-flag reset diso-flag reset flan output disp-flag clear timers clear peaks : ^' return Fig. 2.14 Sigma Spindle Detector Flow Chart.

PAGE 66

61 The main idea behind the REM wave detection algorithm of this system is to get a robust REM wave detection by using both the descriptive REM wave criteria and the EEG channels' background screening with the appropriately chosen time windows. The descriptive REM wave criteria include a slope threshold (350 uv/sec) on the leading edge, a period window (0.2 2.0 sec/half -wave ) , and a peak amplitude threshold (30 uv). Two quiescent testing time windows are applied, one each, to the central (C3-A2) and to the frontal (F1-F7) channel EEGs as is shown in Fig. 2.15. The central channel time window is applied to the preceding and succeeding two seconds from the terminating edge zerocrossing point of the REM wave. The frontal channel time window is applied to the preceding and succeeding one second from the terminating edge zero-crossing point of the REM wave. The REM wave definition is therefore the following. o Period: 0.2 2.0 sec/half-wave o Amplitude: 30 uv o Slope: 350 uv/sec o Quiescent Test Frontal channel: +-1 sec window, peak amplitude ratio > 0.5 (between the biggest waves in the window) Central channel: +-2 sec window, no delta wave

PAGE 67

62 F1-F7 C3-A2 OS-OzPz 1 sec 2 sec 1 sec 2 sec *'"'*' .H T I ifli Fig. 2.15 Time Windows for EEGs Background Pattern Screening

PAGE 68

63 A twolayered screening scheme implements the realtime REM detection algorithm. Each REM wave is detected at the bottom layer through the filter/detector unit with the descriptive waveform criteria, and the EEG channels' background screening is performed at the upper layer with the 0.25 second sampling interval. The functional block diagram of the REM detection is shown in Fig. 2.16. The leading edge slope is defined by the line connecting the leading zero-crossing point and the wave peak point of the signal as is shown in Fig. 2.17. In terms of the global description of the REM wave leading edge, this measurement is found to be a better representation of the edge slope than the measurement where the edge slope is defined as the locally averaged maximum slope value. In terms of accuracy in measuring the edge slope, this global slope measurement is more accurate because the effect of quantization error is greater for the shorter duration measurement. Also in terms of the robustness to the erroneous local variabilities of waves, the global slope measurement is better than locally averaged slope measurement. To eliminate erroneous effects caused by the EOG baseline movements in measuring the leading slope, a "dead-zoning" is first applied to the filtered output signal with an appropriate selection of the dead-zone level. One third of the minimum REM wave peak threshold, i.e. 10 uv, is

PAGE 69

64 f i 1 ter & dead-zoning period & amplitude threshold slope threshold > first layer Frontal quiescenCentral quiescent ± 1 sec ± 2 sec peaks comparison no delta second layer Fig. 2.16 REM Detector Functional Block Diagram.

PAGE 70

65 Fig. 2.17 Illustrations of Dead-zoning and Leading Slooe Measure REM Wave Detection, (a) EOG (LE-A2); (b) Output from the Filter/Dead-zoning Unit.

PAGE 71

66 selected as the dead-zone level, and is found sufficient for the removal of the erroneous effects caused by the EOG baseline movements. This level is not too big for the purpose of an appropriate edge slope representation. To prevent the effect of filter smoothing in measuring the slope, a relatively wide lowpass filter (-3 db at 30 Hz) is used only for the removal of high-end frequency components in the signal [Kt75] . For the second layer, both the central channel (C3-A2) and the frontal channel (F1-F7) are used to test the EEC's quiescent nature during REM sleep resulting in a higher reliability than testing one EEC channel. A fixed amplitude threshold is used for the C3-A2 channel's quiescent test criterion. On the other hand, the relative peak amplitude ratio between the EEC peak and the EOG peak is used for the F1-F7 channel's quiescent test. It is observed that the F1-F7 channel often displays the appearance of delta-like waves, which actually is caused by eye movement activity. These deltalike waves may give an error in testing the F1-F7 channel's quiescent characteristics during REM sleep, although the magnitude level of the delta-like wave at the EEC channel is smaller than the REM waves. It is observed that the frontal ( F1-F7 ) EEG channel shows more REM-related cross channel information than the central (C3-A2) EEG channel. The dominant appearance of

PAGE 72

67 beta on the frontal channel during REM sleep is far more noticeable than it is on the C3-A2 channel. Beta is one of the important waveforms which can be utilized, together with the REM wave detection, for the purpose of accurate demarcation of REM sleep periods. It is also observed that the opposite and synchronized phasic waveform relationship of k-complexes between EOG and EEG channels is more clearly shown on the frontal channel than on the other EEG channels. The Fig. 2.18 shows the flow chart for the first layer processing of the REM detection. The Fig. 2.19 and 2.20 show the flow charts for the second layer processing of the REM detection. Slow Eye Movement Detection The slow eye movements (SEM) cause the EOG channel's slow, rolling fluctuations from the baseline. The SEM is physiologically related with the eyes floating, especially during the transition period from wake to sleep caused by drawsiness. The SEM detection scheme includes a half -wave period window, an amplitude threshold, the central EEG channel's background screening, and a wave leading slope threshold. The flow chart for the SEM detection is basically similar to that of REM detector except for the absence of the peak comparison criteria with the other EEG channels ,

PAGE 73

58 I EOG filter & dead-zoning 0 30 Hz, 10 uv decrease delay countpr ray count^rv^no > 0 ? ^ — ^ clear REMl flag yes increase timer / get new p eak clear overflow flag clear REMl (I'sft layer) flag update peak" set REMl flag delay counter clear REMl flag clear timer clear peak return Fig. 2.18 REM Detector Flow Chart (first layer).

PAGE 74

69 decrease counter: REM peak counter' EEG peak counter output Low clear REM-flag output Low clear REM-flag set flagl yes clear flagl clear flagZ set flag2 set REM-flag output High clear REM-flag clear REM-window( output Low Nj/to Part.B. Fig. 2.19 REM Detector Flow Chart (second layer part A).

PAGE 75

70 from Part. A, shift left REM-window( ) REM-window(l)<— 1 shift left slow-window( ) slow-flag(froIrtal EEG) ^ f 0 ? ^ 1 sec) slow-wir idow(l) 1 ^ peaks (pre update. (EEG & RF r V. 1 sec) M peaks) \f return Fig. 2.20 REM Detector Flow Chart (second layer part B).

PAGE 76

71 K-complexes Detection The main usage of the k-complexes detector in the present system has twofolds: to gain better confidence in scoring sleep stage two by complementing the spindle detector information, and to accurately demarcate the sleep stages two and three. The K-complexes detector uses an algorithm structure similar to that of the REM wave detector, since the detection of the K-complexes also requires the cross-channel information such as the synchronization of peaks, the opposite phasic relationship between EEGs and EOG, and a quiescent criterion. Presently, synchronization of EEGs and EOG is detected with a time window, i.e., 0.75 second, across the EEG and EOG channels (specifically F1-F7, C3-A2, and EOG). The amplitude threshold (50 uv) and the relatively loose period threshold (0.3 1.5 sec/half -wave ) are used as the k-complexes criteria. The current definition of K-complexes in the algorithm requires no delta-like slow and large-amplitude wave within the interval of preceeding and succeeding two seconds from the K-complexes, since the present K-complexes detection algorithm is less reliable in sleep stages three and four. Therefore, the K-complexes information in these sleep stages is not utilized in the sleep scoring process. The more descriptive K-complexes detection criteria, such as the opposite phasic relationship between the cross

PAGE 77

72 EEG channels and the HOG channel, upward going characteristic, and slope criterion should be utilized for a better K-complexes detection. Delta Wave Detection and Measurement The delta detector is implemented with a period window (0.25 1.0 sec/half -wave ) and a relatively low amplitude threshold of 16.7 uv. The amplitude and period values of each delta wave are sent to the host computer through a serial port for a post delta processing. The post delta processing is necessary for an accurate delta summary because the variations of the EEG amplitude level provide poor delta wave detections if the same amplitude threshold is used for all subjects. Individual delta wave amplitude and period are necessary also for the purpose of more quantitative delta studies. The delta amplitude is quantized into 16 levels from 16.7 uv to 100 uv with the 5.2 uv resolution and the period is also quantized into 16 levels from 0.25 to 1.0 second with the 47 msec resolution. Muscle Artifact and EMG Analysis Muscle artifact is associated with the movement of body. The signal consists of high frequency components usually above 30 Hz. The detector consists of a highpass filter, a zero-crossing detector, an amplitude threshold (10

PAGE 78

73 uv), a full-cycle period discriminator (34.3 120 Hz), an average period window, a 60 Hz notch filter, and a pattern specification (6/6/3). The flow chart of the muscle artifact detector is same as that of the spindle detector as is shown in Fig. 2.14, except for the addition of 60 Hz notch filter. The 60 Hz notch filter is operated in conjunction with the average period window. The purpose of this filter is to reduce the effect of 60 Hz environmental noise in detecting muscle artifact. The EMG signal amplitude is described into three levels, i.e., below 10 uv, between 10 uv and 20 uv, and above 20 uv. The EMG discriminator structure is same as the muscle artifact detector but three different amplitude thresholds are applied for the EMG level discrimination.

PAGE 79

CHAPTER III SURVEY ON EXPERT SYSTEMS AND THEIR APPLICATIONS In this chapter, general aspects of expert systems are briefly reviewed with explanations on their historical background, distinguishing features from algorithm-based programs, and architectural aspects. Several expert system application examples are then briefly reviewed with an emphasis on a critical view of the expert system approach, discussing their limitations and future prospects in relation to their application domain characteristics. Expert Systems Overview Knowledge-based expert systems have been drawing considerable interest in the application-oriented artificial intelligence (AI) field since the mid-1960s. This area of Al has concentrated on the construction of a high-performance system which executes tasks at the level of a human expert in a specialized professional domain. Historically, the knowledge-based expert system is a fundamental shift in the application-oriented AI area from the earlier effort to find general methods for problem-solving and use them to create general purpose programs. This strategy to achieve a general problem 74

PAGE 80

75 solver, despite some interesting progress, has produced no significant success because developing general-purpose programs has been too difficult and ultimately fruitless. The main emphasis has shifted to the knowledge imbedded in a problem domain. Researchers have pointed out that the problem-solving power of a program comes from the knowledge it possesses, not just from the formalism and inference schemes it employs. This conceptual shift, thus, leads to the development of special-purpose computer programs, systems that are expert in some narrow problem area. These programs are called expert systems. How to extract, construct, and manipulate the domain knowledge constitutes the major concern in developing an expert system. A new set of principles, techniques, and development tools has emerged and that forms the basis for knowledge engineering [Ba82b, Da82, Ha83b, Wa86] . In comparison with algorithm-based computer programs of the conventional approach, the knowledge-based expert system is best distinguished by the characteristics of the problem domain it is applied to. It is not suitable to apply the algorithm-based conventional programming approach to expert system's problem domains. In these problem domains, tractable algorithmic solutions usually do not exist, since many important tasks are related to the intensive heuristic knowledge of a human expert, thus resisting precise description and rigorous analysis.

PAGE 81

76 Planning, legal reasoning, medical diagnostics, geological exploration, and analysis of military situations exemplify these problems. Contemporary methods of symbolic and mathematical reasoning, which have limited applicability to the area of expert systems, do not provide a means for representing knowledge, describing problems at multiple levels of abstraction, allocating problem-solving resources, controlling cooperative processes, and integrating diverse sources of knowledge in inference. These functions depend primarily on a capacity to manipulate problem descriptions and to selectively apply relevant pieces of knowledge. The expert systems, depending on the application problem domain, are categorized into several types such as interpretation, prediction, diagnosis, design, planning, monitoring, debugging, repair, instruction, and control systems [Ha83b, Ge83, Wa86]. Expert systems are also distinguished by different program structures from the algorithm-based programs of conventional approach. Expert systems typically consist of four major structural components, i.e., a knowledge base, a data base, a control mechanism, and a knowledge-base editor. In the expert system structure, emphasis is on the separation of domain knowledge into a separate knowledge base in the form of modularized pieces of independent knowledge. A control mechanism provides a general strategy in trying to match the relevant findings and data in the

PAGE 82

77 data base to the knowledge base in order to derive a conclusion. Since, most of the domain knowledge is represented by a set of modularized pieces and is constructed in a separate knowledge base, the domain knowledge can be handled in a flexible and manageable way by a knowledge-base editor. The control mechanism can also be in a flexible and generalized form in the knowledge-based expert system structure. The knowledge-base editor, which is a user interface for the manipulation of the overall system, is also a distinguishing feature of an expert system, providing a flexible and transparent environment for the user. Without any modification of programs, a user can access and change the system knowledge through the knowledge-base editor. The expert system often posesses the capability to explain the line of reasoning for a conclusion through the knowledge-base editor. Expert system architectures vary widely depending on their applications. The current techniques and principles of expert systems are based mostly on the relatively small number of early expert systems. Rule-based knowledge representation constitute the major framework of the current expert system technologies [Ha85a] . Although the rule-based knowledge system is appealing because of its general expressive power and well-defined syntaxes, more sophisticated reasoning schemes and architectures are required for many intrinsically harder problem domains.

PAGE 83

78 These domains usually include the problems of a large search space of possible solutions, diversity of knowledge, dimensionality of data or knowledge (e.g. time, space), lack of fixed inference process to solve the problem effectively, handling of interactions among decisions of separate subproblems of the task, need for specialized representations, and incompleteness of data. A frame-based representation [Fi85] and a blackboard model approach [Ha85b] are among the several advanced techniques to cope with these difficulties. A frame provides a structured representation of an object or a class of objects. Each class of objects can be described as a specialization (subclass) of other more generic classes, thus the classes are represented into taxonomies in organizing frames. The frame provides a concise representation of useful relations, and supports a concise definition-by-specialization technique that is easy for most domain experts to use. In addition, this representation structure provides automatic inferences as part of each assertion and retrieval operation. The taxonomic relationships among frames enable descriptive information to be shared among multiple frames via inheritance. The internal structure of frames enables an automatic maintenance of semantic integrity constraints. Production rules by themselves are inadequate to define terms and to describe domain objects and static relationships among objects. Hybrid representations are

PAGE 84

79 constructed by integrating frame and production rule representations, resulting in the consolidation of advantages of both representation techniques. The frames provide a rich structural representation for describing the objects referred to in the rules. The frames support a layer of generic deductive capability about those objects that do not need to be explicitly dealt with in the rules. Frame taxonomy can also be used to partition, index, and organize a system's production rules. The blackboard model of problem solving is a highly structured special case of opportunistic problem solving [Ha79]. In addition to opportunistic reasoning as a knowledge-application strategy, the blackboard model prescribes the organization of the domain knowledge, all inputs, and intermediate and partial solutions needed to solve the problem. The solution space is organized into one or more application-dependent hierarchies. The domain knowledge is partitioned into independent modules of knowledge that transform information on one level of the hierarchy into information on the same or other levels. The knowledge modules perform the transformation using algorithmic procedures or heuristic rules that generate actual or hypothetical transformations. Reasoning is applied within this overall organization of the solution space and task-specific knowledge. in other words, the module of knowledge to apply is determined dynamically, one

PAGE 85

80 step at a time, resulting in the incremental generation of partial solutions. At each step of the knowledge application, either forwardor backward-reasoning methods can be applied. Application Examples And Future Prospects Application Examples One of the applications of expert system technology is in the building of medical consulting systems designed as aids to medical decision making [Sz79, KuBO, Bo84] . Mycin [Sh76], Casnet [We78b] , and Internist [Po75 Po77] are among the examples of early expert systems mostly often referred to as typical expert system applications. The Mycin system was developed to provide consultative advice on diagnosis and therapy for infectious diseases. The system is in a form of a production rule based system, and the exactness and incompleteness of data and knowledge are handled by the Certainty Factor scheme. Casnet is a computer system for medical diagnosis in the treatement of glaucoma. The system represents a disease not as a static state but as a dynamic process modeled by a network of causally linked pathophysiological states. The system diagnoses a patient by determining the pattern of pathophysiological causal pathways present in the patient and identifying this pattern with a disease category. Internist is a consulting program

PAGE 86

81 in the domain of internal medicine. Using the information presented during the consultation, the program tries to discriminate between competing disease hypotheses. Thus, this system is of the type that verifies a hypothesis formation using the system's knowledge represented in the form of a disease tree, or disease taxonomy. The application of expert systems in medicine was motivated by two major reasons. First, the expert computer-based system can provide great benefits by providing thorough and reliable diagnostic services. Considering that most of the errors made by the clinicians are caused by omissions involved in the diagnostic process, the computer can provide reliable diagnostics by an exhaustive consideration of all the possibilities and all the relevant patient's data. The second motivation is related to current research interests in application oriented artificial intelligence. Clinical medicine has been a fertile area for the study of cognitive processes, and diagnosis as a cognitive process has been studied extensively. Historically, statistical analysis and pattern recognition, through a discriminant function based on the Bayesian decision theory, have been used for the development of computer analysis of medical diagnostic problems. The appeal of the statistical method is that the decisions based on such methods are optimal for given criteria. However,

PAGE 87

82 the statistical approach is unsuitable in medical problems, because of various assumptions and simplifications such as independence and mutual exclusiveness of various disease states. These assumptions cannot be suitably validated and the a priori and conditional probabilities required in the analysis are usually not available. The medical decision aid is a typical example of a heuristic domain which includes a large amount of domain specific medical knowledge and physicians' heuristic knowledge obtained through years of practice and experience with many special cases. The physicians provide a diagnosis based on the heuristic knowledge for handling incomplete data and uncertain information. One of the difficulties in designing a medical consulting expert system with a human expert-like performance is that the human expertise comes from his capability to consider and handle a large number of special cases. As the spectrum of a task becomes larger, more and more of special considerations are involved in solving the problem. Moreover, the heuristic knowledge of a physician also involves a kind of commonsense knowledge obtained through a long period of experience and education, and this commonsense knowledge is immense in its nature and is usually very difficult to articulate and represent in a manageable form in a computer.

PAGE 88

83 The development of computer-based consultation systems also brings many formidable social, physical, and ethical problems that must be considered in expert systems' design. These problems include validating the systems, exporting them to hospitals or laboratories, getting physicians and patients to accept them, and determining the level of responsibility for the clinical decisions made with the help of these systems. It is reported that the biases of medical personnel against computers are so strong that systems will eventually be rejected, regardless of performance [St72] . Thus, there exists another barrier in achieving a significant contribution of computer diagnoses consulting systems for routine applications. Several expert system applications have been shown also in the domain of chemistry and other engineering and scientific application areas. Dendral, Prospector, and Hearsay systems are among the well-known expert system examples in those application areas. Dendral [Bu69, Ba82b] generates plausible structural representations of organic molecules with data obtained from mass-spectrographic analysis of unknown molecules and with a set of rules used by an expert chemist to infer constraints on molecular structures from such data. The elucidation of molecular structures is fundamental to the application of chemical knowledge to important problems in biology and medicine. In many circumstances, the powerful analytic

PAGE 89

84 techniques of X-ray crystallography and X-ray fine structure analysis may not be applicable; consequently the analysis process is much too laborious and dependent on the intuitive expertise of chemists. Thus, there exists a legitimate reason for developing a computer aided expert system, and room for a significant contribution of the computer analysis . In this type of molecular structure elucidation or chemical compound synthesis domain, the major problem aspect is a heuristic search in limiting the combinatorial explosion of all the possible structural candidates. However, the process of making up a set of rules about mass-spectrometry proves to be much too involved, since the theory of mass spectrometry is incomplete and the rules about it are inexact and difficult for experts to explicate. The Dendral system's level of performance is generally far less than that of a human expert. This indicates the difficulty involved in codifying the human expert's skilled process in the system. However, there exists a possible contribution of computer analysis in the structure elucidation domain, because a computer can provide a systematic search through the space of possible molecular structures, a systematic use of what it does know to constrain the possible structures, and the calculating power to handle fair amount of calculations involved in the process .

PAGE 90

• 85 Prospector is a computer-based consultation system for mineral exploration [Du79, Ba82b] . The main function of Prospector is to match data from surface geological observations against models of five different types of deposits. In the Prospector system, contextual information and data are treated in the same plane of a probability propagation model, and the control scheme is heavily dependent upon the numerical comparisons. Thus, the system lacks flexibility for considering high level context information and situational considerations in an appropriate way. The model, which is based on the probability propagation, cannot provide a sufficient validation. HEARSAY speech understanding system [ErSOa, Er80b] was developed in the domain of voice chess. This system was constructed with the idea of independent knowledge sources cooperatively solving a problem by posting hypotheses on a global blackboard data structure. Isolating the knowledge sources along functional lines provides efficient modification of the probl emsolving structure of the program, by allowing a free substitution of independent molecular knowledge sources. This molecular structure, where the knowledge sources do not address each other directly, allows great flexibilty as the system evolves and different combinations of knowledge sources and control strategies are tried. This blackboard structure has also been incorporated into several other systems solving diverse

PAGE 91

86 tasks in crystallography, signal interpretation, vision, and psychological modeling. Limitations and Future Prospects Expert knowledge consists of two abstractively described parts. First, it consists of the symbolic descriptions that characterize the definitional, taxonomical, and empirical relationships in a domain and secondly, the procedures for manipulating these descriptions. To achieve a high-level of performance, a human expert's skills also need to be well understood and included in the system's knowledge base. A skilled process is usually related to fast response, efficiency, reduced error, reduced cognitive load, and increased adaptability and robustness. The term skill conveys the idea of appropriate knowledge and its effective use. Most of the current expert systems are based on a limited number of knowledge-representaion and inferencing techniques. However, there exist several fundamental limitations in realizing human-like intelligence in a computer by the present A.I. approaches which are mostly based on the description of features and rule-like inferences. The essential problems related to the immense commonsense knowledge and to the lack of knowledge for human being's recognition processes, such as image-based inference, similarity recognition, and relevance pruning

PAGE 92

87 processes are crucial barriers to overcome the limitation of the current application-oriented A.I. technology in achieving a machine intelligence competetive with a human expert's level of performance. Based on this assertion, A.I and, in particular the expert system approaches are severly criticized in the book by H.L. Dreyfus and S.E. Dreyfus [Dr86] . It states that it is impossible to achieve a computer system which can show human expert level performance incorporating human expertlike processes. This opinion is appealing considering that the fundamental questions raised regarding the human cognitive processes are not answered and most of the current techniques are based on superficial simulations. More importantly, expert level performance is related to the human expert's capability for considering large numbers of special cases including commonsense knowledge obtained through many years of experience and education. If the crucial key to obtaining the human expert-like performance is in the codification of this huge amount of heuristic knowledge, there exists a fundamental barrier to this approach. The heuristic knowledge is inherently difficult to codify into the system, and the amount of knowledge easily exceeds the manageable range of a computer. They also point out that most of the optimistic reports regarding the early expert system examples are misleading, in the sense that, the status of the systems are in reality far

PAGE 93

83 from routine applications. A system's high performance, if any, is not due to machine intelligence and A.I., but to the computer's other superior aspects over human beings in handling data-intensive or calculation-intensive portions of the total task. This is exemplified in the Internist and the Dendral systems. It is also noticeable that no other systems ever attracted more attention than those early application examples mentioned above, implying a difficulty in expert system approach for routine applications. However, there is unfortunately very little that can be offered as a constructive counter-proposal, for solving heuristic problems by a computer with a humanlike level of performance and intelligence, that is not somewhat vague and hand-waving. It is more realistic to accept the A.I. and expert system approach as a new way to solve different aspects of a problem in which conventional approaches cannot offer a suitable solution. Selection of an appropriate domain, then, will be a significant issue in successful applications of the expert system approach. Considering the present status of A.I. and expert system technology, it is more reasonable to select a domain such that the expert system approach can give a complementary contribution to solving the heuristic portion of the whole task. The whole task still involves applications of conventional engineering or computer technologies and/or requires computer's superior calculating power to handle a fair amount of data-intensive

PAGE 94

89 and calculation-intensive portions of the tasks. The problem domain also needs to be selected by taking into consideration the amount of knowledge to be incorporated into the system such that it does not exceed the manageable range of the computer. It is also better to avoid the problems if a validation from the domain professionals is a significant issue, thus it is difficult to be accepted by the domain professionals because of the social, ethical, cultural, and other complicated problems, as shown in the medical diagnostic expert systems.

PAGE 95

CHAPTER IV KNOWLEDGE-BASED TOKEN PROCESSING SYSTEM Sleep EEG Analyzing Domain And Its Design Considerations Sleep EEG Analyzing Domain Human sleep EEG/EOG/EMG data are analyzed by a human scorer through the visual perception and interpretation of the multi-channel data on a polygraph chart. This process is best interpreted as primarily a pattern matching and classification based on heuristically obtained knowledge in the form of templates for the waveforms and sleep stages. The sleep data analysis, thus, falls into an application domain category of data interpretation by classification. The expert classifies a segment of the record into one of the five sleep stages plus awake by interpreting the data according to the perceived information, within the epoch, and other contextual information. In the process of visual scoring, the human expert perceives the occurences of waveform activities, such as alpha, beta, delta, sigma, theta, muscle artifact, rapid eye movement, etc., by a gestalt observation of the multi-channel sleep EEG/EOG/EMG data on a polygraph chart. Most of the patterns and templates do not readily exist in a fixed and definable form, since each segment of 90

PAGE 96

91 the record must be interpreted in a different context according to the individual subject's signal characteristics, several adjacent epoch patterns, and other higher order contextual situations which may affect the sleep stage scoring. Templates have to be articulated and extracted from the equivalent ones that are in the form of the expert's heuristic knowledge which comprises all the contextual data interpretations. The expert's knowledge is obtained by an intensive training process and is gestaltively applied to the visual inspection of the record. In 1968, Rechtschaf f en and Kales created a set of sleep stage scoring criteria that provide a standard reference for sleep stage scoring among researchers [Re68a] . A brief summary of these criteria is listed below. Stage Wake (Stage 0): This stage corresponds to the waking state. It is characterized by alpha activity and/or a low voltage, mixed frequency EEC. Certain subjects may have a virtually continuous alpha activity; other subjects may show little or no alpha activitiy in the record. This stage is usually accompanied by a relatively high tonic EMG, and often REMs and eye blinks are present in the EOG tracing. Stage 1: This stage is defined by a relatively low voltage, mixed frequency EEG with a noticeable theta activity. Stage 1 occurs most often in the transition from wakefulness to the other sleep stages or following body

PAGE 97

92 movements during sleep in normal subjects. During nocturnal sleep. Stage 1 tends to be relatively short, ranging from 1 to 7 min. Scoring of Stage 1 requires an absence of clearly defined K-complexes and sigma spindles. Stage 1, especially following wakefulness, is characterized by the presence of slow eye movements, each of several seconds duration, which are usually most prominent during the early portions of the stage. Rapid eye movements are absent. Tonic EMG levels are usually below those of relaxed wakefulness. The transition from a high-alpha subject wake state to Stage 1 is characterized by a decrease in the amount of alpha activity. Fifty percent (50 %) of the epoch is used as the boundary for separation of Stage 0 and Stage 1. Stage 2: This stage is defined by the presence of Sigma spindles and/or K-complexes and an absence of sufficient delta activity to define the presence of Stage 3 and 4. If less than 3 minutes of the record which would ordinarily meet the requirements for Stage 1 intervene between two epochs of Stage 2, these intervening epochs are to be scored Stage 2, if there is no indication of movement arousal or pronounced increase in muscle tone during the interval in question. Stage 3: Stage 3 is defined by an EEG record in which at least 20 % but not more than 50 % of the epoch consists of delta activity.

PAGE 98

mo] 93 Stage 4: Stage 4 is defined by an EEG record in which )re than 50 % of the epoch consists of delta activity. Intervals between delta waves rarely persist for more than a few seconds in Stage 4, but are usually prominent in Stage 3 epochs. Sigma spindles may or may not be present in Stage 4. Stage REM: Stage REM is defined by the concomitant appearance of relatively low voltage, mixed frequency EEG activity and episodic REMs. The EEG pattern resembles the one described for Stage 1, except the vertex sharp waves are not prominent in Stage REM. Alpha activity is usually somewhat more prominent in Stage REM than in Stage 1 . As with the EEG of Stage 1, there is an absolute absence of Sigma spindles and K-complexes. Stage REM should not be scored in the presence of a relatively elevated mental -submental EMG. These criteria define the patterns of each sleep stage in terms of the number of occurrences of specific waveforms, total time of specific waveform activity in an epoch, and/or the amplitude level description of the record [Ag72, Re68b]. For some waveforms such as sigma spindle, K-complex, and REM, the number of occurrences is used in the criteria. For the other waveforms the total running time of waveform activity within an epoch determines the criteria. The Rechtschaffen and Kales criteria, however, are a minimum set of quantitative specifications which can be used only as

PAGE 99

94 a general guide for sleep stage scoring and training of human scorers. The actual sleep stage scoring is performed by incorporating much more heuristic knowledge for the contextual interpretation. The actual EEG data processing is based on a gestalt perception of waveform activities together with a heuristic interpretation of the record and is not based on a precise and analytic measurement of the data. Design Considerations Several design considerations for developing an expert system for the automated sleep EEG/EOG/EMG signal analysis are discussed in the following. For sleep stage scoring, the on-line data processing aspect of an automated system is a distinguishing factor from other contemporary expert systems. Other related special design considerations like reasoning and processing efficiency, architectural effectiveness, reasoning with uncertainty, and adaptiveness of the system are discussed. To date, expert system design is limited to a few knowledge representations and inferencing methodologies which are exemplified through the earlier applications in the medical domain, chemical structure elucidation domain, and other engineering or scientific application domains [Ba82a, Ba82b, Co81], which are reviewed in Chapter III. If-then production rules provide a very efficient way of

PAGE 100

95 representing knowledge in the application domains where judgemental knowledge constitutes a major portion of the expert domain knowledge [Bu84, Ha85a] . Semantic network and frame based knowledge representation [Fi85] provide a more structured way of representing knowledge when the domain knowledge includes sophisticated interrelationships or hierarchical dependencies among diverse subparts of the knowledge. Most of the current inferencing methodologies are based on a data-driven forward chaining scheme, a goal-driven backward chaining scheme, or other opportunistic methods [Ha79, Ni86] which combine the two basic schemes resulting in improved search efficiency. Several expert system development tools have been also developed to provide a generalized way of expert system development by incorporating these basic knowledge representations and inferencing methodologies. Examples of these tools are EMYCIN, ROSIE, KEE, RLL, AGE, etc., [Wa86] . However, most of these readily available tools bring restrictive flexibilities in system design; thus, applications of these tools are generally confined only to certain limited application domains. In general, the development of an expert system can not totally depend on any of the previously developed design methodologies; it needs to be carefully designed by considering all the problem specific natures and sometimes by employing new representation schemes or inferencing methodologies.

PAGE 101

96 In designing an expert system, several complications may arise. The diversity of domain knowledge is one of the most typical problems which greatly affect the design and applicability of the system. A domain's knowledge is heterogeneous when it consists of several different subtasks combined in complex relations and/or when it employs several layers in terms of depth and fineness of knowledge. The Internist system demonstrates a good example of the knowledge diversity which results from the various subtasks in the domain's knowledge [Po77] . The Crysalis system also exemplifies knowledge diversity related to the wide range of depth and fineness of the domain's knowledge [En79] . The knowledge diversity is, in most cases, handled by the idea of partitioning of knowledge sources, hierarchical construction of the domain knowledge, or a combined scheme of these. The partitioning of the domain knowledge is exemplified in the following systems: Internist, Dendral [Li80], Casnet [We78b] , and Prospector [Du79] . The systems Crysalis, Hearsay [ErSOb], and Casnet are cases of handling the knowledge diversity problem using hierarchical construction of the domain knowledge. Recently, these concepts have been more generalized and referred to as the blackboard structure [Ha79, Ha85b, Ni86] where all levels of knowledge sources and their temporary results are stored and are used as a global data structure for inferring a particular hypothesis.

PAGE 102

97 The sleep EEG analyzing domain shows the problem of knowledge diversity which is associated with the existence of different processing layers when the human scorer's analyzing process is modeled in an expert system. Each processing layer is associated with a different processing time-frame and with a different type of processing knowledge. This layered hierarchy of the domain knowledge calls for designing the system in a layered blackboard model architecture . The sleep EEG analyzing process is broken down into three different layers of hierarchy; recognition of waveform activities at the bottom level; template matching at the next level where each record segment is classified according to the perceived information in the epoch; and contextual interpretation at the highest level of processing. This layered hierarchy provides the basic framework for handling the diverse knowledge sources, increasing the processing efficiency for on-line data processing. The human scorer's visual scoring process is simulated as a single-pass processing for the epoch-wise classification of the entire night's data in a strict bottom-up fashion on the blackboard model hierarchy. in the design of a sleep EEG analyzing expert system, the single-pass processing capability is emphasized for the online processing aspect of the system. However, the human scorer's visual scoring process involves a multi-pass

PAGE 103

98 inspection of data within a certain length of record to analyze an epoch in the context of several adjacent epochs. The human expert's localized multi-pass inspection is reflected in the design by using higher level abstracted data in a sliding window of five consecutive epochs. This single-pass processing model features a noticeable difference between the sleep EEG domain and most other application domains, where a multi-pass problem solving structure is generally important. In these domains, initial or partial solutions are continuously refined further by interactive new inputs from the user and/or by other relevant findings from one of the knowledge sources until a set of solutions for the problem is obtained. The system recognizes and counts the time or number of waveform activities in each 60 second epoch. The waveforms are recognized at the early-processing system as described in chapter II. The number of waveforms is counted for Sigma spindles, K-complexes, and rapid eye movements. The total running time of waveform activity is obtained for alpha spindles, beta spindles, delta, theta spindles, muscle artifact, and slow eye movements. The EMG is described by an amplitude level as high, medium, or low. In obtaining the total running time of these waveform activities within an epoch, an appropriate smoothing should be applied to get a close agreement with the human expert, since smoothing is implicitly employed in the process of the waveform activity

PAGE 104

99 time estimation by the human scorer. In the system, if a waveform activity contains an interval of less than one second without the waveform activity, the interval is considered part of the waveform activity running time. The early-processing always involves the possibility of an error due to the wide variabilities of signal characteristics. Since the detection reliability depends upon the detctor and is different among the detectors, each waveform detector output is associated with one of the three reliability factors, i.e.. High, Medium, and Low. Presently, the alpha and muscle artifact are assigned a high reliability factor, the beta, delta, sigma, SEM, and REM are assigned the medium level reliability factor, and theta activity and Kcomplexes are assigned a low level reliability factor. These reliability factors can be assigned differently through the system's knowledge-base editor. Each record segment (usually 30 or 60 seconds) is classified into one of six stages based on the perceived information in the epoch using the template matching rules. These rules comprise the Rechtschaf f en and Kales sleep stage scoring criteria. The efficiency of an inferencing scheme is closely related to the number of possible solutions. If the size of this solution space is too large, it is almost impossible to find a solution by trying all the possible hypotheses. Examples of this have been shown in chemical molecular elucidation or synthesis problems, and other such

PAGE 105

100 designs and configuration applications. In this case, the data-driven plan-generate-test scheme [Ba82b, Ha83b] is used as a search and pruning scheme to locate the possible answer efficiently as shown in Dendral and Crysalis systems. On the other hand, in the medical domain, an omission of possible hypothesis may create a critical error in terms of diagnosis. Thus, an exhaustive search by attempting all the possible hypotheses is appealing. The Mycin system does the exhaustive search based on the goal-driven backward chaining inferencing. The sleep EEG domain also has a finite set of goal hypotheses and an exhaustive search scheme can be efficiently applied based on the backward chaining inferencing. However, an efficient time-performance of the system is crucial in the design of an on-line processing system. Thus, a dynamic scheduling of searching sequences is devised as a scheme to improve reasoning efficiency using the high level contextual information of the sliding window in the blackboard hierarchy. The dynamic rescheduling of search sequences is most effective in improving reasoning efficiency when it is applied to a higher level knowledge. The overhead when a rescheduling is applied to a lower level knowledge, on the other hand, reduces the efficiency of the overall system, because the search space size is generally much larger for the lower level knowledge. At the top level, the human scorer's scanning process is modeled with a sliding window included in the system.

PAGE 106

101 The human scorer scans the record epoch-by-epoch for an epoch-wise classification. In the epoch-wise classification, the scorer often refers to the information of preceeding and succeeding epochs to classify an epoch in a context of the multiple epochs. The sliding window includes the global features, of several adjacent epochs, for the high level context interpretation. Five consecutive epochs' stage classifications and associated certainty levels constitute the high level abstracted information of the sliding window. These data reflect most of the context information that gives significant influences to the classification of each epoch. Each epoch's stage classification, which is first derived at the intermediate level by the template matching and classification knowledge, can be re-examined in a broader context provided by the sliding window. The window slides epoch-by-epoch representing a scanning procedure similar to that of the human scorer. This sliding window constitutes the highest level information plane of the blackboard and provides a distinctive structural feature resulting with an improvement in processing efficiency. Incompleteness and uncertainty of data, and the lack of information and^ knowledge are inherent problems in many application domains. Appropriate representation and interpretation of these uncertainties are intrinsically difficult problems [Co85] . Most of the methods developed to

PAGE 107

102 date rely on some intuitive and empirical treatment of numerical values using either a probablistic or a non-probablistic approach [Ro72, Ad76, Du79]. These models are designed with an emphasis on end-results in achieving a certain level of performance. A systematic management of the uncertainty problem is also emphasized in designing the models. On the other hand, they lack a theoretical basis and contain certain shortcomings, creating limitations in the application of the models as a general approach. The Bayesian conditional probability theory provides a reasonable groundwork for the treatment of uncertainty, but the assumption of the exclusiveness and the requirement of thorough a priori knowledge make it non-realistic for many applications. In Mycin system, the Certainty Factor (CF) model is developed based on the conceptual framework of confirmation and disconf irmation to represent the level of an expert's belief within a piece of knowledge [Sh75]. Empirical combination functions are devised to obtain a certainty level from more than one evidence. A certain level of threshold (0.2) is used when judging the success of a rule execution in the inferencing step. The Mycin's CF model provides a good systematic approach to overcome several shortcomings of Bayesian conditional probability theory. The Mycin's CF approach does not require the statistical data, but relies on the approximated new variables which

PAGE 108

103 convey an idea of confirmation and disconf irmation of the expert. This technique is not exact, but since the conditional probabilities reflect judgemental knowledge, which is highly subjective, a rigorous application of Bayes' theorem will not necessarily produce accurate cumulative probabilities either. However, the definitions of the CF and its manipulation schema can not provide a concrete theoretical basis, since the technique and its validation are totally based on empirical judgement. The application of the CF model involves several limitations of its own. The suitability of the CF scheme is thus very dependent on the nature of the problem domain. Limitations of the Mycin-like CF model have been well exemplified in the previous research [Ga86] of our laboratory where, the Dempster-Shaf er theory, which provides a generallized model for the Mycin certainty factor scheme, was applied to the sleep EEC scoring problem. The sleep EEC scoring problem was modeled as a finite automata machine and the Dempster-Shafer model provided a numeric measure for the transition mechanism. It was exemplified that the numeric values (certainty factors) and their operations do not provide a sufficient measure of the transition mechanism for a correct sleep stage derivation. In the sleep EEC scoring problem, information from several adjacent epochs significantly influences the classification of an epoch's stage. Thus, the classification of an epoch cannot totally

PAGE 109

104 depend on the transition scheme which is based on the comparison of numeric values. The values reflect the levels of an epoch's matching predefined sleep stages templates. It is necessary to provide a different representation and combining scheme of the certainty factor to provide an appropriate measure for the contextual interpretation. In this research, a new model is designed for the management of uncertainty and incompleteness. The model is designed with a pragmatic reason to overcome several shortcomings of the prior certainty factor models and Bayesian conditional probability theory models which appear when these models are directly applied to the EEG analysis domain. The new uncertainty handling model uses linguistic variables, i.e.. High (H), Medium (M), and Low (L), instead of numeric values. The human expert describes his confidence in data and knowledge by associating it with one of the discrete certainty variables. In this way, the THEN part of each rule is associated with one or more of the certainty variables according to the number of action elements in the THEN part. As an example, ALPHA is described as High-Activity when the sum of alpha waveform running time exceeds 25 seconds within the epoch. However, the certainty level is determined according to the following f uzzif ication scheme; the certainty variable is H if the activity time exceeds 35 seconds, M if the activity

PAGE 110

105 time is between 30 and 35 second, and L if the activity time is between 25 and 30 seconds. These linguistic variables reasonably reflect the human expert's belief level in a natural way and provide a conceivable method of handling uncertainties . A new combination scheme is defined with the discrete certainty variables based on the conceptual framework of balancing the weight with the certainty levels. The detailed description of the combination scheme and its operation are shown with an example in the next section on design and implementation. The newly defined combination scheme and the discrete certainty variables provide a kind of robustness in inferencing steps, preventing propagation of minor variations of certainty levels. In this scheme, the certainty level of an assertion changes only when the combined level of certainties in evidence exceeds a certain level. This combining scheme and the linguistic variables also provide an appropriate way to represent the uncertainties in a form of conceivable template. Since the uncertainty is abstracted and represented as a part of the conceivable template, this uncertainty handling model provides an efficient way to implement the contextual manipulations in the conceptual framework of template matching. The human scorer's sleep EEG analysis is regarded as a process of template matching. The variations in the data

PAGE 111

106 and the lack of information result in ambiguities when an epoch is matched against any of the templates. The ambiguous epoch is understood in terms of the context of several adjacent epochs by the scorer. This scoring process is robust over the minor variations of data in the record. This overall processing, when implemented in a computer, calls for an uncertainty handling model that must reflect the framework of the template matchings for the classifications and contextual interpretations. The comparison of the numerical values of relative likelihood (CF) as implemented in the Mycin system does not represent properly the uncertainty handling of sleep EEG processing, where template matchings and strong contextual reexaminations play a main role in the processing. The operation of the new uncertainty handling model does not rely on analytic number operations. The fact that, an expert's gestalt scanning results in a better performance than a rigorous analysis done by a novice, illustrates that the superiority and efficiency of an expert comes from the robustness of the expert's knowledge for interpreting an uncertain epoch in a context, not from any capability for numerical operations.

PAGE 112

107Design And Implementation Functional Description of the Overall System The entire night's data of recognized waveforms is acquired on-line at the host computer and analyzed by the token processing system for sleep stage scoring. The token processing system is implemented in Common Lisp on the DEC Rainbow PC. The overall functional block diagram of the knowledge-based token processing system is shown in Fig. 4.1, and each functional block is briefly described in the following. Major structural components of the system are a data base, a knowledge base, a knowledge engine (control modules), a knowledge-base editor, and other interfaces for an explanation mechanism as is shown in Fig. 4.1. The data base consists of two parts, a static data base and a dynamic data base. The static data base contains the night ' s token data received from the early-processing system. The dynamic data base contains three layers of temporary data planes, in a blackboard model architecture, which are continuously updated during the system epoch-wise classification of the record. The knowledge base consists of three different layers of knowledge which are each associated with a corresponding data plane in the blackboard. The three layers are a parametric feature extraction knowledge at the bottom level, template matching and classification knowledge at the next

PAGE 113

o c 3 (O > o >> 00 CO (/)
PAGE 114

109 level, and contextual interpretation knowledge at the top level of the knowledge base. The knowledge is represented in if-then production rules effectively reflecting most of the domain knowledge. The primary role of the knowledge engine, which is composed of rule-interpreters and a scheduler, is to derive an epoch's sleep stage using the rules contained in the knowledge base and corresponding data in the blackboard data planes. The scheduler performs a global control over the knowledge base and data planes. It also performs a scheduling of search sequences of the intermediate level inference engine according to the high level contextual information in a sliding window with an improvement in searching efficiency of the system. Explanations about the epoch classification can be explored at any depth and any range of related information in the system. Besides this on-line explanation facility, the system also provides an off-line explanation feature encoded with a series of rule identification codes, which can be used after the system has finished processing the night's data. The expert system development includes the development of an expert system shell with a knowledge-base editor. The knowledge-base editor allows the creation and modification of the knowledge base without any modification of the program by the user. The knowledge-base editor, which is closely related to the knowledge representation

PAGE 115

110 scheme and the system's inferencing mechanism, is designed considering the flexibility for future extension or modification of the system as is described in the next section. A menu-driven user-friendly interface is designed for the editor. Structures of Knowledge Base and Data Base The system's knowledge base is structured in three layers of hierarchy consisting of three different levels of knowledge: the parametric feature extracting knowledge, the template matching and classification knowledge, and the contextual interpretation knowledge. Each level knowledge is refered to as a knowledge source (K.S.) in the layered knowledge base. The knowledge base, thus, consists of the three knowledge sources: parametric level K.S., epoch level K.S., and context level K.S. as illustrated in Fig. 4.2. The blackboard structure of the data base corresponds to the hierarchical organization of the knowledge base. The blackboard contains all the input data and intermediate solution states produced by the knowledge sources; it is hierarchically layered according to the level of analysis. The knowledge sources use this blackboard as a global data base to derive further solutions. The domain knowledge is effectively represented by if-then production rules, since most of the domain knowledge is in the form of judgemental knowledge. Each level

PAGE 116

Ill Contextual Data Plane Context Level K.S. o sliding window o contextual interpretaion o searching path o meta-rules for scheduling Epoch Level K.S. i6inpxaT.es r'xane o wave-activity descriptors o wave-activity descriptors o classification templates generation rules o classification rules Parametric Feature Plane Parametric Level K.S. o waveform occurrence data o procedures to extract o early-processing system parameters reliability factors o queries for user input o user input data o query parameter data Blackboard Knowledge Base Fig. 4.2. The Blackboard Model Hierarchy of The System.

PAGE 117

112 knowledge is associated with a corresponding ruleinterpreter (inference engine). The scheduler globally arranges the rule-interpreters. The layered structures of the knowledge base and blackboard are illustrated in Fig. 4.2. The parametric level knowledge includes procedures which extract parametric data from the user and from the multichannel record. The set of parameter values which constitutes the bottom level parametric data plane includes waveform occurrence values within an epoch, waveform detection reliability factors of the early-processing system, query parameter values, and user input data like sex, age, record number, and date. These parameter values constitute the bottom level of the blackboard. Sample parameters from the parametric feature plane are illustrated in Fig. 4.3. For all the waveform occurence parameters, such as alpha-time, beta-time, sigmacount, etc., the waveform activity time or number of occurence in an epoch is assigned together with the waveform detection reliability factor. The early-processing detection reliability factors are stored in the system knowledge base and can be modified through the knowledgebase editor. The query parameters, such as Subject-alpha, Under-drug, etc., are variables which reflect a priori knowledge of subject dependent characteristics. These parameters can be incorporated in the system's scoring rules according to the purpose. Any number of query parameters.

PAGE 118

113 Alpha-time Delta-time Sigma-count SEM-time Kcom-count Subject-beta Alpha-reliability Delta-reliability Theta-reliability SEM-reliability Kcom-reliability Beta-time Theta-time Muscle-time REM-count Subject-alpha Under-drug Beta-reliability Sigma-reliability Muscle-reliability REM-reliability Fig. 4.3. Sample Parameters of the Parametric Feature Plane. WAKE-W-ACTIVITY ALPHA-ACTIVITY DELTA-ACTIVITY THETA-ACTIVITY SEM-ACTIVITY KCOM-ACTIVITY SLEEP-W-ACTIVITY BETA-ACTIVITY SIGMA-ACTIVITY MUSCLE-ACTIVITY REM-ACTIVITY Fig. 4.4. An Example List of Wave-activity Descriptors.

PAGE 119

114 which can also be modified through the knowledge-base editor, can be defined in advance in the knowledge base with a default value and an associated certainty level. Before the system starts epoch-by-epoch classification of the entire night's data, the system displays the default value of each query parameter, and asks the user if he wants to change the value. In this way, all the a priori knowledge of the subject can be incorporated into the system. Since, these values are input to the system only at the beginning of processing, the system can run without any further user's intervention . The intermediate level knowledge represents the human expert's process of template matching for the classification of each record segment. Templates are mostly articulated and extracted from the human expert's heuristic knowledge and are defined in the form of a production rule. Combined information of several wave-activity descriptors represents a template. This level knowledge also includes the rules which derive the wave-activity descriptors of an epoch using the waveform occurence data in the parametric feature plane of the bottom layer. A sample list of wave-activity descriptors and template matching rules are illustrated in Fig. 4.4 and 4.5, respectively. A rule consists of a rule identification code, a premise, an action, an author, a date, and a justification. The rules are stored internally in a Lisp code and consist

PAGE 120

115 RULE-1: G-14 Premise: ( ( EQ ALPHA-ACTIVITY MED ) ( EQ MUSCLE-ACTIVITY MED) ) Action: ( WAKE-W-ACTIVITY HIGH L) Author: (TAE GYU CHANG) Date: (8-5-1986) Justification: If an epoch is described as both medium ALPHA-ACTIVITY and medium AROUSAL-ACTIVITY, then the epoch is described as high WAKE-W-ACTIVITY. RULE-1: G-44 Premise: ((GT MINUTE 50) (NOT REM LOW)(EQ SLEEP-W-ACTIVITY LOW) (NOT WAKE-W-ACTIVITY HIGH)(EQ DELTA-ACTIVITY LOW) ) Action: (STAGE STAGE5 HIGH) Author: (TAE GYU CHANG) Date: (8 31 1986) Justification: If an epoch after 50 minutes from the begining of the record shows REM, low SLEEP-W-ACTIVITY low DELTA-ACTIVITY, and not high WAKE-W-ACTIVITY, then the epoch is scored as STAGES. Fig. 4.5. Examples of Rules for Classification.

PAGE 121

116 of the following properties. 1) RULE-ID: An atom in a form of a code. 2) PREMISE: A list in a form of (operator parameter value) or in a form of more than one of the three-tuple as ((operator parameter value) (. . .) . .). The predicates of GT, LT, EQ, and NOT constitute the operator set. 3) ACTION: A list in the form of (parameter value certainty-factor) or a list containing more than one of this basic threetuple form. 4) AUTHOR: A list containing the name of the author. 5) DATE: A list in a form of (month date year) 6) JUSTIFICATION: A list containing an English version of the translation to be used in the explanation mechanism. The top level knowledge source includes contextual manipulation rules for the re-examination of an epoch's stage classification and search sequence scheduling rules. These rules use the contextual information in the sliding window. The sliding window contains the information of five consecutive epochs to reflect a local context information in the recordThis information contains epoch sleep stages, derived with the intermediate level template matching knowledge source, and associated certainty levels of the

PAGE 122

117 epoch stages. The contextual manipulation is performed by the process of matching the sliding window with a set of templates in the form of if-then production rules contained in the context level knowledge source. The contextual interpretation rules continuously update the sliding window, possibly more than once each time the window moves one epoch. The process of matching is tried sequentially with the set of templates from the first one to the last one. Whenever a template matches with the sliding window, the window is updated by the rule and the process of matching starts again from the first template. This process stops when there are no more rules which match with the window. The sliding window will then slide one epoch position and receive a new epoch's classification stage and a certainty level from the intermediate level processing, for the next cycle of matching process. If the template size of a contextual manipulation rule is less than five epochs, the template slides through the window from left to right (equivalently from the oldest epoch to the most recent epoch) continuously updating the sliding window whenever a matching occurs. This process cycles again if any matching occurs in the past cycle and stops if no matching occurs in the past cycle. This higher level contextual interpretation capability is especially crucial for a domain like sleep EEG analysis where human experts also perform a lot of

PAGE 123

118 re-examinations in a context of several adjacent epochs' information. The inherent nature of wide variabilities associated with the sleep EEG characteristics adds to the requirement of the contextual manipulation of the data. The information in a sliding window also provides for a heuristic search by rescheduling the searching path of the top-level goal hypotheses, i.e., stages 0 to 5, for the intermediate level template matching and classification knowledge. According to the context information reflected in a sliding window, a reasoning path for the next epoch is rescheduled in the order of the likelihood of each top-level hypothesis. As an example, if the previous record segments show a running streak of stage two, the most likey stage for the next epoch is also stage two and the rest five stages are further arranged in the order of their likelihood, e.g., stage three, one, five, zero, and four. The scheduling is also performed by the template matching of the sliding window against the scheduling rules in the context level knowledge source. The scheduling is performed at the end of every epoch's processing. The matching is tried sequentially on the set of scheduling rules. When a scheduling rule matches with the winow, the new search path information is stored into the contextual data plane. This rescheduling provides a searching efficiency reducing the processing time without any significant overhead of the scheduling process. This searching mechanism feature is

PAGE 124

119 distinctively different from the exhaustive searching method used in the Mycin-like medicine consulting programsIn Fig 4.6, an example of a sliding window is illustrated together with a contextual smoothing rule and a scheduling rule. The top level knowledge base is described by four global variables. These variables allow an access to all of the contained knowledge. These variables are as follows: 1) DOMAIN: The value of this variable is a phrase which describes the knowledge base contained in the system. This phrase is displayed whenever the execution of the system is started. 2) CONTEXT: This variable is a list of atoms, here "list" and "atom" refer to the objects of Lisp programming language, which are the names of the contexts. The knowledge-base editor allows the user to input more than one context name. However, this feature is provided only for the flexibilty of the knowledge-base editor. This system design is based on the hierarchy of the knowledge sources and the blackboard structure. Search spaces are grouped and searching efficiency is obtained by a dynamic rescheduling. Thus, one context is used in the system. 3) PARMGROUPS: This is a list of atoms where each atom is a name of a group of associated

PAGE 125

120 WINDOW WINDOW-PAD s-1 s-2 s-2 s-2 s-2 CERTAINTY-PAD M H H M H Epoch Timeframe 68 Searching Path s-1 s-3 s-4 s-0 s-5 RULE-3 : C-12 Premise: (EQ WINDOW (s-4 H s-3 L s-3 L s-3 L s-4 H)) Action: (WINDOW (s-4 H s-4 L s-4 L s-4 L s-4 H) M) Author: (TAE GYU CHAWNG) Date: (9 7 1986) Justification: If three STAGE 3 epochs which have an ambiguity, equivalently the lowest certainty level, intervene between two clear STAGE 4 epochs, then the intervening epochs are smoothed into STAGE 4 with the lowest certainty level. RULE-2: p-2 Premise: (EQ WINDOW (******** s-2 *)) Action: (PATH (s-2 s-3 s-1 s-5 s-0 s-4) *) Author: (TAE GYU CHANG) Date: (9 7 1986) Justification: If the sleep stage of the last epoch in a running window is s-2, then the searching path for the goal hypotheses is in the order of stage 2, 3, 1 5 o ig. 4.6 An Example of a Running Window (top). Contextual Manipulation Rule (middle) and Scheduling Rule (bottom).

PAGE 126

121 parameters. Also here, the knowledge-base editor allows the user to enter more than one parameter group. 4) RULEGROUPS: A list of atoms where each atom is the name of a group of associated rules. The knowledge-base editor accepts the definition of more than one rule group. A context is described by the following seven properties. 1) TRANS: A text containing an English phrase describing what problem the system is attempting to solve. 2) DISPLAYPROMPT: A list containing an introductory phrase which is displayed directly below the value of DOMAIN whenever the system starts to execute. 3) TEXTS: This is an atom which has as its value the set of names of text lists which are associated with the context. 4) PARMS: This is an atom which has as its value the set of names of parameters associated with the context. 5) CREATEDATE: A list containing the date on which the context is created. This list is of the form (A B C) where A represents the month, B represents the date, and C represents the year.

PAGE 127

122 This date is also used as a default date for the rule insertion or edition. 6) AUTHOR: This is a list containing the full name of the creator of this context. This author name is also used as a default author name during rule insertion or edition. 7) GOALS: A list of parameters for which the system is trying to find values. In this system, one goal parameter exist, i.e., STAGE. A parameter consists of the following properties. 1) TRANS: A list containing an English phrase which is a translation for the parameter. This is used in the explanation mechanism. 2) NEGTRANS: A list containing an English phrase which is a translation for this parameter not being present. This property is associated only with YES/NO type parameters. I) PROMPT: A list containing an English phrase which is used to ask for a value for this parameter. ) DERIVEFIRST: An atom which is either T or NIL. It identifies whether this parameter has to be read from the data base at the bottom layer when it is needed (NIL) or whether the system should search first for a rule that can derive it (T). ) TYPEVALUE: An atom which specifies the type of

PAGE 128

123 the parameter value. 6) EXPECT: A list of expected values. This property is associated with parameters which have TYPEVALUE of ANY only. Not all the property values should be associated with each parameter. Examples of parameters are shown in Fig. 4.7. A text is a list of words which is associated with GOAL parameters. The text has a single property called TEXT which is a list of English words that is displayed if the corresponding rule fires. Inferencing and Control Mechanism Generally, in a blackboard modeling approach, each independent knowledge source is associated with a different reasoning and representation method according to its characteristics. Since most of the EEG domain knowledge can be effectively represented in a form of production rule, allowing a uniformity in the knowledge base structure, the if-then production rule based representation method is used. The parametric feature extraction knowledge at the bottom level includes the procedures which receive user query inputs and extract parameters from token data as is described before. The user query inputs are acquired into the system only at the beginning of the system run to allow a stand-alone running of the system without any

PAGE 129

124 Parameter Name: WAKE-W-ACTIVITY TRANS: The STAGE WAKE related EEG activities of the epoch, equivalently the combined level of activity of ALPHA and MUSCLE artifact, NEG-TRANS: NIL PROMPT: What is the level of STAGE WAKE related EEG activities of the epoch? DERIVEFIRST: T TYPEVALUE: ANY EXPECT: (HIGH MED LOW NONE) Parameter TRANS : Name: SLEEP-W-ACTIVITY The STAGE 2 related EEG activities of the epoch, equivalently the combined level of activity of SIGMA spindle and K-complexes. : NIL What is the level of STAGE 2 related EEG activities of the epoch ? DERIVEFIRST: T TYPEVALUE: ANY EXPECT: (HIGH MED LOW NONE) NEG-TRANS PROMPT : Fig. 4.7. The Example Property Lists for the Parameters of WAKE-W-ACTIVITY and SLEEP-W-ACTIVITY.

PAGE 130

125 intervention . The reasoning method for the intermediate level template matching is a goal-driven backward chaining scheme. The limited niimber of goal hypotheses, i.e., sleep stages, make it very efficient to apply a goal-driven backward reasoning inferencing scheme. It is implemented with the following procedures: 1. The system initially identifies the goal parameter STAGE as a goal for which it is attempting to find values. 2. The system first reads the hypotheses sequence from the hypotheses list. 3. The system then starts to verify those hypotheses sequentially, and it stops when one of the goal hypotheses is verified. 4. To verify a hypothesis, which also includes the goal hypothesis, the system first searches the fact list, which is the intermediate level data plane containing all the derived facts. A hypothesis is a list in a form of three-tuple, i.e., (operator parameter value), consisting a rule premise or a part of a rule premise. A fact in the fact list is a list in a form of three-tuple, i.e., (parameter value certaintylevel). If the hypothesis does not exist in the fact list then the system checks to see if the

PAGE 131

126 DERIVEFIRST property of the parameter of the hypothesis is T or NIL. If it is NIL, the system reads the value from the bottom layer data plane and stores the value in the fact list. If the DERIVEFIRST value is T then the system tries to derive it from other rules. 5. To derive a hypothesis from other rules, the system first gathers all the relevant rules which contain the hypothesis as the action part of the rule. The system then tries all rules to verify that hypothesis. If more than one rule succeeds to verify that hypothesis, then the certainty level is modified according to the combining scheme and is associated with the hypothesis in the fact list. 6. Step 4 and 5 are repeated recursively as many times as needed. 7. Finally the system reports one of the goal hypothesis it verifies with an associated certainty level of the goal hypothesis. If the system fails to verify any of the goal hypotheses in the list, it assigns the default hypothesis as the previous epoch's hypothesis with the lowest certainty level. The rule interpreter scheme of the highest level ledge source is in the form of data-driven template

PAGE 132

127 matching scheme as is described before. The large number of patterns in a window makes it infeasible to apply a goal-driven backward reasoning scheme. A scheduler performs the global control for the execution of the diverse knowledge sources with the dynamic data planes in the blackboard. In Fig. 4.8, the overall looping sequence of the scheduler is illustrated. The scheduler triggers the rule interpreters in a strict bottom-up fashion in the hierarchy of the knowledge base. At the beginning of a new epoch processing, it triggers a procedure to set up the dynamic data base of the blackboard with waveform occurence parameters. The sliding window is also updated by sliding one epoch. Then it triggers the rule interpreter of the intermediate level template matching knowledge. The rule interpreter performs the reasoning in a backward chaining scheme to derive a conclusion with a set of rules and information on the same or other level of the dynamic data base of the blackboard. All the derived conclusions and intermediate results are stored in the dynamic data base. The derived final conclusion is stored in the sliding window for the next level processing. The scheduler triggers the contextual smoothing rules and search sequence scheduling rules. Finally, the scheduler triggers a procedure to attatch the result of the epoch processing to the report at the end of the epoch processing.

PAGE 133

128 ± Data Dispatch Parametric Extrac Feature tion \ / Channel Descripors Generation & Matching N / Contextual Interpretation Scheduling Fig. 4.8. The Looping Sequence of The Scheduler.

PAGE 134

SI c 129 Reasoning with Uncertainty A well trained human scorer possesses the capability of processing the record in a robust and efficient way. The human expert can easily score prominent epochs which contain clear features of any sleep stage by perceiving the global pictures of the record. He, then, has the capability of interpreting ambiguous epochs in the context of the record. This robust processing capability is especially important nee the sleep EEG data include wide variations in haracteristics . The certainty factor model must be designed to reflect the above human expert's robust processing in interpreting ambiguous epochs in a context of the record. In the system, the lack of information in data and the uncertainties in related rules lead to a low confidence level in matching or even to a matching failure. Deduced information of an epoch, i.e., the matching result and its associated certainty level, is subject to the surrounding data context. Thus, the information within an epoch can be interpreted differently depending on the context. The role of the certainty level information is to provide a measure of matching closeness, which is then used in the next higher level contextual manipulation. The certainty factor model of the sleep EEG analyzing expert system includes the linguistic variables, i.e.. High, Medium, and Low. These represent the certainty level of a

PAGE 135

130 piece of knowledge and a corresponding weight value is associated with each variable as is shown in Fig. 4.9. Certainty Variable Weight Value High + 1 Medium 0 Low 1 Fig. 4.9. Certainty variables and associated Weight values. The certainty variables and their combination functions are defined based on the concept of weight balancing. A rule is stated with an associated certainty level in the action part based on the assumption that all the premise elements have a medium level of support. In other words, the certainty level of a rule premise is balanced under the assumption that each element of the rule premise is true with a Medium level of certainty. The action part certainty level is modified during rule execution according to the combined certainty level of all the premise elements. A combined weight (CWT) value of a rule premise is defined as follows. CWT = Wi N

PAGE 136

131 where CWT: the normalized combined weight W^: the weight value of the i-th condition in the rule premise N: total number of conditions in the rule premise + 1 Increase + 1/3 0 No change 1/3 1 Decrease CWT CF Modification Fig. 4.10. Certainty level combining scheme. The combining scheme is illustrated in Fig. 4.10. If more than one rule is executed, the certainty variable which has the highest certainty level is chosen for the action part of the rule. This scheme is reasonable because of the implied mutual exclusiveness of the templates. The following example illustrates the operational of the certainty factor scheme.

PAGE 137

132 IF: 1) The Sigma spindle activity is high, and 2) The delta wave activity is low, and 3 ) The wake wave activities are low THEN: The sleep stage of the epoch is STAGE 2 with a HIGH level of certainty. The above rule implies that, the sleep stage of an epoch would be scored as STAGE 2 with a high level of certainty under the assumption that all the conditions of the IF part of the rule are satisfied with a medium level of certainty. The certainty level for the epoch will be modified according to the combining scheme with the actual certainty level of each condition in the rule premise. If the epoch contains one Sigma spindle, five seconds of alpha, and one second of delta, then the associated certainty levels for each condition in the rule premise are LOW, LOW, and HIGH, respectively. CWT = (Wl + W2 + W3)/3 = (-1 1 + l)/3 = 1/3 Thus, the certainty level of the rule action part is modified by decreasing the level by one, i.e., from High to Medium.

PAGE 138

133 Explanation Mechanism The data processing nature of the problem requires special consideration in the explanation mechanism design. The explanation mechanism can be accessed any time, during system execution and also after the system has finished the processing. The system stores all the rule identification codes of the executed rules in each epoch in the order of the fired sequence. When an explanation is requested for a certain epoch, the system displays all the parameters associated with the fired rules in the epoch. The user, then, can trace the line of reasoning by selecting the parameters from the menu screen as is shown in appendix B. The system displays the explanation of how the system has derived the parameter, by displaying translated phrases with a list of the rule as is illustrated with an example in appendix A.

PAGE 139

CHAPTER V SYSTEM EVALUATION AND RESULTS The complete system must be developed and tested with a large amount of data from a wide range of subjects by incrementally incorporating relevant rules into the system's knowledge base. The waveform recognition parameters, which are mostly based on the previous waveform parameters of the SAHC system, were established by evaluating the system with a wide range of subject records. A total of 112 rules are included in the token processing system's knowledge base, and these rules are listed in Appendix B. Presently, most of the classification rules are based on the Rechtschaf f en and Kales sleep staging criteria. In order not to lose the objectivity of the analysis, wide deviations from the Rechtschaff en and Kales criteria are avoided as much as possible in this initial system evaluation. However, the developed system has the flexibility to allow modification and incorporation of more rules in the system through the knowledge-base editor. Potentially, there exists room for further performance improvement by elaborating on the system's knowledge base with a variety of contextual information. The developed system is tested with a selected set of 16 records of subjects from 5 to 70 years old. The number 134

PAGE 140

135 of subjects included in the present analysis is not sufficient to provide a thorough system performance validation. The validation is an involved process including processing of much more data from range of subjects. The system also requires a validation by professionals in sleep EEG research and its clinical application areas. Thus, the system development and validation require coorperation of the domain professionals as is generally true in other cases of expert system development. However, from the results of the 16 records, a general perspective for problems can be illustrated with discussions on the nature of errors, the system performance and its limitations, and ways of further improvements. Experimental Procedures This study selects a total of 16 records of normative subjects in three age groups. Group 1 contains seven records of six subjects between 5 to 18 years old. Group 2 contains five records of four subjects in the age range of 25 to 34 years old. Group 3 contains four records of four subjects from 47 to 70 years old. Entire night's EEG/EOG data have been recorded on a 1-inch Sangamo FM tape recorder Model 3500 at 15/16 inch/sec. These data have been replayed at the same speed for processing. Three EEG channels, i.e., F1-F7, C3-A2, and 03-OzPz, and one LE-A2 EOG channel are

PAGE 141

136 used in the analysis. The channel assignments for the waveform detectors are described in chapter II. The four channels are calibrated at the same level such that 50 [uv] EEG/EOG corresponds to 1.5 [v] for the input to the earlyprocessing system. A 10 Hz sinusoidal calibration signal is recorded at the beginning and end of each subject's EEG/EOG recording to indicate the 50 [uv] EEG level. Each channel signal is digitized at a 480 Hz sampling rate by a 12 bit A/D converter. All the channels are first lowpass filtered by a digital filter with a 3-db cutoff frequency at 120 Hz. All the waveform detection parameters, except the delta amplitude threshold level, are kept the same as described in chapter II for all the subjects' records. Since the delta amplitude level is generally much higher for a young subject group, a different delta amplitude threshold, i.e., 77.2 [uv], is used for all the subjects of 13 years old and younger. The same threshold of 16.7 [uv] is used for all the rest of the higher age subjects. The token data of all subjects processed from the early-processing system are online acquired with a PDP-11/23+ system and stored on RL02 disks. These data of 16 records constitute a data-bank which can also be used in the future for other types of token analysis purposes. The token data for one night generally occupies around 300 Kbytes. Minute-wise summarized token data files are sent to the Rainbow PC for the sleep stage scoring. The number of waveforms per minute

PAGE 142

137 is counted for sigma spindles, K-complexes, and REM waves. The total waveform activity time per minute is counted for alpha spindles, beta spindles, delta waves, theta spindles, muscle artifact, and slow eye movements. If a waveform activity contains an interval of less than one second without the waveform, the interval is considered part of the waveform activity time in the minute. Results The system performance is ultimately measured in terms of a man-machine agreement for the record scoring. The supplementary sleep staging manual of H.W. Agnew, Jr., and W.B. Webb [Ag72] is used as a referential guide for the system's performance evaluation. The manual provides instructions for signal recording, scoring criteria, and training procedures for human scorers. They obtained a set of calibration records, which includes six normative subjects in the range of 27 to 34 years old, by an initial scoring with at least 90 % agreement between two experienced scorers and by final adjustments among the laboratory personnels. The set was used as the standard calibration records in that laboratory to test a human scorer's performance level. According to the manual, which is also true in most cases for other laboratories, a 90 % agreement rate with the standard calibration records is required for

PAGE 143

138 the qualification of satisfactory scoring by a scorer. It is noticeable that the required agreement rate among experienced scorers is 90 %, thus, it is the desired reference for an automated system's performance evaluation. Average agreements, across the six stages, of four trained human scorers with the six standard calibration records are reported in the manual and are illustrated in Table 5.1. The system's processing results are compared with the independent human scorer's results obtained from the sleep research laboratory of the Baylor College of Medicine, Houston, Texas. Table 5.2 shows a list of man-machine average agreement for the records. Overall average agreement of the 16 records is 83.6 %. The four records, 10719, 10114, 10067, and 10889, show especially poor agreement of far below 80 %. On the other hand, all the other 12 records are relatively evenly distributed between 82.4 and 92.8 %, as is illustrated in Fig. 5.1, giving a total average of 87.2 %. it is exemplified from the distribution of subject percentage agreement that the scoring agreement is quite subject dependent. Table 5.3 shows the percentage agreement and the classification epochs' distribution across the sleep stages for the 16 records. The percentage agreement tables for each subject record are attatched in appendix C. The overall epoch-by-epoch percentage agreement is 83.6 %, which is obtained by dividing the diagonal sum of the epochs by

PAGE 144

139 Table 5.1. Percentage Agreement With The Standardized Scoring Across Six Calibration Records By Stage of Sleep And By Scorer [Ag72] . Sleep Stages Scorer Mean Across Scorers 1 2 3 4 0 94 93 96 98 95 1 79 83 80 79 80 2 95 95 94 92 94 3 32 35 59 70 49 4 65 91 89 90 84 5 96 93 97 99 96

PAGE 145

140 Table 5.2. List of Average Agreement for Each Record. Group Average A^r. (%) Age Records stages 3 and 4 together Group 1 5-18 10719 (5) 76.0 77.5 10150 (8) 83.5 91.1 10158 (13) RQ Q 10109 (13) 88.0 91.9 10146 (13) 84.4 93.8 10114 (13) 72.9 81.8 11623 (18) 91.5 98.0 Group 2 25 34 11769 (25) 89.0 93.6 11717 (27) 86.7 91.4 11747 (29) 82.4 90.7 10244 (34) 82.8 84.0 10256 (34) 92.6 94.0 Group 3 43 79 10067 (43) 72.4 72.4 11771 (53) 85.2 85.2 10889 (53) 72.4 72.4 11740 (70) 92.8 92.8 Total 83. 6 87.3

PAGE 146

141 (a) O OO — 1 — * 1 t 0 o o o o GOO OOO OO 70 80 ' ' 1 ' ' 90 — 1 100 (b) o o o H 1 i 1 1— I 0 0 o « 1 1 1—4 0 0 0 OOO OOO o 70 lo ' ' jc 100 (%) Fig. 5.1. Distribution of Percentage Man-Machine Agreement for the 16 Records, a) With separation of stages 3 and 4; b) without separation of stages 3 and 4.

PAGE 147

142 Table 5.3. Man-machine Agreement for the Total 16 Records (5 to 70 years old) . Machine Score (false negative) \stage staoes. 0 1 2 3 4 5 total agr. (%) 0 436 A n u X / 4o4 90 . 1 1 52 117 67 0 0 52 288 40.6 2 44 37 3425 130 3 169 3808 89.9 3 0 1 85 212 21 4 323 65.6 4 0 1 19 260 484 5 769 *89.4 62.9 5 79 45 113 7 0 1641 1885 87.1 total 611 237 3713 609 509 1878 7557 agr(%) 71.4 49.4 92.2 34.8 *84. 95.1 1 87.4 83.6 87.3 * no stages 3 and 4 discrimination. (false positive)

PAGE 148

143 the total sum of the epochs in the table. Each stage-wise agreement should be considered in terms of both false positive and negative errors. The stage-wise percentage agreement shown at the right most column reflects the false negative error for each sleep stage, if the agreement rate is substracted from 100 %. Similarly, the agreement rate of each stage shown at the bottom row of the table also reflects the false positive errors, if the agreement rate is substracted from 100 %. Stage 2 scoring shows about 90 % agreement in both directions resulting in around 10 % of false negative and positive errors. However, there exists room for further improvement of Stage 2 scoring by allowing changes in the Sigma spindle detection parameters and filter structure mainly according to a subject's age group. Since, Stage 2 epochs occupy a significant portion of the total epochs, further increase in Stage 2 scoring will give a significant improvement in terms of total system average agreement. The REM stage (Stage 5) scoring shows also a relatively high agreement above 87 % in both directions resulting in less than 13 % of false negative and positive errors. Thus, the scoring of stages 2 and 5, which together occupy three fourths (3/4) of the total record, show relatively higher man-machine agreements than the scoring of the other stages. It is significant that the REM sleep stage becomes one of the most reliably scored stages. The

PAGE 149

144 accuracy of REM stage scoring is relatively subject dependent upon the variations of EEG characteristics. Most of the REM stage scoring errors result from the mis-scoring with Stage 2. For some subjects, REM waves appear abnormally together with sigma spindles causing trouble in separating the REM sleep and Stage 2. In this study, this type error gives a decrease of 3.7 % in the total average agreement. The portion of 61.3 % of these errors results from the three poorly scored records, i.e., 10067, 10114, and 10719. If an EMG channel is incorporated in the analysis, the REM sleep scoring can be improved further especially resulting in a better separation of REM sleep and wake. The REM detection algorithm sometimes gives a false positive REM detection during NREM sleep requiring a further elaboration. However, this does not cause serious trouble since the false REM detections do not occur often, thus in most cases, the false REM detections can be easily overrided by the context information of other delta or sigma spindle. Thus, the Stage 5 scoring errors of the system are related more to the variations of subject EEG characteristics than to the false positive REM wave detection. The most significant error is the poor classification of Stage 3. The Stage 3 scoring shows a significant amount of both false positive and negative errors associated with the Stage 2. Stage 3 scoring also shows a significant amount of false positive error resulting from mis-scoring of

PAGE 150

145 Stage 4 as Stage 3. The Stage 3 scoring errors associated with stages 2 and 4 contribute 6.7 % of error in the total average agreement, i.e., 496 mis-scored epochs among the total of 7557 epochs. Thus, it is noticeable that, without the Stage 3 error, the total average agreement is above 90 % for the system. However, the correct scoring of Stage 3 is an inherently difficult problem, since the nature of Stage 3 is a type of transition between Stage 2 and Stage 4. More importantly, the human scorers also show a very low percentage agreement for the sleep stage 3 classification. This is illustrated in Table 5.1 which is obtained from the standard of Agnew and Webb. The four human scorers' average percentage agreement of Stage 3 classification with the six standard calibration records is only 49 %. It is important to point out that the man-machine percentage agreement for the Stage 3 classification cannot become as high as the other stages regardless of the machine's high accuracy, because it is obtained from epoch-by-epoch comparison of the machine scoring with the relatively unreliable human scoring. It is, thus, very doubtful that the system can get a significant improvement in scoring Stage 3 by any other deliberate analysis of delta data. It even seems unrealistic to evaluate an automated system's performance by the present scheme of epoch-by-epoch comparison with the human scoring results. In an effort to improve sleep stage scoring accuracy, machine scoring is often reported without

PAGE 151

146 separating the sleep stages 3 and 4. The overall percentage agreement of this system, without separation of the stage 3 and 4, is 87.3 %. Stage 1 scoring shows a low agreement. This error is associated with the relatively low reliability of human scorer's Stage 1 classification. A Stage 1 epoch does not involve any clear waveform or activity, rather it is defined as a transient state between the other stages; therfore, it is susceptible to erroneous classification into the adjacent epoch stages. Another reason for the low agreement of the Stage 1 scoring is that the Stage 1 scoring rules have not been articulated sufficiently in terms of contextual interpretation. Thus, the rules related to Stage 1 scoring need to be elaborated further by considering various contextual factors in the record. However, since the total number of Stage 1 epochs is only about 3.8 % of the total epochs, the significance of Stage 1 error in the system performance is very small. Stage 0 shows a high agreement rate of 90.1 % with a small false negative error (100 90.1 = 9.9 %). On the other hand. Stage 0 shows a reduced agreement of 71.4 % with a false positive error of 28.6 %. This false positive classification error in the Stage 0 mostly results from the present system's scheme of handling arousal epochs. Arousal epochs are reported as Stage 0 instead of classifying the epochs into the surrounding epoch's stage as is the case in

PAGE 152

147 human scoring. The separate arousal epoch reporting is desirable since the information on arousals is often important for clinical applications. By incorporating an EMG channel in the analysis, a significant portion of Stage 0 error, specifically the misclassif ication of a REM epoch as Stage wake, can be reduced. If this portion of error is removed, the agreement of the Stage 0 scoring is also 84.3 % with a 15.7 % of false positive error. However, 82.3 % (65 from the total of 79 epochs) of this type error, i.e., misclassif ication of a REM epoch as a Stage wake, comes from a single record, 10889, which shows the lowest scoring agreement among the subjects. Thus, if this error from the record 10889 is not counted in the statistics, the false negative error of REM stage scoring is reduced to below 10 %, giving a 90.2 % average agreement. The youngest age group records show the average agreement of 83.1 % as shown in Table 5.4. This age group shows 88.4 % of high agreement without separation of stages 3 and 4. The young adult age group records show the highest average agreement rate of 86.8 % as shown in Table 5.5. If stages 3 and 4 are not separated, then the agreement of this age group is 90.2 %. The oldest age group records reflect the lowest average agreement of 80.6 % as shown in Table 5.6. However, among the four subjects in this age group, two subjects' records, 10067, and 10889, belong to the previously described four records which display a large

PAGE 153

148 Table 5.4. Manold) machine Agreement for Group 1 (5 to 18 years Machine Score V X a J. o \stag6 stagev 0 1 2 3 4 5 total agr. (%) 0 190 13 2 0 1 1 207 91.8 1 20 63 24 0 0 3 110 57.3 2 7 27 1464 55 3 96 1652 88.6 3 0 1 37 116 10 4 168 69.0 4 0 n \J 1 1 181 395 5 592 *92.4 66.7 5 1 28 71 7 0 771 878 87.8 total 218 132 1609 395 409 880 3607 agr(%) 37.2 47.7 91.0 32.3 *91 96.6 .4 87.6 83.1 88.4 * no stages 3 and 4 discrimination. (false positive)

PAGE 154

149 Table 5. 5. Manold) machine Agreement for Group 2 (25 to 34 years Machine Score /'Pal nega txve ; Nstage stages. 0 1 2 3 4 5 total agr. (%) 0 57 2 1 0 0 3 63 90.5 1 7 20 9 0 0 8 44 45.5 2 7 2 1037 74 0 14 1134 91.4 3 0 0 34 84 11 0 129 65.1 4 0 1 X Q O 79 89 0 177 *85.9 50.3 5 4 4 22 0 0 602 632 95.3 total 75 29 1111 237 100 627 2179 agr(%) ' 76 .0 69.0 93.3 35.4 *78. 89.0 0 96.0 86.7 90.8 * no stages 3 and 4 discrimination. (false positive)

PAGE 155

150 Table 5.6. Manold) machine • Agreement for Group 3 (47 to 79 years Machine Score ( false negative ) N^tage stag0«v 0 1 2 3 4 5 total agr . ( % ) 0 189 21 1 0 0 3 214 88.3 1 25 34 34 0 0 41 134 25.4 2 30 8 924 1 0 59 1022 90.4 3 0 0 14 12 0 0 26 46.2 4 0 0 0 0 0 0 0 5 74 13 20 0 0 268 375 71.5 total 318 76 993 13 0 371 1771 agr(%) 59.4 44.7 93.1 92.3 72.2 80.6 (false positive)

PAGE 156

151 variability of EEG characteristics with fair amount of human scoring errors. Thus, it is not reasonable to represent the system's performance for this age group by the above agreement rate. Discussions The disagreement between the automated sleep EEG scoring and human scoring is, as reviewed in the last section, is very much dependent upon the subject's EEG characteristics. Moreover, the cause of error is compounded in its nature as a result of several erroneous factors involved in the process. Among the major factors are: the early-processing system's detection accuracy, probable insufficiency of information extracted from the EEG/EOG data, insufficient codification of sleep stage scoring rules, subject abnormality or wide variations in EEG characteristics, and human scorers errors and inconsistency involved in visual scoring. Unfortunately, straightforward solutions for the above listed problems are not readily available. Much of the problem is related to the human visual recognition process, which is not fully understood in several aspects in terms of application of machine intelligence and A.I. techniques. However, discussions on further improvements can be given based on the exposed limitations of the present methodology.

PAGE 157

152 Presently, the EEG analyzing system is functionally divided in two independent parts, i.e., the early-processing system for waveform recognition and the token processing system for sleep stage scoring. The serial communication link between the two is only for the sending of token data from the early-processing system to the token processing system. The waveform recognition parameters in the earlyprocessing system are fixed. However, it is neccessary to provide parameter adaptability so the system can automatically change values according to the information obtained from the higher level token processing system or from other subject data. Thus, there can exist a kind of feedback loop such that the two separate parts can operate together constituting an incorporated single system. This idea is reasonable since most of the waveform activity parameters vary from subject to subject and even within a subject record. The human scorer has the capability to adapt his perception to detect the occurences of waveforms in the record. The issue of keeping the waveform parameters constant to obtain an objectivity in the computer analysis is refutable. The EEG analyzing problem is based on the detection of waveform occurrences in the record. The objectivity of computer analysis results from a consistent detection of each waveform occurence without giving false positive or negative errors. The waveform parameters provided in various reports and standards must be regarded

PAGE 158

153 as only marginal references for the definitions of waveforms. As long as the computer can detect the waveform occurences in a consistent way closely agreeing with a human scorer, it cannot be a question of whether the waveform parameters are kept constant or changed according to other related contexts. The objectivity in terms of waveform parameter values can be an issue for other aspects of computer analysis of EEG data, such as a study of individual waveform characteristics. A self-adaptive mechanism is necessary to reflect a wider context of information during the data processing. Presently, a sliding window provides a partial solution to reflect local context information in the analysis. The contextual smoothing of the third layer processing, using the sliding window scheme, presently contributes a 2.8 % increase in the total average agreement of 83.6 %. But, a more flexible and powerful mechanism can adaptively modify system knowledge reflecting various context information in a wider record range. This is required to achieve a high manmachine agreement, since the human scorer has the capability to reflect the wide range of context information to the visual scoring of the EEG data. Sleep stage 3 scoring is the most serious task involved in the development of an automated scoring system. Unfortunately, the problem of low agreement of sleep stage 3 scoring cannot be solved readily, and it is not clear even

PAGE 159

154 if it can be solved at all. The most typical example of inconsistency in analyzing delta data is shown by W.B. Webb [We86] . He suggests the delta amplitude threshold not be used, since without it basically the same amount of stage 3 and 4 sleep is observed in the old age group subjects giving a consistency among the various other age groups. If human scorers can not provide a consistent and reliable scoring for this stage, it seems more realistic to incorporate a human scorer's re-examination process of sleep stage 3 scoring with the computer analysis results during the system performance evaluation. The present scheme of blind epochby-epoch comparison is not suitable for performance evaluation of the system's stage 3 scoring.

PAGE 160

CHAPTER VI CONCLUSION The new approach to automated sleep analysis, a knowledge-based expert system on top of a heuristic signal analysis technique, is applied to the development of an automated computer system for the multichannel EEC analysis. The new approach provides a different way of computer analysis of sleep EEC data based on the idea of simulating the human visual inspection process. The approach also provides a different user involving environment which features transparency in accessing the system's knowledge and flexible system operation. These features are important in a domain like sleep EEC analysis. The system is designed for intensive signal processing applications, where a great amount of data requires special considerations of processing efficiency and on-line processing and monitoring features. The rescheduling of the searching sequence, and the architectural aspects using a blackboard model and a sliding window scheme are considered in the design for the processing efficiency and the on-line monitoring application of the system. The waveform recognition system is designed and implemented in a digital environment including the detectors for all waveforms such as REM, K-complexes, SEM, 155

PAGE 161

156 alpha, beta, delta, sigma, theta, muscle artifact, and EMG level. The knowledge-based token processing system, including the shell and the knowledge-base editor, is developed by employing sleep stage scoring rules simulating the human scorer's process of interpretation and classification of the sleep data. The developed system is tested with a randomly selected set of 16 records of subjects from 5 to 70 years of age. The system shows an average man-machine agreement of 83.6 % for the records. The performance level of the present system is not considered as high as that of a human expert. However, the system's performance cannot be simply judged based on the agreement result with the selected records, since performance validation is an involved process including processing of much more data of a wide range of subjects. The system elaboration and validation also requires incorporation of the domain professionals in the process. Potentially, there exists room for further improvement of performance by elaborating on the system's knowledge base with additional contextual information and also by further elaboration of waveform parameters. Presently, most of the classification rules are based on the Rechetscheffen and Kales sleep staging criteria. In order not to lose the objectivity of the analysis, wide deviations from the Rechtschaf f en and Kales criteria, are avoided as

PAGE 162

157 much as possible in the development of the system's knowledge base. Besides the necessity of further rule elaborations, there exist several potential limitations which can be addressed for further improvement of the system performance. The waveform parameters of the early-processing system have to be adjusted according to higher level context information obtained from the token processing system or from other subject data. This will significantly improve the earlyprocessing system's detection reliability. Next, a selfadaptive learning mechanism is necessary to reflect wider contextual information during the data processing as discussed in chapter V. The sleep stage 3 scoring problem does not seem to have any feasible solution, although a good agreement of this stage scoring is significant in terms of total system performance. As discussed in chapter V, it is realistic to incorporate a human scorer's re-examination process of stage 3 scoring with the computer scoring results for the performance evaluation of an automated system, since human scorers do not provide a reliable and consistent sleep stage 3 scoring. The sleep EEG analyzing problem provides an excellent example of a heuristic, knowledge-intensive problem domain for the application of an expert system approach. Objective EEG models and analytic problem solving models are not available, but the analysis relies on the visual inspection

PAGE 163

158 and application of the heuristic interpretation rules of a well-trained human expert. In the heuristic domain like sleep EEG analysis, application of any analytic signal processing approach has significant limitations in designing an automated computer analyzing system, and an algorithm-based conventional programming approach is not appropriate to cope with the heuristic, potentially varying knowledge of the sleep EEG analysis. Expert system approach proposes a new way of solving the problem in this type of domain by providing a flexible way to codify human expert's heuristic knowledge and skilled process with an appropriate representation in the system. However, there exist several fundamental limitations in realizing human-like intelligence in a computer by the present A.I. approaches which are mostly based on the description of features and rule-like inferences. The essential problems related with the immense commonsense knowledge and the lack of knowledge for some of human being's recognition processes, such as image-based inferences, similarity recognition, and relevance pruning processes are crucial barriers to overcome the limitation of current application-oriented A.I. technology in achieving a really competetive machine intelligence at a human expert's performance level. H.L. Dreyfus and S.B. Dreyfus address this issue and conclude that, with the current A.I. and expert system technologies, it is impossible to implement a

PAGE 164

159 competetive intelligence in a computer with a human expert's level of performance [Dr86] . He extends his arguments such that, the human expert's high level performance is ultimately related to his capability of handling a huge amount of special cases in different contexts, and this capability has been obtained through years of practice and experience. This then becomes a fundamental barrier in implementing the heuristic knowledge in a computer, since in most cases codification of the knowledge is inherently a difficult task and the amount of knowledge can easily exceed the manageable range of a computer. The same fundamental questions, i.e., wether or not the human scorer's sleep EEG analyzing knowledge is sufficiently codifiable and wether or not the amount of knowledge to show human expert level performance is within the manageable range of a computer, are not yet answered. However, considering that presently no alternative ways of solving the problem of heuristic domain without using A.I. technology can be offered, it is realistic to accept the expert system approach and to expect its potential contributions in handling this type of heuristic problem by providing a way of incrementally codifying the expert knowledge. The current study shows a new application of the A.I and expert system approach to the EEG signal processing application domain. Although, this research focuses on the automated sleep stage scoring, future research must

PAGE 165

160 incorporate other aspects of an expert system to provide an intelligent environmental tool in the sleep research and its clinical applications areas. All the extracted waveform « activity information can be further manipulated by the computer in various ways according to the user's interests and can be provided for the user in a user-friendly intelligent manner. A natural language interface capability is important for a successful application of the expert system as an intelligent research tool for sleep research and related clinical applications. Since, significant information which can be further extracted from the stored tokens is not limited to any small number of fixed manipulations of the token data, the system must interact with the user in a highly flexible way responding to the user's requirement of wide ranging information exploration. The system must incorporate all the token processing knowledge which can extract potentially significant information from the tokens. The approach and the techniques developed in this study can be further applied to other similar signal processing application domains like, signal processing and interpretation of other biological data, pulmonary disease diagnosis, patient monitoring and anaesthesia, neurology, and sonar signal analysis and monitoring applications.

PAGE 166

APPENDIX A SYSTEM OPERATION EXAMPLE SCREENS The system shows a top level selection menu at the beginning of the system operation. The top level menu is illustrated in Fig. A.l. A user can select one of the previously created knowledge bases or create a new knowledge base by menu selection. If the user selects one of the previously developed knowledge bases, the knowledge base is read into the system from the disk. The next level menu is then displayed on the screen as is shown in Fig. A. 2. If the user selects to create a new knowledge base, the system will interactively ask for values of the global variables. After the completion of the global data input, the system will lead to the next level menu for further editing of the knowledge base. The menu shown in Fig. A. 2 is the top level menu for a selected knowledge base. The user can run and test the knowledge base or, he can edit the knowledge base rules, parameters, variables, detector reliabilities, and query input parameters. The user can also simply list the knowledge base, and can save the modified or created knowledge base into a disk. The editing of the knowledge base is guided by user friendly menus and prompts. Examples 161

PAGE 167

162 Sleep Analyzing Expert Syatem i University of Florida Knowledge base name: Enter a number from the following MENU: 1. EEG-KBASE 2. TEST-KBASE 3. Create a new knowledge-base 4. Exit Fig. A.l. Top Level Menu for Knowledge Base Selection. TOPMENU: Shell a EB<5-KBASE Enter a number from the following MSmt 2. RULES 3 . PARAMETERS 4. VARIABLES 5. DETECTOR-N-QUERY-PARMS 6. SAVE 7. QUIT Fig. A. 2. Main Menu for Activity Selection.

PAGE 168

163 of screens for rule editing sessions are shown in figures A. 3, A. 4, A. 5, and A. 6. At the beginning of the execution, the system asks several questions for user inputs such as subject's age, record number, and date. Then the system displays all the query input parameters one by one with a default value as is illustrated in Fig. A. 7. If the user wants to change the value of the query parameter, he can input a new value with an associated certainty level. The system then starts running and displays epoch-wise classification results on the screen. An example of monitoring screen, while the system is running, is illustrated in Fig. A. 8. Rule identification codes for all the related rules in each epoch classification are also displayed together with the classification result. The left most column displays the final score for each epoch. A running window containing the data of the current five epochs, is also shown at the bottom of the screen. Contextual manipulations are performed at the end of every epoch classification when new epoch data is attached to the window making all the data slide one epoch position. The first epoch in the running window becomes the final Classification result. This is listed together with the epoch-wise classification rules executed before contextual smoothing. During the execution, a user can interrupt at any time, if he wants an explanation of the results, by pressing

PAGE 169

164 EJtpsrt System Shell : :^ EBG-KBASS TOPMENU: Rules RULE-GROUP: CHANNELS-GROUP Edit CHANNELS -GROUP rules: Modify Enter a number the follov;lng 2. DELETE-RULE 3. MODIFY-RULE 4. LIST-RULE 5. QUIT Fig. A. 3. Menu for Rule Edition. Tnnv,n.K,„ B*pert System Shell ;j S?:g~KBASE TOPMENU: Rules RULE-GROUP: CHANNELS -GROUP Edit CHANNELS-GROUP rules: Modify *** Insert rule-6 of ALPHA-CH *** Rule i . d. : iiii^er f n" an atom f CJjriti , OR^^ Fig. A. 4. An Example Screen for Rule Insertion.

PAGE 170

165 Ikpert System Shell TOPMENU: Rules RULE-GROUP: CHANNELS -GROUP Edit CHANNELS -GROUP rules: Modify *** Insert rule-6 of ALPHA-CH *** Rule i.d: A-38 Enter Premise:
PAGE 171

166 Date: (12-16-1986) Subject Number: 11717 Subject Age: 27 *** Query Parameters *** E^ert Systeai SheXt tt BB*3-KBASE (SUB-ALPHA HIGH M) Fig. A. 7. Query Parameter Values Input.

PAGE 172

167 tem Shelo. Date : ^^^2-16^1986) Subject Number: 11717 Subject Age: 27 *** Query Parameters *** (SUB-ALPHA High M) (UNDER-DRUG NO H) *** Running Sleep Stage Scoring *** EEC-KBASE :*** waxt for RUN *** Press F to pause or S to stop running: 1 2 3 4 5 (S-0 H) (S-0 H) (G-12 G-24 A-2 G-14 G-62 G-66) (S-0 H) (S-0 H) (G-12 G-24 A-1 G-8) (S-0 H) (S-0 H) (G-12 G-24 A-1 G-8) (S-0 M) (S-0 M) (G-12 G-63 A-2 G-14 G-62 G-24 G-43 G-17 G-34 ) (S-0 H) (S-0 H) (G-12 G-63 A-1 G-8 G-24 G-43 G-17 G-34) ($-0 S-0 S-1 S-1 S-2) liiifcif ' u h: m m ) iS-2 S-1 S-3 S-4 S-0 S-5} 10 (S-2 M) (G-19 G-56 G-11 G-15 G-5 G-6 G-9 G-31 G-41 G-16) Fig. A. 8. Monitoring Screen During System Run.

PAGE 173

168 the 'P' key. The system provides an explanation mechanism such that the user can trace a rule execution line. User can trace the line of reasoning by selecting all the related parameters listed as is shown in Fig. A. 9 for an example. An example of explanation is shown in Fig. A. 10.

PAGE 174

169 ^.C I^xpert System Shell : Date: (12-16-1986) Subject Number: 11717 Subject Age: 27 Sex: Male *** Running Sleep Stage Scoring *** I. STAGE 2. WAKE-WAVE 3. DELTA 4. REM 5. MINUTE 6. SLP-WAVE 7. ALPHA 8. ALPHA-TIME 9. DELTA-TIME 10. SIGMA II. SIGMA-COUNT 12. REM-COUNT 13. QUIT Fig. A. 9. Parameter Selection Menu for Explanations, Date: (12-16-1986) Subject Number: 11717 Subject Age: 27 Sex: Male *** Running Sleep Stage Scoring *** ^^:#ss CR to continue : *** Explanation minute : 4 *** Rule ID : G-12 IF : (EQ WAKE-WAVE HIGH) THEN : (STAGE STAGEO M) °5 J?® determined as STAGEO by the above rule G-12 such that : If an epoch's wake related wave activity IS high, then the epoch is scored as STAGEO with a medium level of certainty. Fig. A. 10. An Example of Explanation.

PAGE 175

APPENDIX B SYSTEM RULES The current system contains 112 domain rules in the knowledge base of the system. The system can be further elaborated by adding more rules. In this appendix, the rule base of the current system As listed in a summarized table form. Each waveform activity time or number of occurence per minute is associated with the reliability factor of thje waveform detection, and these data constitute a part of the bottom level data plane. Each wave activity of the epoch is then described as High, Medium, and Low according to the level of occurrence in each epoch. The waveform activity description is associated with a certainty level, either High, Medium, or Low, which is obtained by f uzzif ication rules within the range of each activity level and also by taking into account the associated waveform detector reliability level. A set of f uzzif ication scheme employed in the system is shown in Table B.l for each wave activity. New wave-activities can be further defined by combining the above descriptor information. These new descriptors can give further abstraction of the waveform activity data. This system includes the descriptors of Wake-W-Activity and Sleep-W-Activity . The Wake-W-Activity 170

PAGE 176

171 Table B.l. Fuzzif ication Scheme Table. Activity Level High Medium Low Certainty Level H M L M L M H Alpha >35 >30 >25 25 > > 10 <10 <5 Beta >20 >15 >10 10 > > 5 <5 Delta >30 >20 20 > > 12 <12 <8 Sigma >3 2 1 0 Theta >20 >15 >10 10 > > 5 <5 Muscle >30 >20 20 > > 5 <5 SEM >15 15 > > 5 <5 REM >3 2 1 0 Kcomp >3 2 1 0

PAGE 177

172 reflects the level of the wave activities which are closely related with subject's wake state, i.e., combined information of alpha-activity and muscle-activity. The Sleep-W-Activity is obtained by combining the wave-activity information of sigma spindle and k-complexes. The definition rules for the Wake-W-Activity and the Sleep-WActivity are shown in Table B.2 and B.3, respectively. Each epoch is then classified into one of the stages according to the classification rules. The classification rules provide templates for each stage. The templates are obtained by the combination of the wave act ivity descriptors. Each epoch is matched with one of the templates in the classification rules and those epochs which fail to match are classifieded as the previous epoch's stage with the lowest certainty level. Contextual manipulation will be performed on the intermediate classification results at the next level processing. The classification rules for each stage include the Rech t s che f f en and Kales classification standards. Key features of each stage are described in the following and summarized in Table B.4. Stage 0: Stage 0 templates are obtained by the appearance information of muscle artifact, alpha activity, and/or eye movement activity. The other wave activities, such as sigma spindle, k-complexes, and delta activities are used as the exclusive information for Stage 0 templates. EMG level is

PAGE 178

173 Table B.2. Combination Rules for Wake-W-Activity. High Medium Low H M L M L H AlphaActivity H H NH M M NH L MuscleActivity H NH H M L L NH Table B.3. Combination Rules for Sleep-W-Activity . High Medium Low H M L M L M H SigmaActivity H M L L KcompActivity _ _ _ H NH

PAGE 179

174 Table B.4. Classification Templates. Activity Descriptor Stage 0 Stage 1 Stage 2 M L M M H M L Wake-Wave H M H L L L H Sleep-Wave L L L H M Alpha Beta H Delta L L L L L L L Sigma Theta Muscle SEM REM L L L L L Kcomp L MINUTE <50 <50 Delta-time >8 Subject-Alpha Under-drug Pre-1 S-5 s-5 s-5 Pre-2 s-5 : Underline indicates the negation of the element.

PAGE 180

175 Table B . 4--continued Activity Descriptor Stage 3 Stage 4 Stage 5 L M H M L L Wake-Wave H H H L L L Sleep-Wave L L L L Alpha Beta H Delta M H L L L L Sigma Theta Muscle SEM REM L Kcomp MINUTE >50 >50 Delta-time Subject-Alpha Under-drug Pre-1 s -5 Pre-2 S-5 Note: Underline indicates the negation of the element.

PAGE 181

176 high for the Stage 0 epochs and this EMG information gives a clear separation between Stage 0 and REM sleep (Stage 5). Since the EMG channel is not utilized in the system, the level of alpha activity (for a high alpha subject) and muscle artifact provide the separation of REM sleep from Stage 0 in addition to the REM activity information. Stage 1: Stage 1 templates are mostly based on the absence of wave activities such as Wake-W-Activity, Sleep-W-Activity , delta Activity, and REM activity, and occurrence of beta activities. Stage 2: The appearances of sigma spindles and k-complexes provide the clearest and most reliable templates for Stage 2. Since the sigma spindle and k-complexes also appear in Stage 3 and Stage 4, the delta activity level is used to determine the separation of Stage 2 from stages 3 and 4. Stage 3, 4: The level of delta activity provides the templates for the stages 3 and 4. Stage 5: The REM, the beta activity, and the disappearance of the other wave activities, constitute the major part of Stage 5 templates. If the EMG is utilized in the system, the low

PAGE 182

177 EMG amplitude information can provide a supplementary condition for the Stage 5. The contextual smoothing with a running window plays an important role in simulating the human scorer's visual inspection of the record. The contextual manipulation is performed with the running window of five consecutive epochs. Abstracted high level information, i.e., a classification stage and a matching certainty level of each epoch, provides the context data for the window. The major contextual manipulation rules are briefly described in the following and are summarized in Table B,5. o The Stage 1 epochs which precede within five epochs from the REM Stage are all smoothed into Stage 5. o Stage 1 epochs which succeed REM epochs are all smoothed into Stage 5. o Stage 1 epochs are smoothed into the surrounding epoch's stage, if the Stage 1 epochs are not associated with the highest certainty level and appear in less than three consecutive epochs, o Stage 2 epoch which is surrounded by the REM epochs is smoothed into Stage 5, if the matching level of Stage 2 is not high, o If Stage 2, 3, 4 epochs, which are associated with the lowest matching certainty level, are surrounded by the other stage epochs which are

PAGE 183

178 Table B.5. Contextual Smoothing Rules. — > V s— u (s-0 * s— u ^ s-0 * S-1 ^ s-0 L s-0 * s-0 * s-0 * ) s-0 *) — > ( r\ ie \ s-u ^ (s-0 * r\ -le s-u ^ s-0 * S-2 * S-0 L s-0 * s-0 * s-0 *) s-0 *) — > \ s — U ^ (s-0 * r~, n\ -ie S-U " s-0 * S-0 * S-0 L s-0 * s-0 * s-0 * ) s-0 *) — > I s— u ^ (s-0 * S-1 ^ S-0 L s-0 * ) s-0 *) — > / o O * V S-z ^ (s-2 * „ O 4* s-z * S-2 * s-1 * s-2 L s-2 * s-2 * s-2 *) s-2 *) — > ^ e5 O * \ s — ^ (s-2 * ^ o ^ S-Z * S-2 * S-3 * s-2 L s-2 * s-2 * s-2 *) s-2 *) — > (s-2 * (s-2 * S-2 * s-2 * s-5 * s-2 L s-2 * s-2 * o Z / s-2 *) — > (s-3 * (s-3 * s-3 * s-3 * s-2 * s-3 L s-3 * s-3 * s-3 *) s-3 *) (s-0 * (s-0 * s-1 * s-0 L s-1 * s-0 L s-0 * s-0 * s-0 *) s-0 *) — > (s-0 * (s-0 * s-2 NH s-0 L s-0 *) s-0 *) — > (s-5 * (s-5 * s-5 * s-5 * s-0 * s-5 L s-5 * s-5 * s-5 *) s-5 *) Note: The symbol * indicates any certainty level.

PAGE 184

179 Table B.5 — continued — > ( s-1 (s-5 L S-2 * s-2 * s-5 *) s-5 *) — > ( s-0 (s-0 * S-2 NH S-0 L s-2 NH s-0 L s-0 s-0 * * s-0 s-0 *) *) — > ( s-0 (s-0 * s-5 * S-0 L s-0 *) s-0 *) — > ( s-0 (s-0 NL * s-5 * s-0 L s-5 * s-0 L s-0 s-0 * * s-0 s-0 *) *) — > ( S-4 (s-4 * s-4 * s-4 * s-3 * s-4 L s-4 s-4 * * s-4 s-4 *) *) — > (s-4 (s-4 NL * s-3 * s-4 L S-3 * s-4 L s-4 * * IS — A s-4 ) *) — > (s-3 (s-3 * * s-1 * s-3 L s-3 *) s-3 *) (s-3 (s-3 * * s-2 * s-3 L s-2 * s-3 L s-3 s-3 * * s-3 s-3 *) *) — > (s-5 (s-5 * * s-2 * s-5 L s-2 * s-5 L s-5 s-5 * * s-5 s-5 *) *) — > (s-2 (s-2 * * s-1 * s-2 L s-2 *) s-2 *) (s-2 * s-1 * s-1 * s-2 * s-2 *) ---> (s-2 * s-2 L s-2 L s-2 * s-2 *) Note: The symbol * indicates any certainty level.

PAGE 185

180 Table B.5 — continued — > (s-2 * (s-2 * s-2 s-2 * s-1 S-2 * L S-1 s-2 * L S-2 s-2 NL) * ) — > (s-2 NL (s-2 * s-1 s-2 * L s-1 s-2 * L s-1 s-2 * L s-2 s-2 NL) * ) — > (s-2 * (s-2 * s-2 s-2 * * s-0 s-2 * L s-2 s-2 * * s-2 s-2 *) *) — > (s-2 * (s-2 * s-5 s-2 * L s-5 s-2 * L s-2 s-2 * * s-2 s-2 *) *) — > (s-4 * (s-4 * s-1 s-4 * L s-4 s-4 *) *) ( s-4 NL (s-4 * S-4 * L s ^ s-4 ic L S-4 s-4 * S-4 S-4 *) *) (s-2 * (s-2 * s-2 s-2 * * s-4 s-2 * L s-2 s-2 * * s-2 s-2 *) *) — > (s-5 * (s-5 * s-5 s-5 * * s-2 s-5 * L s-5 s-5 * * s-5 s-5 *) *) (s-3 * (s-3 * s-3 s-3 * * s-4 s-3 * L s-3 s-3 * * s-3 s-3 *) *) — > (s-1 * (s-5 L s-5 s-5 *) *) (s-5 * s-1 *) > (s-5 * s-5 L)

PAGE 186

181 Table B.5 — continued — > (s-3 (s-3 NL * S-4 s-3 * L S-4 s-3 * L S-3 s-3 * * s-3 s-3 *) *) — > (s-2 (s-2 NL * s-3 s-2 * L s-3 s-2 * L s-2 s-2 * * s-2 s-2 *) *) — > (s-4 (s-4 * * s-3 s-4 * L s-4 s-4 * * s-4 s-4 *) *) — > (s-3 (s-3 * * s-4 s-3 ie L s-3 s-3 * * s-3 s-3 *) *) — > ( s-3 (s-3 * * s-2 s-3 * L s-3 s-3 * * s-3 s-3 *) *) (s-5 (s-5 * * s-2 s-5 * L s-5 s-5 * * s-5 s-5 *) *) — > (s-2 (s-2 * * s-5 s-2 * L s-2 s-2 * * s-2 s-2 *) *) — > (s-2 (s-2 * * s-3 s-2 * L s-2 s-2 * * s-2 s-2 *) *) — > (s-5 (s-5 * * s-3 s-5 * L s-5 s-5 * * s-5 s-5 *) *) Note: The symbol * indicates any certainty level.

PAGE 187

182 associated with a high matching certainty level, and the length of the intervening epochs are less than three, then the intervening epochs are smoothed into the surrounding epoch's stage. Rescheduling of the searching path is performed by the arrangement of the goal hypotheses of the intermediate level classification rules. The system can employ many number of scheduling rules which can assign each searching path by matching against the running window. However, for simplicity, the current scheme employs only six rules and the searching paths are empirically decided as the followings depending on the last epoch's stage of the window. Search Path (s-0 s-1 s-2 s-3 s-5 s-4) (s-1 s-2 s-5 s-0 s-3 s-4) (s-2 s-3 s-1 s-5 s-0 s-4) (s-3 s-4 s-2 s-0 s-5 s-1) (s-4 s-3 s-2 s-0 s-5 s-1) (s-5 s-1 s-2 s-0 s-3 s-4) Last epoch's Stage stage 0. stage 1 . stage 2. stage 3. stage 4. stage 5.

PAGE 188

APPENDIX C MAN-MACHINE SCORING AGREEMENT TABLES FOR EACH SUBJECT RECORD 183

PAGE 189

184 Table C. 1. Manold) machine Agreement Table for 10719 (5 years Machine Score ( false negative) stage stage 0 1 2 3 4 5 total agr. (%) 0 37 2 1 0 0 0 40 92.5 1 0 7 0 0 0 8 7 100.0 2 2 9 168 28 2 54 263 63.9 3 0 0 0 10 7 0 17 58.8 4 0 n \j r\ U 2 120 0 122 *100.0 98.4 5 0 15 23 6 0 137 181 75.7 total 39 33 192 46 129 191 630 agr(%) 94 .9 21.2 87.5 21.7 *79 93.0 .4 71.7 76.0 *77.5 no stages 3 and 4 discrimination . (false positive)

PAGE 190

• 185 Table C.2. Manold) machine • Agreement Table for 10150 (8 years Machine Score ( false negative ) stage 0 stage 1 2 3 4 5 total agr. (%) 0 15 0 0 0 0 0 15 100.0 1 7 4 5 0 0 0 16 25.0 2 1 3 207 11 0 1 223 92.8 3 0 0 0 31 1 0 32 96.9 0 0 42 94 0 135 *100.0 69.1 5 1 7 13 1 0 120 142 84.5 total 24 14 225 85 95 121 564 agr(%) 62.5 28.6 92.0 35.5 *93 98.9 .3 99.2 83.5 *91.1 * no stages 3 and 4 discrimination . (false positive)

PAGE 191

186 Table C. 3. Manold) machine Agreement Table for 10158 (13 years Machine Score ( false negative) stage stage 0 1 2 3 4 5 total agr. (%) 0 53 0 0 0 1 0 54 98.1 1 3 0 6 0 0 0 9 0.0 2 1 0 263 12 0 0 276 95.3 3 0 0 2 24 1 0 27 88.9 4 0 0 n 1 65 0 66 *97.8 98.5 5 0 0 28 0 0 66 94 70.2 total 57 0 299 37 67 66 526 agr(%) 93 .0 88.0 64.9 *87. 97.0 100 5 .0 89.5 *89.9 no stages 3 and 4 discrimination. (false positive)

PAGE 192

187 Table C.4 • Manold) machine Agreement Table for 10109 (13 years Machine Score ( false negative ) stage stage 0 1 2 3 4 5 total agr . ( % ) 0 5 1 0 0 0 0 6 83.3 1 2 3 8 0 0 0 13 23.1 2 1 10 230 1 0 2 244 94.3 3 0 0 6 8 1 0 15 53.3 4 0 0 2 17 32 2 53 *85.3 60.3 5 0 2 1 0 0 134 137 97.8 total 8 16 247 26 33 138 468 agr(%) 62. 5 18.8 93.1 30.8 *100 97.0 .0 97.1 88.0 *91.9 * no stages 3 and 4 discrimination. (false positive)

PAGE 193

188 Table C .5. Manold) machine Agreement Table for 10146 (13 years Machine Score ( false negative) stage stage 0 1 2 3 4 5 total agr . ( % ) 0 26 2 0 0 0 0 28 92.9 1 1 12 2 0 0 0 15 80.0 2 1 0 190 3 0 3 197 96.4 3 0 0 17 29 0 0 46 63.0 4 0 0 0 47 48 0 95 *87.9 50.5 5 0 1 1 0 0 118 120 98.3 total 28 15 210 79 48 121 501 agr(%) 92.9 80.0 90.5 36.7 100.0 97.5 84.4 *93 .3 *93.8 * no stages 3 and 4 .discrimination. (false positive)

PAGE 194

189 Table C. 6. Manold) machine Agreement Table for 10114 (13 years Machine Score ( false negative ) stage stage 0 1 2 3 4 5 total agr. (%) 0 8 8 1 0 0 1 18 44.4 1 3 30 3 0 0 3 39 76.9 2 1 5 197 0 0 35 238 82.8 3 0 1 10 n \j U 4 15 0.0 4 0 0 9 46 8 3 66 *66.7 12.1 5 0 3 5 0 0 137 145 94.5 total 12 47 225 46 8 183 521 agr(%) 66 .7 63.8 87.6 0.0 100. *100.0 0 74.9 72.9 *81.8 no stages 3 and 4 discrimination (false positive)

PAGE 195

190 Table C.7. Man-machine Agreement Table for 11623 (18 years old) . Machine Score ( false negative ) stage stage 0 1 2 3 4 5 total agr. (%) 0 50 0 0 0 0 0 50 100.0 1 4 7 0 0 0 0 1 1 X J. o . D 2 0 0 209 0 1 1 211 99.1 3 0 0 2 14 0 0 16 87 S 4 0 0 0 26 28 0 54 *97.1 51.9 5 0 0 0 0 0 59 59 100.0 total 54 7 211 40 29 60 401 agr(%) 92 .6 100.0 99.1 35.0 *98. 96.6 98 6 .3 91.5 *98.0 * no stages 3 and 4 discrimination . (false positive)

PAGE 196

191 Table C.8. Man-machine Agreement Table for 11769 (25 years old) . Machine Score ( false negative) stage 0 1 2 3 4 5 total agr.(%) 0 9 1 1 0 0 2 13 69.2 1 1 4 4 0 0 2 11 36.4 2 0 0 219 18 0 0 237 92.4 3 0 0 0 22 1 0 23 95 . 7 4 0 0 0 21 10 0 31 *100.0 32.3 5 0 1 0 0 0 155 156 99.4 total 10 6 224 61 11 159 471 agr(%) 90 .0 66.7 97.8 36.1 *75 90.0 .0 97.5 89.0 *93.6 * no stages 3 and 4 discrimination (false positive)

PAGE 197

192 Table C.9. Manold) machine Agreement Table for 11717 (27 years Machine Score ( false negative ) siiaae stage 0 X z 3 4 5 total agr. (%) 0 13 0 0 0 0 0 13 100.0 1 0 6 2 0 0 0 8 75.0 2 0 0 204 1 0 9 214 95.3 o n 0 11 7 2 0 20 35.0 4 0 1 8 18 20 0 47 *70.1 42.6 5 0 0 4 0 0 115 119 96.6 total 13 7 229 26 22 124 421 agr(%)100 .0 85.7 89.1 26.9 *97. 90.9 9 92. 7 86.7 *91.4 * no stages 3 and 4 discrimination (false positive)

PAGE 198

193 Table C.IO. Man-machine Agreement Table for 11747 (29 years old) . Machine Score (false negative) stage 0 5 total agr.(%) stage 0 11 0 0 0 0 0 11 100.0 1 1 3 2 0 0 0 6 50.0 2 5 1 204 1 0 1 212 96.2 3 0 0 21 6 0 0 27 22.2 *73.1 4 0 0 0 37 14 0 51 27.5 5 1 0 8 0 0 127 136 93.4 total 18 4 235 44 14 128 443 agr(%) 61.1 75.0 86.8 13.6 100.0 99.2 82.4 *98.3 *90.7 * no stages 3 and 4 discrimination. (false positive)

PAGE 199

194 Table C .11. Manold) machine • Agreement Table for 10244 (34 years Machine Score ( f al=5«i \ -1. CL A. O C stage stage 0 1 2 3 4 5 total agr. (%) 0 20 0 0 0 0 0 20 100.0 1 4 5 0 0 0 3 12 41 . 7 2 1 1 195 43 0 2 242 80.6 3 0 0 0 30 5 0 35 85.7 4 0 0 0 0 0 0 0 *100.0 5 3 1 10 0 0 101 115 87.8 total 28 7 205 73 5 106 424 agr(%) 71.4 71.4 95.1 41.1 0.0 95.3 82.8 *44. 9 *84.0 * no stages 3 and 4 discrimination (false positive)

PAGE 200

195 Table C.12. Man-machine Agreement Table for 10256 (34 years old). Machine Score ( false negative ) stage stage 0 1 2 3 4 5 total agr . ( % ) 0 4 1 0 n n u X D 56.7 1 1 2 r\ U u o 7 28 . 6 2 1 0 215 11 0 2 229 93.9 3 0 0 2 19 3 0 24 79.2 4 0 0 0 3 45 0 48 *97.2 93.8 5 0 2 0 0 0 104 106 98.1 total 6 5 218 33 48 110 420 agr(%) 66. 7 40.0 98.6 57.6 *86 93.8 .4 94.5 92.6 *94.0 * no stages 3 and 4 discrimination. (false positive)

PAGE 201

196 Table C. 13. Manold) machine • Agreement Table for 10067 (43 years Machine Score V. -L u JO stage stage 0 1 2 3 4 5 total agr. (%) 0 62 8 0 0 0 1 71 87.3 1 5 9 12 0 0 31 57 15.8 2 0 3 148 1 0 43 195 75.9 3 0 0 0 12 0 0 12 100.0 4 0 0 0 0 0 0 0 5 3 6 13 0 0 100 122 82.0 total 70 26 173 13 0 175 457 agr(%) 88.6 34.6 85.5 92.3 57.1 72.4 (false positive)

PAGE 202

197 Table C. 14. Manold) machine • Agreement Table for 11771 (53 years Machine Score ( f 1 y -iu ^ o c stage stage 0 1 2 3 4 5 total agr. (%) 0 5 11 1 0 0 0 17 29.4 1 2 9 13 0 0 7 31 29.0 2 4 2 308 0 0 12 326 94.5 3 0 U 1 A 14 0 0 0 14 0.0 4 0 0 0 0 0 0 0 5 0 0 0 0 0 58 58 100.0 total 11 22 336 0 0 77 446 agr(%) 45.5 40.9 91.7 75.3 85.2 (false positive)

PAGE 203

198 Table C .15. Manold) machine • Agreement Table for 10889 (53 years Machine V ia X stage stage 0 1 2 3 4 5 total agr.(%) 0 20 1 0 0 0 2 23 87.0 1 11 0 5 0 0 3 19 0.0 2 23 0 243 0 0 2 268 90.7 3 0 0 0 0 0 0 0 4 0 0 0 0 0 0 0 5 65 3 6 0 0 55 129 42.6 total 119 4 254 0 0 62 439 agr(%) 16.8 0.0 95.7 88.4 72.4 (false positive)

PAGE 204

199 Table C . 16 . Manold) machine • Agreement Table for 11740 (70 years Machine Score ( false negative ) stage stage 0 1 2 3 4 5 total agr. (%) 0 102 1 0 0 0 0 103 99.0 1 7 16 4 0 0 0 27 59.3 2 3 3 225 0 0 2 233 96.6 3 0 0 0 0 0 0 0 4 0 0 0 0 0 0 0 5 6 4 1 0 0 55 66 83.3 total 118 24 230 0 0 57 429 agr(%) 86.4 66.7 97.8 96.5 92.8 (false positive)

PAGE 205

REFERENCES Ad76 J.B. Adams, "Probabilistic Reasoning and Certainty Factors," Mathematical Bioscience, Vol. 32, 1976, 177-186. Ag72 H.W. Agnew, Jr. and W.B. Webb, "Sleep Stage Scoring," The Journal Supplement Abstract of American Psychological Association, Manuscript No. 293, May 1972. As55 E. Aserinsky and N. Kleitman, "The Types of Ocular Motility Occurring in Sleep, " Journal of Applied Physiology, Vol. 8, No. 1, July 1955, 1-10. Au81 J.I. Aunon and D.G. Childers, "Signal Processing in Evoked Potential Research: Averaging and Modeling, " Critical Reviews in Bioengineering, Vol. 5, No. 4, July 1981, 323-367. Ba79 J.S. Barlow, "Computerized Clinical Electroencephalography in Perspective," IEEE Trans. Biomed. Eng., Vol. 26, No. 7, July 1979, 377-391. Ba82a A. Barr and E.A. Feigenbaum, The Handbook of Artificial Intelligence, Vol. 1, William Kaufman Inc., Los Altos, California, 1982, 141-222. Ba82b A. Barr and E.A. Feigenbaum, The Handbook of Artificial Intelligence, Vol. 2, William Kaufman Inc., Los Altos, California, 1982, 175-222. Bi73 R.G. Bickford, J. Brimm, L. Berger, and M. Aung, "Application of Compressed Spectral Array in Clinical EEC," In: Automation of Clinical Electroencephalography, Ed. by P. Kellaway and J. Peterson, Raven Press, New York, 55-64. Bi78 CD. Binnie, B.C. Batchelor, P. A. Bowring, C.E. Darby, L. Herbert, D.S.L. Lloyd, D.M. Smith, G.F. Smith, and M. Smith, "Computer-Assisted Interpretation of Clinical EEGs," Electroenceph . Clin. Neurophysiol . , Vol. 44, 1978, 575-585. Bo77 G. Bodenstein, and H.M. Praetorius, "Feature Extraction from the Electroencephalogram by Adaptive Segmentation," Proc. IEEE, Vol. 65, No. 5, May 1977, 200

PAGE 206

201 642-652. Bo80 J.R. Bourne, B. Hamel, D. Giese, G.M. Woyce, P.L. Lawrence, J.W. Ward, and P.E. Teschan, "The EEG Analysis System of the National Cooperative Dialysis Study," IEEE Trans. Biomed. Eng., Vol. 27, No. 11, Nov. 1980, 656-664. Bo84 J.R. Bourne, "Expert Systems in Medicine: A Biomedical Engineering Perspective, " Critical Reviews in Biomedical Engineering, Vol. 12, No. 2, 1984, 95-129. Bu69 B.C. Buchanan, G.L. Sutherland, and E.A. Feigenbaum, "Heuristic Dendral: A Program for Generating Explanatory Hypotheses in Organic Chemistry, " Machine Intelligence 4., Edinburgh University Press, Edinburgh, 1969, 209-254. Bu70 B.G. Buchanan, G.L. Sutherland, and E.A. Feigenbaum, "Rediscovering Some Problems of Artificial Intelligence in the Context of Organic Chemistry." Machine Intelligence 5., Edinburgh University Press, Edinburgh, 1970, 253-280. Bu84 B.G. Buchanan and E.H. Shortliffe, Rule-based Expert Systems : The Mycin Experiments of the Stanford Heuristic Programming Project, Addison-Wesley Publishing Company, Reading, Massachusetts, 1984. Ca76 R.E. Carhart and D.H. Smith, "Applications of Artificial Intelligence for Chemical Inference XX. Intelligent Use of Constraints in Computer-assisted Structure Elucidation, " Computers and Chemistry Vol.1, No. 79, 1976. Ch73 D.G. Childers, W.M. Mesa, and O.S. Halpeny, "A Neuronal Population Model for Simulation of Spatio-Temporal Evoked EEG," IEEE Trans., Syst., Man and Cyber., Vol. 3, No. 4, July 1973, 336-348. Ch81 D.G. Childers and J.I. Aunon, "Spectral Analysis: Prediction and Extrapolation, " Critical Reviews in Bioengineering, Vol. 6, No. 2, Sept. 1981, 133-175. Co65 J.W. Cooley and J.S. Tukey, "An Algorithm for the Machine Calculation of Complex Fourier Series," Mathematics of Computation, Vol. 19, 1965, 297-301.' Co67 Cooley, P.A.W. Lewis, and P.D. Welch, "Historical Notes on the Fast Fourier Transform, " IEEE Trans. Audio Electracoust . , Vol. 15, June 1967,

PAGE 207

202 76-79. Co77 B.A. Cohen and A. Sances, "Stationarity of the Human Electroencephalogram," Med. & Biol. & Computers, Vol. 15, 1977, 513-518. Co80 R. Cooper, J.W. Osselton, J.C. Shaw, EEG Technology, Butterworths & Co. Ltd., London, England, 1980. Co81 P.R. Cohen and E.A. Feigenbum, The Handbook of Artificial Intelligence, Vol. 3, HeurisTech Press, Stanford, California, 1981. Co85 P. Cohen, A. Davis, D. Day, M. Greenberg, R. Kjeldsen, S. Lander, and C. Loiselle, "Representativeness and Uncertainty in Classification Systems," The AI Magazine, Fall 1985, 136-149. Cr85 O.D. Creutzfeldt, G. Bodenstein, and J.S. Barlow, "Computerized EEG Pattern Classification by Adaptive Segmentation and Probability Density Function Classification. Clinical Evaluation," Electroenceph. Clin. Neurophysiol . , Vol. 60, 1985, 373-393. Da82 R. Davis, "Expert System: Where Are We? And Where Do We Go From Here?" The AI Magazine, Vol. 3, No. 2, Spring 1982, 135-173. De75 H.E. Degler, Jr., J.R. Smith, and F.O. Black, "Automatic Detection and Resolution of Synchronous Rapid Eye Movements," Computers and Biomedical Research, Vol. 8, 1975, 393-404. Dr86 H.L. Dreyfus and S.E. Dreyfus, Mind Over Machine, The Free Press, A Division of Macmillan, Inc., New York, N.Y. 10022, 1986. Du73 G. Dumermuth and T. Keller, "EEG Spectral Analysis by Means of FFT," In: Automation of Clinical Electroencephalography, Raven Press, New York, 1973, 145-160. Du79 R. Duda, J. Gaschnig, and P.E. Hart, "Model Design in the PROSPECTOR Consultant System for Mineral Exploration, " Expert Systems in Micro-electronic Age, Edinburgh University Press, Edinburgh, 1979, 153-167. En77 R.S. Engelmore and H.P. Nii, "A Knowledge-based System for the Interpretation of Protein X-ray Crystallographic Data, " Heuristic Programming Project Rep. No. HPP-77-2, Computer Science Dept., Stanford University, 1977.

PAGE 208

203 En79 R.E. Engelmore and A. Terry, "Structure and Function of the CRYSALIS System," IJCAI, Vol. 6, 250-256. ErSOa L.D. Erman, F.H. Hayes-Roth, V.R. Lesser, and D.R. Reddy, "The Hearsay-II Speech Understanding System: Integrating Knowledge to Resolve Uncertainty, " ACM Computing Surveys, Vol. 12, No. 2, June 1980, 213-253. ErSOb L.D. Erman and V.R. Lesser, "Hearsay-II Speech Understanding System: A Tutorial." In: Trends in Speech Recognition. Ed. by W. Lea, Prentice-Hall, Englewood Cliffs, N.J., 1980, 361-381. Fe80 I. Feinberg, G. Fein, and T.C. Floyd, "Period and Amplitude Analysis of NREM EEC in Sleep: repeatabilty of results in young adults, " Electroenceph. Clin. Neurophysiol . , Vol. 48, 1980, 212-221. Fe71 P. B.C. Fenwick, P. Hiche, J. Dollimore, and G.W. Fenton, "Mathematical Simulation of the Electroencephalogram Using an Autoregressive Series, " Biomedical Computing, Vol. 2, 1971, 281-307. Fi85 R. Fikes and T. Kehler, "The Role of Frame-Based Representation in Reasoning, " Communications of the ACM, Vol. 28, No. 9, Sep. 1985, 904-920. Ga86 S. Gala, "Application of the Dampster Shafer Evidence Theory in Sleep EEG Data Analysis," Master's Thesis, the Electrical Engineering Department at the University of Florida, 1986. Ge75 A.S. Gevins, C.L. Yeager, S.L. Diamond, J-P. Spire, G.M. Zeitlin, and A.H. Gevins, "Automated Analysis of the Electrical Activity of the Human Brain (EEG): A progress report," Proceedings, IEEE, Vol. 63, No. 10, Oct. 1975, 1382-1399. Ge80 A.S. Gevins, "PAttern Recognition of Human Brain Electrical Potentials," IEEE Trans. Pattern Analysis and Machine Intelligence, Vol. 2, No. 5, Sept. 1980 383-404. Ge83 W.B. Gevarter, "Expert Systems: Limited But Powerful," IEEE Spectrum, Vol. 20, No. 8, Aug. 1983 39-44. Ha79 B. Hayes-Roth, H. Hayes-Roth, S. Rosenschein, and S. Cammarata, "Modeling Planning as an Incremental, Opportunistic Process," IJCAI, Vol. 6, 1979, 375-383!

PAGE 209

204 Ha83a Ha83b Ha85a J. Hasan, "Differentiation of Normal and Disturbed Sleep by Automatic Analysis," Turku, Finland, ACTA PHYSIOLOGICA SCANDINVICA Supplementum 526, 1983. F. Hayes-Roth, D.A. Waterman, and D.B. Lenat, Building Expert Systems, Addison-Wesley Publishing Co. Inc., Reading, Massachussetts, 1983. F . Hayes -Roth , Communications of 921-932. "Rule-Based Expert the ACM, Vol. 28, No. 9, Systems, " Sep. 1985, Ha85b B. Hayes-Roth, "Blackboard Architecture for Control," Journal of Artificial Intelligence, Vol. 26, 1985, 251-321. Ha86 W. Haustein, J. Pilcher, J. Klink, and H. Schulz, "Automatic Analysis Overcomes Limitations of Sleep Stage Scoring," El ec troencephy and Clinical Neurophysiology, Vol. 64, 1986, 364-371. Hj70 B. Hjorth, "EEC Analysis Based on Time Domain Properties," Electroenceph. Clin. Neyrophysiol . , Vol. 29, 1970, 306-310. Is75a A. Isaksson and A. Wennberg, "Visual Evaluation and Computer Analysis of the EEC; A Comparison, " Electroenceph. Clin. Neurophysiol . , Vol. 38 1975 79-86. Is75b A. Isaksson and A. Wennberg, "An EEC Simulator A Means of Objective Clinical Interpretation of EEC, " Electroenceph. Clin. Neurophysiol., Vol. 39 1975 313-320. Is81 Ja58 A. Isaksson, A. Wennerberg, and L.H. Zetterberg, "Computer Analysis of EEC Signals with Parametric Models," Proc. IEEE, Vol. 69, No. 4, Apr. 1981, 451461 . H.H. Jasper, "Report of the Committee on Methods of Clinical Examination in Electroencephalography, " Electroenceph. Clin. Neurophysiol., Vol. 10 370 1958. Ja71 Ja81 L. Jacobs, MD, M. Feldman, MD, and M.B. Bender, MD "Eye Movements During Sleep," Arch. Neurol., Vol'. 25' Aug. 1971, 151-159. B.H. Jansen, J.R. Bourne, .and J.W. Ward, "Spectral Decomposition of EEC Intervals Using Walsh and

PAGE 210

205 Fourier Transform, " IEEE Trans, 28, No. 12, Dec. 1981, 836-838. Biomed. Eng., Vol. Kt81 Ku80 La70 La80 Jo75 S.C. Johnson, "YACC: Yet Another Compiler-Compiler," Comp. Sci. Tech. Rep. No. 32, Bell Laboratories, Murray Hill, New Jersey, July 1975. P.Y. Ktonas and A. P. Gosalia, "Spectral Analysis vs. Period-Amplitude Analysis of Narrowband EEC Activity: A comparison based on the Sleep Delta Frequency Band," Sleep, Vol. 4, No. 2, 1981, 193-206. C.A. Kulikowski, "Artificial Intelligence Methods and Systems for Medical Consultation," IEEE Trans. PAMI, Vol. 2, No. 5, Sept. 1980, 464-476. L.E. Larson and D.O. Walter, "On Automatic Methods of Sleep Staging by EEC Spectra," Electroenceph . Clin. Neruphysiol . , Vol. 28, 1970, 459-467. H. Larsen and D.C. Lai, "Walsh Spectral Estimates with Applications to the Classification of EEG Signals," IEEE Trans. Biomed. Eng., Vol. 27, No. 9, Sep. 1980, 485-492. M.E. Lesk and E. Schmidt, "LEX A Lexical Analyzer Generator," Comp. Sci. Tech. Rep., No. 39, Bell Laboratories, Murray Hill, New Jersey, Oct. 1975. C.N. Lee, "Expert System Design and Implementation for Multichannel Sleep EEG Signal Processing," Ph.D. Dissertation, University of Florida, 1985. R. Lindsay, E.G. Buchanan, E.A. Feigenbaum, and J. Lederberg, DENDRAL, New York, McGraw-Hill, 1980. G. Logothetis, "A Table Generator for SLR(l) Parsers, " A. Lubin, L.C. Johnson, and M. Austin, "Discrimination among states of consciousness using EEG spectra," Psychophysiology, Vol. 6, 1969, 122. W.B. Martin, L.C. Johnson, S.S. Viglione, P. Naitoh, R.D. Joseph, and J.D. Moses, "Pattern Recognition of EEG-EGG as a Technique for All Night Sleep Scoring, " Electroenceph. Clin. Neurophysiol . , Vol. 32 1972 417-427. Ma76 J. Makhoul, "Linear Prediction: A Tutorial Review " Proc. IEEE, Vol. 63, April 1975, 561-580. Le75 Le85 Li80 Lo84 Lu69 Ma72

PAGE 211

206 Mc75 J. A. McEwen and G.B. Anderson, "Modeling of Stationarity and Gaussianity of Spontaneous EEG Activity," IEEE Trans. Biomed. Eng., Vol. 22, 1975, 361-368. Mc82 J. McDermott, "Rl: A Rule-Based Configurer of Computer Systems," Artificial Intelligence, Vol. 19, 1982, 39-88. Me81 V. Melle, System Aids in Constructing Consultation Programs, UMI Research Press, Ann Arbor, Michigan, 1981. Mi79 D. Michael and J Houchin, "Automatic EEG Analysis: A Segmentation Procedure Based on the Autocorrelation Function, "Electroenceph. Clin. Neurophysiol . , Vol. 46, 232-235. Ni86 H.P. Nii, "Blackboard Systems: The Blackboard Model of Problem Solving and the Evolution of Blackboard Architectures," The AI Magazine, Summer 1986, 38-53. Nu81a P.L. Nunez, "Electric Fields of the Brain," Oxford University Press, Oxford, New York, 1981, 349-400. Nu81b P.L. Nunez, "A Study of Origins of the Time Dependencies of Scalp EEG: I-Theoretical Basis," IEEE Trans. Biomed. Eng., Vol. 28, No. 3, Mar. 1981, 271280. Op75 Po75 Po77 Pr77 A.V. Oppenheim and R.W. Schafer, Digital Signal Processing, Prentice-Hall Inc., Englewood Cliffs, New Jersey, 1975. H.E. Pople, J.D. Myers, and R.A. Miller, "Dialog Internist: A Model of Diagnostic Logic for Internal Medicine," Proc. IJCAI, Vol. 4, 1975, 849-855. H. Pople, "The Formation of Composite Hypotheses in Diagnostic Problem Solving-An Exercise in Synthetic Reasoning." IJCAI, Vol. 5, 1977, 1030-103. H.M. Praetouries, G. Bodenstein, and O.D. Creutzfeldt, "Adaptive Segmentation of EEG records: A new approach to automatic EEG analysis," Electroenceph. Clin. Neurophysiol., Vol. 42 1977 84-94. Pr85 D.S. Prerau, "Selection of an Appropriate Domain for an Expert System, " The AI Magazine, Summer 1985 26-30. •9-

PAGE 212

207 Pr86 J.C. Principe and J.R. Smith, "Design and Implementation of Linear Phase FIR Filters for Biological Signal Processing," IEEE Trans. Biomed. Eng., Vol. 33, No. 6, June 1986, 550-559. Re68a A. Recht scha f f en and A. Kales, "A Manual of Standardized Terminology, Techniques and Scoring System for Sleep Stages of Human Subjects, " Public Health Service, U.S. Govermnent Printing Office, Washington D.C., 1968. Re68b A. Remond, "The Importance of Topographic Data in EEC Phenomena and an Electric Model to Reproduce Them, " Electroenceph. Clin. Neurophysiol . , Suppl. 27, 1968, 27-49. Ro72 P. Ross, "Computers in Medical Diagnosis," CRC Critical Review of Radiological Science, Vol. 3, 1972, 197-243. Sa80 A.C. Sanderson, J. Segen, and E. Richey, "Hierarchical Modeling of EEG Signals," IEEE Trans. Pattern Analysis and Machine Intelligence, Vol. 2, No. 5, Sep. 1980, 405-415. Sh73 E.H. Shortliffe, S.G. Axline, B.G. Buchanan, T.C. Merigan, and S.N. Cohen, "An Artificial Intelligence Program to Advise Physicians Regards Antimicrobial Therapy," Comput. Biomed. Res., Vol. 6, 1973 544-560. Sh76 E.H. Shortliffe, Computer-based Clinical Dicision Aids: MYCIN, Elsevier, New York, 1976. Sh79 E.H. Shortliffe, B.G. Buchanan, and E.A. Feigenbaum, "Knowledge Engineering for Medical Decision Making: A Review of Computer-based Clinical Decision Aids," Proc. IEEE, Vol. 69, Sep. 1979, 1207-1224. Sm71 J.R. Smith, M.J. Cronin, and I. Karacan, "A Multichannel Hybrid System for Rapid Eye Movement Detection (REM Detection)," Computers and Biomedical Research, Vo. 4, 1971, 275-290. Sm74 J.R. Smith, "Automated Analysis and Detection of EEG Spikes," IEEE Trans. Biomed. Eng., Vol. 21 No. 1 Jan. 1974, 1-7. Sm75 J.R. Smith, W.F. Funke, W.C. Yeo, and R.A. Ambuehl, "Detection of Human Sleep EEG Waveforms," Electroenceph. Clin. Neurophysiol., Vol. 38 1975 435-437.

PAGE 213

208 Sm78a J.R. Smith, I. Karacan, and M Analysis of the Human Sleep Sleeping, Vol. 2, 1978, 75-82. . Yang, "Automated EEG, " Waking and Sm78b J.R. Smith, "Computers in Sleep Research," Critical Reviews in Bioengineering, Vol. 3, Dec. 1978, 93-148. Sm79 J.R. Smith, I. Karacan, and M. Yang, "Automated Measurement of Alpha, Beta, Sigma, and Theta Burst Characteristics," Sleep, Vol. 1, No. 4, 1979. Sm81 W.D. Smith, "Walsh versus Fourier Estimators of the EEG Power Spectrum," IEEE Trans. Biomed. Eng., Vol. 28, No. 11, Nov. 1981, 790-793. Sm86 J.R. Smith, "Automated Analysis of Sleep EEG Data," Chapter 4, Handbook of Electroencephalography and Clinical Neurophysiology, Vol. 2, Elsevier Science Publishing Co., Inc., New York, 1986, 131-147. St82 M. Stefik, J. Aikins, R. Balzer, J. Benoit, L. Birnbaum, F. Hayes-Roth, and E. Sacerdoti, "The Organization of Expert Systems, A Tutorial," Artificial Intelligence, Vol. 18, No. 2, 135-173. Sz79 P. Szolovits and S.G. Pauker, "Computer and Clinical Decision-Making: Where, How, and for Whom?," Proc. IEEE, Vol. 67, 1979, 1224-1226. To75 H. Tong, " Autoregressive Model Fitting with Noisy Data by Akaike's Information Criterion," IEEE Trans. Inform. Theory, Vol. 21, July 1975, 476-480. Wa86 D.A. Waterman, A Guide to Expert Systems, AddisonWesley Publishing Co., Reading, Massachusetts, 1986. We70 P.C. Welch, "The Use of Fast Fourier Transform for the Estimation of Power Spectra," IEEE Trans. Audio Electroacoust. , Vol. 15, June 1970, 70-73. We78a H. Weiberg and R. Cooper, "The Recognition Index: A Pattern Recognition Technique for Noisy Signals," Electroenceph. Clin. Neurophysiol . , Vol. 33, 1972 608-613. We78b S.M. Weiss, C.A. Kulikowski, S. Amarel,and A. Safir, "A Model-based Method for Computer-aided Medical Decision-making," Artificial Intelligence, Vol. 11 1978, 145-172. Wi74 R.L. Williams, I. Karacan, and C.J. Hursch,

PAGE 214

209 Electroencephalography (EEG) of Human Sleep: Clinical Applications, John Wiley and Sons, New York, 1974. W.C. Yeo and J.R. Smith, "Walsh Power Spectra of Human Electroencephalogram," In Proc. Appl . Walsh Functions, 1972, 159-162.

PAGE 215

BIOGRAPHICAL SKETCH Tae Gyu Chang was born on November 13, 1955, in Seoul, Korea. He received his bachelor's degree in electrical engineering from the Seoul National University, Seoul, Korea, in 1979. He received his Master of Science degree in electrical engineering from the Korea Advanced Institute of Science and Technology, Seoul, Korea, in 1981. He worked for Hyundai Engineering Company and Modern Electrosystems , Inc., for three years as a control and computer systems design engineer. He has been studying for a Ph.D. in electrical engineering at the University of Florida since August 1984. 210

PAGE 216

I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. Jack R. Smith, Chairman Professor of Electrical Engineering I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. Donald G. Childers Professor of Electrical Engineering I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. Antonio A. Arroyo Associate Professor of Electrical Engineering I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. S.Y. Oh Assistant Professor of Electrical Engineering

PAGE 217

I certify that I have read this study and that in my opinion it conforms to acceptable standards of scholarly presentation and is fully adequate, in scope and quality, as a dissertation for the degree of Doctor of Philosophy. George Logothetis Assistant Professor of Computer and Information Sciences This dissertation was submitted to the Graduate Faculty of the College of Engineering and to the Graduate School, and was accepted as partial fulfillment of the requirements for the degree of Doctor of Philosophy. May 1987 Dean, College of Engineering Dean, Graduate School