<%BANNER%>

A Memory-Optimized Architecture for ECG Signal Processing

Permanent Link: http://ufdc.ufl.edu/UFE0042752/00001

Material Information

Title: A Memory-Optimized Architecture for ECG Signal Processing
Physical Description: 1 online resource (104 p.)
Language: english
Creator: PENG,CHUNG-CHING
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2011

Subjects

Subjects / Keywords: Electrical and Computer Engineering -- Dissertations, Academic -- UF
Genre: Electrical and Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Given an increasing demand for wearable personal healthcare devices, automatic heart beat monitoring device along with wireless interconnection is vital for providing real-time heart diagnostics and services. This dissertation investigates low-cost, automatic real-time architecture for electrocardiogram (ECG) beat classification. Based on the nature shape of human heart beat, Euclidean distance is used to detect and cluster the heart beats. The proposed architecture using a Heart Beat Cache, establishes a small table to dynamically capture and record recent heart beats. The information recorded in the cache provides necessary matching attributes to clustering the heart beats. To adapt the dynamic change of heart beat shapes, proper merging and aging of the recorded heart beats in cache increase the detecting accuracy as well as reduce the space requirement. With a small number of recorded heart beats, the proposed Heart Beat Cache consumes low power and can fit into small medical devices.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by CHUNG-CHING PENG.
Thesis: Thesis (Ph.D.)--University of Florida, 2011.
Local: Adviser: Peir, Jih-Kwon.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2011
System ID: UFE0042752:00001

Permanent Link: http://ufdc.ufl.edu/UFE0042752/00001

Material Information

Title: A Memory-Optimized Architecture for ECG Signal Processing
Physical Description: 1 online resource (104 p.)
Language: english
Creator: PENG,CHUNG-CHING
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2011

Subjects

Subjects / Keywords: Electrical and Computer Engineering -- Dissertations, Academic -- UF
Genre: Electrical and Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Given an increasing demand for wearable personal healthcare devices, automatic heart beat monitoring device along with wireless interconnection is vital for providing real-time heart diagnostics and services. This dissertation investigates low-cost, automatic real-time architecture for electrocardiogram (ECG) beat classification. Based on the nature shape of human heart beat, Euclidean distance is used to detect and cluster the heart beats. The proposed architecture using a Heart Beat Cache, establishes a small table to dynamically capture and record recent heart beats. The information recorded in the cache provides necessary matching attributes to clustering the heart beats. To adapt the dynamic change of heart beat shapes, proper merging and aging of the recorded heart beats in cache increase the detecting accuracy as well as reduce the space requirement. With a small number of recorded heart beats, the proposed Heart Beat Cache consumes low power and can fit into small medical devices.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by CHUNG-CHING PENG.
Thesis: Thesis (Ph.D.)--University of Florida, 2011.
Local: Adviser: Peir, Jih-Kwon.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2011
System ID: UFE0042752:00001


This item has the following downloads:


Full Text

PAGE 1

1 A MEMORY OPTIMIZED ARCHITECTURE FOR ECG SIGNAL PROCESSING By CHUNG CHING PENG A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCT OR OF PHILOSOPHY UNIVERSITY OF FLORIDA 201 1

PAGE 2

2 201 1 Chung Ching Peng

PAGE 3

3 I dedicate this dissertation t o my parents Chien Feng Peng and Chiu Chen Cheng and my loving wife Hsiao Hsiang Hsu

PAGE 4

4 ACKNOWLEDGMENTS I thank my advisor Dr. Jih Kwon Peir for his unlimited support and encouragement both professionally and personally starting even before I came to USA. Under his guidance, I learned so much about research and academia in general and specifically computer architectur e. During the darkest time of my journey, he reminded me to remain optimistic and n ever give up. I am eternally grateful. My committee Dr. Harris, Dr. Wheeler, and Dr. Li, helped me throughout my research by providing valuable feedback. I am especially gra s support during the difficult time when I change d my research direction. Many labmates and friends in both CISE and ECE have helped me throughout the long period of time I spent at UF. Seonho Huang who graduated in 2007 taught me e verything about ASIC design flow. Dr. Hong Yu who graduated in 2010 showed me the power of persistence on many occasions Pawan Sabharwal, who was the co author of many of my publications, helped me revise paper draft s and drew some of my chip layout. Zhim ing Xiao, Chun Ming Tang, Deepak Bhati a, Qiuzhong Wu, and Tanuj Aggarwal have helped me complete several chip designs and testing

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ ............... 4 LIST OF TABLES ................................ ................................ ................................ ........................... 7 LIST OF FIGURES ................................ ................................ ................................ ......................... 8 ABSTRACT ................................ ................................ ................................ ................................ ... 11 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .................. 13 1.1 The ECG Signal ................................ ................................ ................................ ................ 15 1.2 Arrhythmia Recognition ................................ ................................ ................................ ... 17 1.3 Current State of the Art in Automatic Beat Classifier ................................ ..................... 18 1.4 Motivation ................................ ................................ ................................ ......................... 19 1.5 Dissertation Structure ................................ ................................ ................................ ....... 20 2 QRS COMPLEX DETECTION ................................ ................................ ............................. 22 2.1 QRS Complex ................................ ................................ ................................ ................... 22 2.2 QRS Detection System ................................ ................................ ................................ ..... 24 2.2.1 Preprocessing Feature Extraction Stage ................................ ............................... 24 2.2.2 Decision Stage ................................ ................................ ................................ ...... 26 2.3 MIT BIH Arrhythmia Database ................................ ................................ ....................... 28 2.3.1 Introduction ................................ ................................ ................................ .......... 28 2.3.2 Annotations ................................ ................................ ................................ .......... 29 2.4 Matlab Simulations Results ................................ ................................ .............................. 29 2.4.1 Simulation Setup ................................ ................................ ................................ .. 29 2.4.2 Detection Accuracy ................................ ................................ .............................. 30 2.4.3 Discussions ................................ ................................ ................................ ........... 31 3 BEAT CLASSIFICATION ................................ ................................ ................................ .... 34 3.1 Types of Arrhythmia ................................ ................................ ................................ ......... 34 3.2 Clustering Analysis ................................ ................................ ................................ ........... 38 3.2.1 Distance Distribution ................................ ................................ ........................... 38 3.2. 2 Principal Component Analysis ................................ ................................ ............. 47 3.3 Morphological Variation ................................ ................................ ................................ .. 55 4 ARCHITECTURE ................................ ................................ ................................ .................. 56 4.1 Computational Model ................................ ................................ ................................ ....... 56 4.2 Clustering Architecture ................................ ................................ ................................ ..... 57

PAGE 6

6 4.2.1 Cache ................................ ................................ ................................ .................... 57 4.2.2 Reuse Distance of Detected Heart Beats ................................ .............................. 59 4.3 Heart Beat Cache ................................ ................................ ................................ .............. 74 4.3.1 Cache Organization ................................ ................................ .............................. 74 4.3.2 Cache Operations ................................ ................................ ................................ 74 4.3.3 Power Management ................................ ................................ .............................. 76 4.4 Matlab S imulation Results ................................ ................................ ................................ 76 4.4.1 Simulation Setup ................................ ................................ ................................ .. 76 4.4.2 Evaluation of Classification Performance ................................ ............................ 77 4.4.3 Normal Beats ................................ ................................ ................................ ........ 78 4.4.4 Premature Ventricular Contraction ................................ ................................ ...... 78 4.4.5 Other Abnormal Beats ................................ ................................ .......................... 78 4.4.6 Cache Size Usage ................................ ................................ ................................ 79 4.5 Discussions ................................ ................................ ................................ ....................... 79 5 HARDWARE IMPLEMEN TATION ................................ ................................ .................... 88 5.1 ECG System ................................ ................................ ................................ ...................... 88 5.2 RTL Implementation ................................ ................................ ................................ ........ 89 5.2.1 QRS Detector ................................ ................................ ................................ ....... 89 5.2.2 Beat Classifier ................................ ................................ ................................ ...... 90 5.2.3 Functional Simulations ................................ ................................ ........................ 91 5.3 FPGA Implementation ................................ ................................ ................................ ...... 93 5.3.1 Synthesis ................................ ................................ ................................ .............. 93 5.3.2 Memory ................................ ................................ ................................ ................ 93 5.4 Optimizations ................................ ................................ ................................ .................... 94 5.4.1 FIFO size ................................ ................................ ................................ .............. 94 5.4.2 Cluster Size ................................ ................................ ................................ .......... 94 5.4.3 Number of Templates ................................ ................................ ........................... 94 6 CONCLUSION ................................ ................................ ................................ ....................... 96 6.1 Conclusive Comments ................................ ................................ ................................ ...... 96 6.2 Future Lines of Research ................................ ................................ ................................ .. 97 6.2.1 Explore More Databases ................................ ................................ ...................... 97 6.2.2 Intervals and Waves ................................ ................................ ............................. 97 6.2.3 Heart Rate Variability ................................ ................................ .......................... 97 LIST OF REFERENCES ................................ ................................ ................................ ............... 98 BIOGRAPHICAL SKETCH ................................ ................................ ................................ ....... 104

PAGE 7

7 LIST OF TABLES Table page 2 1 Overview of the QRS detection algorithms in the literature ................................ .............. 22 2 2 Patient records in the MIT BIH Arrhythmia Database ................................ ...................... 29 2 3 Heart beat annotations appearing in the MIT BIH Arrhythmia Database ......................... 30 2 4 QRS detection results using the Pan Tompkins algorithm (Records #100 #124) ........... 32 2 5 QRS detection results using Thresholding method (Records #100 #124). ...................... 32 3 1 O verview of annotated beat types ................................ ................................ ...................... 35 4 1 Clus tering results for Normal beats ................................ ................................ ................... 80 4 2 Cluste ring results for Premature Vent ricular Contraction (PVC) beats ............................. 82 4 3 Clustering results for Paced beats ................................ ................................ ...................... 84 4 4 Clustering resul ts for Left B undle Branch Block (LBBB) beats ................................ ....... 84 4 5 Clustering results for Right B undle Branch Block (RBBB) beats ................................ ..... 84 4 6 Cluster ing results for f usion of Paced and Normal beats ................................ ................... 84 4 7 Number of cache entries used d uring the simulation ................................ ......................... 85 4 8 Performance co mparison results for two beat classification algorithms ........................... 86 5 1 FPGA synthesis report ................................ ................................ ................................ ....... 93

PAGE 8

8 LIST OF FIGURES Figure page 1 1 Block diagram of a wireless vital signs monitor system. ................................ ................... 14 1 2 Schematic representation of normal ECG signal in sinus rhythm ................................ .... 15 1 3 Nomenclature for different morphologies of QRS complexes ................................ .......... 16 1 4 Normal heart beat. ................................ ................................ ................................ .............. 17 1 5 Examples of arrhythmia. ................................ ................................ ................................ .... 18 2 1 Basic flow diagram for Thresholding and Pan Tompkins algorithms ............................... 25 2 2 Block diagram of the b and pass filter ................................ ................................ ................ 26 2 3 Matlab simulation results of the preprocessing stage using MIT BIH Arrhythmia Database ................................ ................................ ................................ ............................ 27 2 4 Sample ECG plot in MIT BIH Arrhythmia Database ................................ ........................ 28 2 5 Performance comparison using MIT BIH Arrhythmia Database ................................ ..... 33 3 1 Normal beat, PVC, an d APB ................................ ................................ ............................. 36 3 2 LBBB, RBBB, and Paced Beat ................................ ................................ .......................... 37 3 3 Snapshots of abnormal beats using the MIT BIH Arrhythmia Database .......................... 38 3 4 The distance distribution plots simulated using records #100 to #124 in the MIT BIH Arrhythmia Database ................................ ................................ ................................ ......... 40 3 5 The distance distribution plots simulated using records #200 to #234 in the MIT BIH Arrhythmia Database ................................ ................................ ................................ ......... 43 3 6 The PCA plots for records from #100 to #124 in the MIT BIH Arrhythmia Database ..... 49 3 7 The PCA plots for records from #200 to #234 in the MIT BIH Arrhythmia Database ..... 52 3 8 Representative beats during a stress test ................................ ................................ ............ 55 4 1 A generalized computational model for the bi omedical signals classification .................. 57 4 2 Analogy between no rmal cache and Heart Beat Cache ................................ ..................... 58 4 3 An illustration of the definition of r euse distance in cache behavior ................................ 60

PAGE 9

9 4 4 The reuse distance plots from record #100 to #124 in the MIT BIH Arrhythmia Database ................................ ................................ ................................ ............................ 61 4 5 The reuse distance plots simulated using records #200 to #234 in the MIT BIH Arrhythmia Database ................................ ................................ ................................ ......... 65 4 6 The reuse distance plots simulated using record #208 in the MIT BIH Arrhythmia Database ................................ ................................ ................................ ............................ 69 4 7 The reuse distance plots simulated using record #221 in the MIT BIH Arrhythmia Da tabase ................................ ................................ ................................ ............................ 71 4 8 The reuse distance plots simulated using record #228 in the MIT BIH Arrhythmia Database ................................ ................................ ................................ ............................ 72 4 9 The distribution of the summed reuse distance for all the detected heart beats in the MIT BIH Arrhythmia Database ................................ ................................ ......................... 73 4 10 System block diagram and detaile d Heart Beat Cache organization ................................ 75 4 11 Heart beats samples taken from record #207 in MIT BIH Arrhythmia Database ............. 75 4 12 Simulation setup illustration using the record #106 from MIT BI H Arrhythmia Database ................................ ................................ ................................ ............................ 77 4 13 Clustering performance results for normal beats. A) Records #100 t o #123. B) Records #200 to #234 ................................ ................................ ................................ ........ 81 4 14 Clustering performance results for PVCs. A) Records #100 to #119. B) Records #200 to #233. ................................ ................................ ................................ ..................... 83 4 15 Number of used cache entries for 48 data sets ................................ ................................ ... 86 4 16 Performance comparison for two beat classification algorithms ................................ ....... 87 5 1 Sin gle ECG processor block diagram ................................ ................................ ................ 88 5 2 Detai led ECG processor block diagram ................................ ................................ ............. 89 5 3 Single channel p rocessor datapath block diagram ................................ ............................. 90 5 4 Timing diagra m for the processor pipeline datapath and the detailed Heart Beat Cache read/write oper ations of the memory controller ................................ ...................... 91 5 5 Function simulation wave forms for QRS detector in NCSim ................................ ........... 92 5 6 Function simulation wavefor ms for beat classifier in NCSim ................................ ........... 92 5 7 ZBT SRAM interface ................................ ................................ ................................ ......... 93

PAGE 10

10 5 8 Typical heart beat waveform ................................ ................................ ............................. 95 6 1 Sample plot for APB ................................ ................................ ................................ .......... 97

PAGE 11

11 Abstract of Dissertation Presented to the Graduate School of the Universi ty of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy A MEMORY OPTIMIZED ARCHITECTURE FOR ECG SIGNAL PROCESSING By Chung Ching Peng May 2011 Chair: Jih Kwon Peir Major: Electrical and Computer En gineering Given an increasing demand for wearable personal healthcare devices, automatic heart beat monitoring device along with wireless interconnection is vital for providing real time heart diagnostics and services. This dissertation investigates low c ost, automatic real time architecture for electrocardiogram (ECG) beat classification. Based on the nature shape of human heart beat, Euclidean distance is used to detect and cluster the heart beats. The proposed architecture using a Heart Beat Cache, esta blishes a small table to dynamically capture and record recent heart beats. The information recorded in the cache provides necessary matching attributes to clustering the heart beats. To adapt the dynamic change of heart beat shapes, proper merging and agi ng of the recorded heart beats in cache increase the detecting accuracy as well as reduce the space requirement. With a small number of recorded heart beats, the proposed Heart Beat Cache consumes low power and can fit into small medical devices. The prop osed architecture was simulated in Matlab using the MIT BIH Arrhythmia Database We found that all the major beat types were successfully clustered and recognized. Based on the simulation results, the classification of normal beats shows consistent perform ance (89.92%). Among all abnormal beat types, premature ventricular contractions (PVCs) yield more

PAGE 12

12 inconsistent results due to their larger shape variation and the p ossibility of PVCs in some records containing more than one shape type. Finally, the RTL implementation is developed and described. A cycle based timing diagram for accessing the He art Beat Cache is also analyzed

PAGE 13

13 CHAPTER 1 INTRODUCTION The demand for w earable personal healthcare devices has recently risen significantly due to the advances in biomedical engineering, nanotechnology and microelectronics and also due to the need for timely access to healthcare, medical information and medical resources We arable devices allow the monitoring of vital signs at home, and enable wirelessly transmitting of physiological data to a remote diagnostic center. Vital signs could include an e lectrocardiography (ECG) data, heart rate, temperature (skin or ambient) brea thing rate and both diastolic and systolic blood pressure. Studies have shown that t he number of people suffering from heart disease is rising [1]. In this work, we focus on monitoring the ECG signal, which is a cruc ial tool for cardiologists to diagnose the condition of a heart For patients with heart problems, the occurrence of irregular heartbeat s (arrhythmia) is an indication of potential heart failure requiring immediate treatment. Prevention of heart failure relies on constant tracking of the patien professional diagnosis based on those measurements in timely manner. Traditional ECG holter systems record the signals continuously for to only 48 hours maximum, and only then the data is retrieved for analysis. Any events occurring outside these monitoring periods are missed. Also, the data is analyzed only after a considerable delay. To achieve a faster diagnosis and treatment, any cardiac event recorded should be forwarded automatically over the internet to physicians wh o have been immediately alerted by the system. Thus, it is necessary to design a system to recognize arrhythmia characteristics in real time. The identification of an arrhythmia event involves detecting heart beats automatically, discriminating abnormal be ats from normal beats, and identifying the type of arrhythmia. Therefore, wireless

PAGE 14

14 monitoring devices should integrate functions of signal sensing, real time ECG signal processing and telemetry to other centralized devices that are connected to network F igure 1 1 demonstrates a next generation wireless vital sign monitoring system. Devices can be either integrated with adhesive ECG sensor pads that are attached to the chest or implanted underneath the skin for measuring additional signals (such as blood p ressure records). To reduce the power consumption for wireless transmission, bandwidth reduction, which only transmits the most important results, is required. The extracted signals can then be transmitted wirelessly to a larger handheld device containing more complex algorithms and wireless turn forwards necessary extracted information to the surrounding wireless network. Devices targeted for implantation hav e additional requirements for safety, reliability, and cost effectiveness Most importantly the devices need to have ultra low power dissipation for a long battery lifetime and minimum size for wearability. Figure 1 1. Block diagram of a wireless vital signs monitor system.

PAGE 15

15 1.1 The ECG Signal The electrocardiogram (ECG) measures the electrical activity of the heart, and is probably the most representative and noninvasive test to diagnose cardiac disorders. As Figure 1 2 shows, the ECG signal consists of a number of diagnostic critical segments or waves: the P wave, the QRS complex, and the T wave [2 5]. The P wave is generated during the atrial depolarization, which occurs when the blood is squeezed from the atria to the ventricles. Figure 1 2. Sche matic representation of normal ECG signal in sinus rhythm. Reproduced from [3]. The QRS complex, which can be decomposed into Q R and S wave segments, is generated when the ventricles depolarize and squeeze the blood from the right ventricle to the aor ta. The QRS complex has a period between 80ms and 120ms (longer than P wave). However, not every QRS complex actually contains a Q wave, an R wave, and an S wave; from Figure

PAGE 16

16 1.3, it is clear that there are many variations of a QRS complex. The problem of QRS complex detection has been studied for several decades. A review of existing methods can be found in [ 6 ]. Finally, the T wave is generated during the period of time when the ventricles repolarize in preparation for the next heart beat. Figure 1 3. N omenclature for different morphologies of QRS complexes. Reproduced from [3]. Critical information about the heart is obtained by studying the shape and size of the waves, the time between each wave, and the beat rate. In addition to these three basic wave forms, other segments and certain time intervals shown in Figure 1 2 also contain important information about heart conditions. The ST segment lasts from the end of the S wave to the beginning of the T wave, and corresponds to the period of ventricular dep olarization. The PR segment represents the atrioventricular conduction time, which is used as a reference for

PAGE 17

17 comparison with the ST segment. The QT interval represents the time of ventricular systole, and is measured between the beginnings of the Q wave t o the end of the T wave [7]. 1.2 Arrhythmia Recognition An arrhyth mia is an irregular heartbeat or abnormal heart rhythm and real time diagnosis is crucial for high risk cardiac patients. There are four main types of arrhythmias premature beats, supraven tricular arrhythmias, ventricular arrhythmias, and bradyarrhythmias [7]. Premature beats is the most common type of arrhythmias (Figure 1.5) and can be categorized into premature ventricular contraction (PVC) and atrial premature beat (APB), depending on t heir origin. With ECG data, PVC is characterized by abnormally shorter spacing between subsequent waves, while APB can usually be characterized by an early P wave. In rare cases, a n APB, just like a PVC can trigger a more serious arrhythmia such as an atr ial flutter or atrial fibrillation. Figure 1 4. Normal heart beat [8]. Supraventricular arrhythmia is a fast heart rate that originates in the upper chambers (atria). The most common are atrial fibrillation or flutter and atrioventricular n odal reentry tachycardia Likewise, ventricular arrhythmia is a fast heart rate or iginating in the lower chambers ( ventric les) B radyarrhythmia is a slow heart rate due to problems with the SA node's pacemaker ability, or else an i nterruption in electrical conduction t hrough the natural electrical pathways of the heart.

PAGE 18

18 Figure 1 5. Examples of arrhythmia [8]. 1. 3 Current State of the Art in Automatic Beat Classifier Armato et al have proposed a feasibility study and development of an arrhythmia recognition syste m using Kohenen Self Organizing Map, a neural network method [ 2 ]. The algorithm was implemented into Xilinx Virtex 4 FX12, and used 90% of the FPGA resource. One drawback of their architecture is the large memory space required (estimated to be 73 KB). Ano ther disadvantage is the need for offline training. Approximately 70% of the data sets from the MIT BIH Arrhythmia Database were used for training, and the remainder was used for classification. Because training is not always an option for long term vital signs monitoring in the home environment, another solution needs to be found.

PAGE 19

19 Freescale Semiconductor and Monebo Technologies have jointly proposed an on a with t he Monebo ECG software and Freescale [10 ]. The Monebo algorithm is able to handle a wide range of ECG applications, and targets various commercial platforms from workstations to microcontrollers. The minimum memory size required is 18 KB. In the hierarchy of next generation vital signs monitorin g systems, this solution can serve as part of the centralized processing unit in a handheld device, and thus provide in depth analysis whenever abnormal beats are detected. Besides the above mentioned hardware approaches, there are many existing software a lgorithms for automatic beat classification, including linear discriminants [11], back propagation neural networks [12], and learning vector quantization [13]. However, they are usually used for the clinical instruments, where the high computation complexi ty and the large memory capacity are available. 1.4 Motivation Although clinical instrument allows more sophisticated algorithms to improve the accuracy, it is becoming increasingly necessary to use the portable medical devices with stringent power constra int This is because rising healthcare costs combined with a more longer living population has created demand for new applications an d therapies relying on portable medical devices that are wirelessly linked to base stations However, there are several cha llenges in the development of automatic beat classification algorithms and their implementations, the most critical one being the trade off between algorithmic c omplexity and power consumption. The leakage power becomes increasingly significant in deep sub micron technologies. With the low operating frequency typical for ECG processing applications the dynamic power is not as domin ant as the leakage power. Static random access memory (SRAM) often contributes a

PAGE 20

20 large part to the leakage power consumption; t herefore a memory optimized architecture is essential to reduce the total power consumption. Another challenge of automatic heart beat classification is the variability in ECG readings between people, and even different measurement form the same person. A lthough different types of arrhythmias have characteristics which are common to all patients, a training process is still required for most of the algorithms to accommodate the inherent variability. Still, an unsupervised ECG monitoring system is preferred for the home environment. The goal of this work is thus to design a memory optimized architecture for next generation wireless ECG telemetry, that does not require supervised training. In the previous sections we have seen that a digital ECG processor pl ays an important role in wearable healthcare devices. The cornerstone of this work is to transform complicated beat classification algorithms (usually implemented in software) into hardware friendly operations. This study has three aims. First, we seek to develop a low power, real time online digital beat classification architecture that not only achieves transmission bandwidth as low as the QRS detector, but also enables on chip real time ECG data analysis. Second, we investigate the nature of ECG signals and outline the required hardware properties for a fully unsupervised and automated ECG processor implementation. Third, we attempt to provide an ECG signal processing platform capable of partially identifying heart failure symptoms based on beat classific ation. 1.5 Dissertation Structure The dissertation proposal is organized as follows. Chapter 2 provides the introduction and performance comparison for both fixed threshold algorithm and Pan Tompkins algorithm for QRS detection

PAGE 21

21 Chapter 3 observe s the clu stering properties of heart beats through analytical tools such as Euclidean distance distribution and PCA. Then, the temporal properties of the heart beat shape changes is demonstrated Chapter 4 explore s energy saving architectures for online beat classi fication that do not sacrifice performance. A low power architecture for real time online beat classification is proposed. It does not require any training period and has the capab ility to quickly respond to changing beat shapes Chapter 5 presents a n imp lementation of a single channel self contained ECG processing architecture which consists of the QRS spike detector and beat classifier presented in Chapters 2 and 3. The proposed VLSI a rchitecture was implemented in register transfer l evel (RTL) and funct ionally verified using ECG data sets. Chapter 6 gives conclusive comments and the direction for future works that can further improve the performance of the proposed single channel ECG processor architecture.

PAGE 22

22 CHAPTER 2 QRS COMPLEX DETECTIO N 2 .1 QRS Complex For an automatic ECG diagnosis algorithm, retrieving characteristic points is essential. The QRS complex is the most significant characteristic waveform set of the cardiac cycle and corresponds to the depolarization of the ventricles in the heart. The position, duration and shape of the QRS complex provide important information about the performance of the heart. T he accuracy of a QRS complex detection algorithm is determined by the number of correctly detected beats (as opposed to false posi tives or missed beats). False detections can result in an incorrect diagnosis (such as heart rate variability). This is a significant problem since ECG recordings may contain noise, sudden changes in QRS amplitudes, and muscle artifacts [32]. Many algorith ms have been proposed for QRS complex detection; an overview and comprehensive performance evaluations are described in [6]. General approaches employ techniques such as threshold crossing, single derivatives, digital filters, adaptive and matched filters, template correlation, frequency domain analysis, maximum a posteriori (MAP) estimation methods, and the hidden Markov model. For real time analysis, derivative based algorithms, with relatively inexpensive computations, are often preferred. These algorith ms do not require Table 2 1. Overview of the QRS detection algorithms in the literature [6]. Algorithm Types Most Referred Works (Software) Implementation (Hardware) Computation Complexity Threshold Crossing Miller et al (1971) Zemva (2007) -Xilinx Vir tex II Low Derivative Based Pan Tompkins (1985) Hamilton Tompkins (1986) Shukla (2008) -Xilinx Spartan Low Wavelet Based Mallat & Hwang (1992) In et al .(2008) -Xilinx SysGen Medium Neural Network Hu et al .(1993) De Rossi(2009) -Xilinx Virtex 4 High Hidden Markov Models Briller et al .(1990) High

PAGE 23

23 training or modifications specific to each patient. Other algorithms borrow ideas from pattern recognition and neural network field [33]. More recently, wavelet transform methods have been used for QRS detection in recent years [11], which exploits the time frequency properties of each ECG wavelet. The disadvantages of wavelet methods are poor performance due to the sensitivity to the onset and the offset step. Real time QRS detector can be consider ed to be the first step toward an automatic ECG analysis system. In this work, we seek to adopt a QRS detection algorithm from the literature with low hardware complexity and reasonable detection accuracy. Table 2 1 categorized different QRS detection appr oaches including the representative works as well as their hardware implementation. Only until recent years, people started to implement these software algorithms into hardware [ 6 0 6 1]. It is clear that threshold crossing and derivative based approaches ha ve lower hardware complexity compared to others. Therefore, the purpose of this chapter is to compare the performance of the representative works from these two approaches and discuss the tradeoff between performance and hardware cost. In threshold crossin g approach, we used an algorithm Thresholding with the threshold being the mean value of the signals added with 3 times of standard deviation value. In derivative based approach, we used a well established QRS detection algorithm Pan Tompkins proposed by Pan and Tompkins in 1985 [43]. The Pan Tompkins algorithm has a high accuracy for various beat morphologies, and has been widely used in the past decades [6]. The MIT BIH Arrhythmia Database is a popular choice for evaluation of QRS complex detection algo rithms because of its ease of use and popularity [ 5 9]. We will use the same ECG data sets from this database to simulate both the Thresholding and Pan Tompkins algorithms in Matlab. Then, the performance comparison will be carried out.

PAGE 24

24 2.2 QRS Detection Sy stem For both Thresholding and Pan Tompkins algorithms, there are two stages in the algorithmic structure: a preprocessing feature extraction stage and a decision stage. The first stage includes linear filtering and a nonlinear transformation, while the s econd stage includes peak detection and a set of decision logic for QRS detection. A basic flow diagram is given in Figure 2.1. 2.2.1 Preprocessing Feature Extraction Stage During the preprocessing stage the signal is filtered, its derivative is calculated it is squared and then integrated by a moving window algorithm. The preprocessing formulas listed in this section are taken from the Pan Tompkins algorithm [43], as they are commonly used signal processing techniques. The output waveforms from each proce ssing step are generated using Matlab and shown in Figure 2.3. A heart beat is defined between two R waves. The typical frequency range of the QRS complex is from 10 Hz to 25 Hz. To suppress other signal components and noise sources (such as power line int erference), a band pass filter is required prior to the actual detection. The band pass filter consists of a high pass filter and a low pass filter. The high pass filter reduces the presence of P and T waves and baseline drift, while the low pass filter removes artifacts. The transfer function of the s econd order low pass filter is and its difference equation is The cutoff frequency is about 20Hz and the gain is 36, and the filter processing delay is 6 samples.

PAGE 25

25 The transfer f unction of the high pass filter is while its difference equation is The low cutoff frequen cy of this filter is about 5 Hz and the gain is 32. Figure 2 1. Basic flow diag ram for Thresholding and Pan Tompkins algorithms. The differentiator further suppresses the low frequency P and T waves, and provides a large gain to the high frequency components arising from the steep slopes of the QRS complex. The transfer function of the five point differentiator is Its difference equation is Squaring is a nonlinear operation that forces a positive result and emphasizes large differences resulting from the QRS complexes. In addition, small differences from P and T waves are suppressed.

PAGE 26

26 Multiple peaks are generated from the previous derivative based operation within the duration of a single QRS complex. Integration smoothes these peaks with a moving wind ow integration filter. 2.2.2 Decision Stage For the Threhsolding algorithm, the decision stage simply compares the band pass filtered signal with the previously calculated threshold, which is defined to be the mean value of the signals added with three times of standard deviation value. For the Pan Tompkins algorithm, there are more steps involved. The peak detector compares the signal with two adaptive thresholds (T1 and T2), which are updated after each detected beat based on some criteria. T1 is set for the filtered ECG, and T2 is set for the signals produced by the moving window integration. Thresholds T1 and T2 are continuously updated after every beat to adapt to the rises and falls of the R wave pea k amplitudes. Processing delays are considered when estimating the average R R intervals. If an R wave is not detected after the maximum time interval, the algorithm will repeat the search to find a possible R wave candidate by using a new set of lower thr esholds. More details are given in [43] about how to configure the two thresholds and implement the decision rules algorithms using a Matlab simulation. Figure 2 2. Block diagram of the band pass filter.

PAGE 27

27 Figure 2 3. Matlab simulation results of t he preprocessing stage using MIT BIH Arrhythmia Database From top to bottom: two cycles of a ECG signal; output of the low pass filter; output of high pass filter; output of the derivative based operator; output of the squared signal; and the result of th e final integrator.

PAGE 28

28 2.3 MIT BIH Arrhythmia Database Assessing the quality of different QRS detection algorithms is often difficult due to the lack of an ECG data set with known annotations, which label the characteristics of the actual heart beats. 2.3.1 I ntroduction The MIT BIH Arrhythmia Database is a collection of holter recordings obtained by the Beth Israel Hospital Arrhythmia Laboratory between 1975 and 1979. The patient records are listed in Table 2 2. A sample plot from the database is shown in Figu re 2.4. The recordings were taken from 48 patients, who were 25 men (aged 32 to 89 years) and 23 women (aged 23 to 89 years). Each recording is 30 minutes long with a sampling rate of 360 Hz and 11 bit resolution over a 10 mV range. The MIT BIH Database is frequently used for evaluating the performance of QRS complex detection and beat classification algorithms. We used this database because it is widely used (more than 500 citations) and easier to compare the results with other works. Figure 2 4. Sample ECG plot in MIT BIH Arrhythmia Database

PAGE 29

29 Table 2 2. Patient records in the MIT BIH Arrhythmia Database (*) indicates Paced Beats; (**) indicates Left Bundle Branch Block; (***) indicates Right Bundle Branch Block. Record Record Patient Normal / Paced / B BB Special (PVC ..etc) Record Record Patient Normal / Paced / BBB Special (PVC ..etc) 100 Male, age 69 2239 34 200 Male, age 64 1743 861 101 Female, age 75 1860 5 201 Male, age 68 1625 375 102 Female, age 84 2028(*) 4 202 Male, age 68 2061 75 103 Male, age N/A 2082 2 203 Male, age 43 2529 451 104 Female, age 66 1380(*) 20 205 Male, age 59 2571 85 105 Female, age 73 2526 46 207 Female, age 89 1543(**) 789 106 Female, age 24 1507 520 208 Female, age 23 1586 1369 107 Male, age 63 2078(*) 59 209 Male, a ge 62 2621 384 108 Female, age 87 1740 35 210 Male, age 89 2423 227 109 Male, age 64 2492(**) 40 212 Female, age 32 2748(***) 0 111 Female, age 47 2123(**) 1 213 Male, age 61 2641 610 112 Male, age 54 2537 2 214 Male, age 53 2003(**) 259 113 Female, a ge 24 1789 6 215 Male, age 81 3195 168 114 Female, age 72 1820 59 217 Male, age 65 1786(*) 422 115 Female, age 39 1953 0 219 Male, age N/A 2082 205 116 Male, age 68 2302 110 220 Female, age 87 1954 94 117 Male, age 69 1534 1 221 Male, age 83 2031 396 118 Male, age 69 2166(***) 122 222 Female, age 84 2062 421 119 Female, age 51 1543 444 223 Male, age 73 2029 576 121 Female, age 83 1861 2 228 Female, age 80 1688 365 122 Male, age 51 2476 0 230 Male, age 32 2255 1 123 Female, age 63 1515 3 231 Female, age 72 1568(***) 5 124 Male, age 77 1531(***) 88 232 Female, age 76 397(***) 1383 233 Male, age 57 2230 849 234 Female, age 56 2700 53 2.3.2 Annotations For each record, annotations were recorded by two cardiologists working independently. Th ey noted detector missed beats, rhythm labels, signal quality labels, abnormal beats and comments. The types of annotation are listed in Table 2.3. 2.4 Matlab Simulations Results 2.4.1 Simulation Setup First, a program reads in the selected ECG data set s and their attributes from the MIT BIH Arrhythmia Database to Matlab workspace. Then, the data is re sampled from 320 Hz to 250Hz,

PAGE 30

30 and Matlab functions implemented for the Thresholding and Pan Tompkins algorithms are executed. Finally, the QRS detection r esults are compared against the ground truth (annotations) to evaluate the performance. Table 2 3. Heart beat annotations appearing in the MIT BIH Arrhythmia Database Index Beat Types Index Beat Types 1 Normal Beat 22 Comment annotation 2 Left bundle br anch block beat 23 Measurement annotation 3 Right bundle branch block beat 24 P wave peak 4 Aberrated atrial premature beat 25 Left or right bundle branch block 5 Premature ventricular contraction 26 Non conducted pacer spike 6 Fusion of ventricular an d normal beat 27 T wave peak 7 Nodal (junctional) premature beat 28 Rhythm change 8 Atrial premature contraction 29 U wave peak 9 Premature or ectopic supraventricular beat 30 Learning 10 Ventricular escape beat 31 Ventricular flutter wave 11 Nodal (j unctional) escape beat 32 Start of ventricular flutter/fibrillation 12 Paced beat 33 End of ventricular flutter/fibrillation 13 Unclassifiable beat 34 Atrial escape beat 14 Signal quality change 35 Supraventricular escape beat 15 Not specified 36 Not s pecified 16 Isolated QRS like artifact 37 Non conducted P wave (blocked APB) 17 Not specified 38 Fusion of paced and normal beat 18 ST change 39 Waveform onset 19 T wave change 40 Waveform end 20 Systole 41 R on T premature ventricular contraction 21 Diastole 2.4.2 Detection Accuracy Essentially, we are interested in how well the algorithm performs in discriminating normal beats from abnormal beats and recognizing different types of abnormal beats. There are several

PAGE 31

31 terms defined for serving this purpose. TP (true positives) is the number of correct detections. FP (false positives), sometimes called false alarms is the number of incorrect detection. FN (false negatives), sometimes called misses is incorrect miss detection. For each beat type, the performance is measured using the following parameters: Sensitivity is defined as the amount of correctly detected beats for each annotated beat. Positive Predictivity is the probability that a positive detection is correct. Note that the total number of an notated beats is the sum of TP and FN and the total number of detected beats is the sum of TP and FP 2.4.3 Discussions Tables 2 4 and 2 5 list the simulation results for Pan Tompkins and Thresholding algorithms using record #100 to #124 in MIT BIH Arrhyt hmia Database For some records, the detection accuracy is too low due to a misalignment of the positions sequence calculated by the algorithm in comparison to the annotation files. To correct these errors, human visual inspection and manual annotation pos ition change is required, which is a time consuming task. Therefore, we eliminated these records in the subsequent performance comparisons. Figure 2 5(a) shows the sensitivity comparison between these two algorithms. Among 14 records, Pan Tompkins algorith m outperforms the Thresholding algorithm in 9 records. This is because there are more missed R peaks using Thresholding causing large FN number. In Pan Tompkins algorithm, the search back logic using lower second threshold reduces the number of missed R p eaks.

PAGE 32

32 Table 2 4. QRS detection results using the Pan Tompkins algorithm (Records #100 #124). Record Annotated Beats Detected Beats TP FP FN Sensitivity TP/(TP+FN) Predictivity TP/(TP+FP) 100 2273 2252 2028 224 245 89.22% 90.05% 101 1874 1834 1747 87 127 93.22% 95.26% 102 2191 2248 1767 481 424 80.60% 78.60% 103 2090 2154 2052 102 38 98.18% 95.26% 104 2310 1736 1571 165 739 68.01% 90.50% 105 2691 2553 868 1685 1823 32.23% 34.00% 106 2098 1982 1306 676 792 62.25% 65.89% 107 2140 2315 2023 292 117 94.53% 87.39% 108 1824 2374 1459 915 365 79.99% 61.46% 109 2534 2652 1496 1156 1038 59.04% 56.41% 111 2133 2025 1978 47 155 92.73% 97.68% 112 2549 2704 1458 1246 1091 57.20% 53.92% 113 1795 2059 1752 307 43 97.60% 85.09% 114 1890 1222 1047 175 843 5 5.40% 85.68% 115 1961 2178 1922 256 39 98.01% 88.25% 116 2420 2650 1565 1085 855 64.67% 59.06% 117 1538 1526 1235 291 303 80.30% 80.93% 118 2301 2139 1734 405 567 75.36% 81.07% 119 2094 1910 1140 770 954 54.44% 59.69% 121 1875 2012 1826 186 49 97.39% 90.76% 122 2478 2439 1114 1325 1364 44.96% 45.67% 123 1519 1924 774 1150 745 50.95% 40.23% 124 1633 1624 1405 219 228 86.04% 86.51% Table 2 5. QRS detection results using Thresholding m ethod (Records #100 #124). Record Annotated Beats Detected B eats TP FP FN Sensitivity TP/(TP+FN) Predictivity TP/(TP+FP) 100 2273 2270 1888 382 385 83.06% 83.17% 101 1874 1863 1853 10 21 98.88% 99.46% 102 2191 1820 1787 33 404 81.56% 98.19% 103 2090 2075 2025 50 65 96.89% 97.59% 104 2310 1051 967 84 1343 41 .86% 92.01% 105 2691 2517 215 2302 2476 7.99% 8.54% 106 2098 1735 1258 477 840 59.96% 72.51% 107 2140 1966 1951 15 189 91.17% 99.24% 108 1824 1255 1208 47 616 66.23% 96.26% 109 2534 2372 134 2238 2400 5.29% 5.65% 111 2133 1854 1841 13 292 86.31% 99.3 0% 112 2549 2536 25 2511 2524 0.98% 0.99% 113 1795 1792 1761 31 34 98.11% 98.27% 114 1890 608 569 39 1321 30.11% 93.59% 115 1961 1948 1921 27 40 97.96% 98.61% 116 2420 1927 730 1197 1690 30.17% 37.88% 117 1538 1532 1527 5 11 99.28% 99.67% 118 2301 2 220 1592 628 709 69.19% 71.71% 119 2094 1984 1158 826 936 55.30% 58.37% 121 1875 1858 1808 50 67 96.43% 97.31% 122 2478 2472 487 1985 1991 19.65% 19.70% 123 1519 1516 1000 516 519 65.83% 65.96% 124 1633 1571 1524 47 109 93.33% 97.01%

PAGE 33

33 Figure 2 5 (b) shows the predictivity comparison. Among 14 records, Thresholding algorithm outperforms Pan Tompkins algorithm in 12 records. This is because Pan Tompkins produces more FPs due to T waves and noise that are mistaken as QRS complexes. However, as stated before, the Thresholding algorithm has lower hardware complexity. Further studies on the tradeoff between performance and hardware implementation cost will be forthcoming. a b Figure 2 5. Performance comparison using MIT BIH Arrhythmia Database (a) se nsitivity; (b) positive predictivity.

PAGE 34

34 CHAPTER 3 BEAT CLASSIFICATION Heart beat classification is an essential part of ECG signal processing. When a QRS complex is detected, its morphology change can be analyzed and associated with cardiac problems. The c lassification techniques fall into three main categories time domain, frequency domain, and time frequency domain. For the time domain, the shape correlation is a frequently used approach. For frequency domain, t he Fast Fourier Transform (FFT) is the mos t popular method to obtain the frequency components of the ECG record [ 68 ] For the time frequency domain, a widely used beat classification technique is wavelets decomposition [11]. As stated in Chapter 1, a low cost, automatic real time approach to clas sify heart beats is desired. Here we propose and evaluate a time domain method that correlates each detected beat with a previously st ored template beat. The difference from the template beat is calculated and then used to cluster similar beats together. N ormally, heart beats from a healthy ECG have similar shapes and thus only one cluster is formed. However, when abnormal beats or noises are present, more different clusters will be formed. The formation of additional clusters indicates potential heart prob lems, and thus a need to generate appropriate warnings to notify patients and remote healthcare centers. 3.1 Types of Arrhythmia An arrhythmia is an irregular heartbeat the heart may beat to o fast (tachycardia), too slow (bradycardia), too early (prematu re contraction) or too irregularly (fibri llation). Arrhythmias are heart rhythm problems they occur when the electrical impulses to the heart that coordinate heartbeats are not working properly, making the heart beat too fast, too slow or inconsistently The occurrence of abnormal beats is useful in diagnosing arrhythmia. Annotated beats that frequently appear in the MIT BIH Arrhythmia Database are Normal Beat ( #1), PVC (Premature

PAGE 35

35 Ventricular Contraction, #5), APB (Atrial Premature Beat, #8), RBBB (Righ t Bundle Branch Block Beat, #3), LBBB (Left Bundle Branch Block Beat, #2), Paced Beat and Fusion (Fusion of Ventricular and Normal Beat, #6). Table 3 1 categorizes the characteristics of these annotated beats and their medical indication. Snapshots of eac h annotated beat are shown in Figures 3 1 and 3 2. The waveforms are acquired, overlapped, and plotted in Matlab using the MIT BIH Arrhythmia Database In each plot, the unit on the x axis represents the number of samples of the detected waveforms (sampled at 250Hz) The y axis is set to display a scaled version of the original voltage amplitude sensed by the recording devices. The amplitude swing in each waveform is from 2 to 2 Table 3 1. Overview of annotated beat types. Beat Types # of Records Characte ristics Medical Indication Normal Beat 40 Normal QRS Healthy human heart Premature Ventricular Contraction (PVC) 37 Widened QRS, Bizarre shape Three or more PVCs occur in a row indicate ventricular Tachycardia Atrial Premature Beat (APB) 26 Early P wave When occurring frequently, treatment is necessary. Right Bundle Branch Block (RBBB) 6 Widened S wave Warning sign of other, more serious heart conditions. Left Bundle Branch Block (LBBB) 4 Broad R wave Longer QRS Same with RBBB Paced Beat 4 Pacing spike followed by a wide QRS Pacemaker in use Fusion (Paced & Normal) 2 Flat P Q and S waves Pacemaker in use

PAGE 36

36 Normal Beat ( MIT BIH #100) Premature Ventricular Contraction ( MIT BIH #105) Atrial Premature Beat ( MIT BIH #200) Figure 3 1. Normal beat, PVC, and APB.

PAGE 37

37 Left Bundle Branch Block Beat ( MIT BIH #109) Right Bundle Branch Block Beat ( MIT BIH #118) Paced Beat ( MIT BIH #102) Figure 3 2. LBBB, RBBB, and Paced Beat.

PAGE 38

38 Figure 3 3. Snapshots of abnormal beats using the M IT BIH Arrhythmia Database Figure 3 3 shows snapshots of different beat types that are in the same record. For example, in record 105 there are PVC and normal beats presented. Also, from the plots we can observe that PVCs usually have wider shapes. It is clear that the shapes are different for differ ent beat types, suggesting the shape correlation method that will be discussed in the next chapter. 3.2 Clustering Analysis 3 .2.1 Distance Distribution Euclidean distance is a frequently used approach in image of video analysis to measure the similar ity of two vectors In [67], the Euclidean distance method was used for ECG beat detection. Here we proposed to use this method for the heart beat classification. The Euclidean distance d between two shapes and is defined as followed,

PAGE 39

39 We determine the similarity between heart beat shapes be calculat ing their Euclidean distance to a common template beat. To evaluate the beats clustering using the Euclidean distance method, we performed simulations using the MIT BIH Arrhythmia Database and Matlab. The template beats were arbitrarily chosen from those b eats within the majority beat and Paced) were simulated. The Euclidean distances of all the beats with the template beats were computed and recorded, and then the his tograms of the distance distribution were plotted. In Figures 3 4 and 3 5, a total of 48 plots illustrate the beat clustering for each record in the database. In each plot, the scale on the x axis represents the results of Euclidean distance calculation b etween the two beat shapes that have amplitude swings of 4. On the y axis is the number of beats that belong to the predefined distance ranges. From the plots, we can clearly see the distance distribution of different beat types that are in the records. S o me distances are closely clustered within a relatively small range, indicating that the majority of the beats in the clusters have very similar shapes. On the other hand, the presence of wide clusters indicate s that the beats within the se clusters have lar ge shape variations.

PAGE 40

40 A B C D E F Figure 3 4. The distance distribution plots simulated using records #100 to #124 in the MIT BIH Arrhythmia Database A) Record #100. B) Record #101. C) Record #102. D) Record #103. E) Record #104. F) Record #105. G) Record #106. H) Record #107. I) Record #108. J) Record #109. K) Record #111. L) Record #112. M) Record #113. N) Record #114. O) Record #115. P) Record #116. Q) Record #117. R) Record #118. S) Record #119. T) Record #121. U) Record #122. V) Record #123. W) Record #124.

PAGE 41

41 G H I J K L M N Figure 3 4. Continued

PAGE 42

42 O P Q R S T Figure 3 4. Continued

PAGE 43

43 U V W Figure 3 4. Continued A B Figure 3 5. The distance distribution plots simulated using records #200 to #234 in the MIT BIH Arrh ythmia Database A) Record #200. B) Record #201. C) Record #202. D) Record #203. E) Record #205. F) Record #207. G) Record #208. H) Record #209. I) Record #210. J) Record #212. K) Record #213. L) Record #214. M) Record #215. N) Record #217. O) Record #219. P) Record #220. Q) Record #221. R) Record #222. S) Record #223. T) Record #228. U) Record #230. V) Record #231. W) Record #232. X) Record #233. Y) Record #234.

PAGE 44

44 C D E F G H Figure 3 5. Continued.

PAGE 45

45 I J K L M N Figure 3 5. Continued.

PAGE 46

46 O P Q R S T Figure 3 5. Continued.

PAGE 47

47 U V W X Y Figure 3 5. Continued. 3 .2.2 Principal Component Analysis Principal component analysis ( PCA ) transforms a number of correlated variables into a smaller number of uncorrelated variables (p rincipal compon ents) [66] For ECG signals, most of

PAGE 48

4 8 the variability is preserved even after redu cing the dimensionality of heart beat signal datasets Previous studies have shown that heart beat classification is a common application using PCA in which a subset of the pr incipal components serves as features which are used to distinguish between normal sinus beats and abnormal waveforms [ 63 ]. PCA consists of the following steps: store the original data, subtract the mean f rom each of the data dimensions, calculate the cov ariance matrix calculate the eigenvectors and eigenvalues of the covariance matrix, and choose components and forming a feature vector. Applying PCA to a set of pre recorded heart beats, the associated pattern of principal components reflects the degree o f morphologic beat to beat variability When the eigenvalue associated with the first principal component is much larger than those associated with other components, the set exhibits a low morphologic variability, whereas a slow fall off of the principal c omponent values indicates a large variability [ 63 ]. PCA was performed on the 48 data sets in the MIT BIH Arrhythmia Database using Matlab. The resulting 48 plots are shown in Figures 3 6 and 3 7. The beats belong to the same beat type is plotted in the sam e color Thus, for each cluster in the graph we can read off the compactness of each cluster and the distances between clusters. From the PCA plots, we made the first observation: f or data sets with only one ma jor beat type, the clusters are clearly compac t. For example, there are 21 records that include only the normal beat type: 100, 101, 103, 105, 108, 112, 113, 114, 115, 117, 121, 122, 123, 202, 205, 209, 219, 220, 222, 230, and 234. All the records show compact clustering except the 230, in which there are two different shapes presented. The second observation is that the redundancies are revealed in some of the records that have only normal beats. T he redundanc y indicates the degree of correlation between PCs and is

PAGE 49

49 reflected in the shapes of the clu sters. For example, there are several record s show long stripe shaped cluster s including 100, 101, 103, 115, 121 and 200. Th e s e type s of shape s are the results of high degree of correlation between two of the ir PCs. Finally for data sets with two or more major beat types, the results vary. For example, PVCs in the database have larger beat shape variation, resulting in PVC clusters which are diffuse and inseparable from the normal beat clusters in several cases. On the other hand, clusters of RBBBs and LB BBs show clear distinction from the normal beat clusters, as show n in record 207, 212, and 231. A B C D Figure 3 6. The PCA plots for records from #100 to #124 in the MIT BIH Arrhythmia Database A) Record #100. B) Record #101. C) Record #102. D) Record #103. E) Record #104. F) Record #105. G) Record #106. H) Record #107. I) Record #108. J) Record #109. K) Record #111. L) Record #112. M) Record #113. N) Record #114. O) Record #115. P) Record #116. Q) Record #117. R) Record #118. S) Record #119. T) Record #121. U) Record #122. V) Record #123. W) Record #124.

PAGE 50

50 E F G H I J K L M N Figure 3 6. Continued.

PAGE 51

51 O P Q R S T U V W Figure 3 6. Continued

PAGE 52

52 A B C D E F G H Figure 3 7. The PCA plots for records from #200 to #234 in the MIT BIH Arrhythmia D atabase A) Record #200. B) Record #201. C) Record #202. D) Record #203. E) Record #205. F) Record #207. G) Record #208. H) Record #209. I) Record #210. J) Record #212. K) Record #213. L) Record #214. M) Record #215. N) Record #217. O) Record #219. P) Reco rd #220. Q) Record #221. R) Record #222. S) Record #223. T) Record #228. U) Record #230. V) Record #231. W) Record #232. X) Record #233. Y) Record #234.

PAGE 53

53 I J K L M N O P Figure 3 7. Continued.

PAGE 54

54 Q R S T U V W X Y Figure 3 7. Continued.

PAGE 55

55 3.3 Mor phological Variation ECG signals have non stationary and easily disturbed features. Even for a single person, the amplitude and shape of detected heart beats can change over time due to changes in the propagation medium between heart and sensor, measuremen t variability in ECG recordings affecting the amplitude levels, implanted devices drift, and mental status changes. Consider the stress test plot in Figure 3 8, computed for sequential thirty second intervals during a 28 minute interval, as an example. The plot shows representative beats (i.e., average beats) during the stress test [ 65 ]. Clearly changes in mental status can cause the heart beat shapes to change drastically. The solution to the morphological variation is dynamically adapting to the shape ch anges. Whenever shape changes occur, new clusters needs to be created immediately and given sufficient time and space to grow. The clusters with old beat shapes need to be gradually flushed out of the system to free up the space for future clusters. Fig ure 3 8. Representative beats during a stress test

PAGE 56

56 CHAPTER 4 ARCHITECTURE 4.1 Computational Model In response to the growing demand for the need to transmit ECG data wirelessly, heart beat classification can be performed onsite and on line in a hardware sol ution, instead of in software as traditionally performed. Previous studies show that online digital beat classification is feasible on the FPGA platform [ 2 ] To design efficient hardware architecture for heart beat classification, we need to first develop an underlying computational model based on a fundamental understanding of the nature of ECG signals. ECG, along with neural spikes and respiration signals, are recurrent biomedical signals. Figure 4 1 illustrates a top down flowchart for the classification process of these recurrent biomedical signals. First, the detection stage captures the occurrences of the recurrent signal. Second, the feature extraction stage converts the original signal into a set of features. The features, extracted by arithmetic log ic units, can then be weighted by different coefficients depending on the classification algorithms. These weighted features can be considered as identifications for the signals. Third, the system searches the prior knowledge, recorded in the form of memor y, to determine if the identifications satisfy the criteria for different clusters. If there are any recorded criteria met in the memory, they will be updated to reflect the changes caused by newly detected signals. If there are no satisfied criteria, a se t of newly created features will be merged into the existing knowledge. Finally, the system resets to the initial state and is ready for the next signal. Based on this model, the use of memory elements in the system is inevitable, whether for storing coeff icients or the classification criteria set. The size of the memory needed is determined by the complexity of the classification algorithms.

PAGE 57

57 Figure 4 1. A generalized computational model for the biomedical signals classification 4.2 Clustering Architec ture 4.2.1 Cache In Sections 3.2.1 and 3.2.2, we observed the clustering properties of heart beats through analytical tools such as Euclidean distance distribution and PCA. In Section 3.2.3, we demonstrated the temporal properties of the heart beat shape c hanges. Combining the generalized computational model described in Section 4.1 with the requirements of ultra low power consumption and small memory space for implantable medical devices, we propose a hardware solution that benefits from the cache concept in Computer Architecture. Cache holds a crucial role in Computer Architecture and has a history spanning several decades. Caches are used as a temporary storage area to reduce the effective latency between the

PAGE 58

58 Figure 4 2. Analogy between normal cache a nd Heart Beat Cache processor and main memory by providing rapid access to recently used or frequently needed data. Furthermore, the small on chip cache reduces the overall system hardware cost. Before directly applying the cache concept to the heart beat classification problem, we review some important underlying principles. The most fundamental is the principle of locality which considers both temporal locality and spatial locality. Temporal locality (locality in time) states that a data block that is r eferenced will tend to be referenced again soon. Spatial locality (locality in space) states that if a data block is referenced, data blocks whose addresses are close will tend to be referenced soon [62]. There is thus a need to find the locality properti es within the heart beat clustering process. In beat classification applicat ions, only those detected real beats are stored for faste r classification of future beats. Multiple detected beats of the same types are used to form a clust er. Instead of storing the entire set of heart beat waveforms in one cluster, only the extracted informati on such as distance range, beat occurrenc e frequency, and number of beat s ar e stored in a cache entry. Temporal locality is exploited so that frequently occurring beat types are maintained in the cache with high priority. Spatial l locality is different from the traditional

PAGE 59

59 definition in the cache concept where t he data accessing behavior causes the nearby cache blocks to be fetched from the main memory to the cache. Here, t hose beat s with Euclidean distance similar to the template beat are clustered together in the cache Nonetheless, t he similarity lies in the fact that the spatially close data are grouped together in the cache We proposed the use of cache in the applicat ion of beat classification, which is extremely memory optimized compared to previous proposed architecture [30], in which a larger memory size is used for adopting the neural network algorithm s 4.2.2 Reuse Distance of Detected Heart Beats To gain further insight in the applicability of recording recent heart beats for heart beats classification, we study the reuse behavior using real ECG databases. Caches take advantage of data reuse. Reuse distance, measured as the number of unique addresses accessed sinc e the last reference to the requested data, is a useful metric of cache behavior [64] Figure 4 3 is an illustration of the reuse distance. Reuse distance is directly related to the origin of cache misses. The cache hit/miss ratio the number of cache hits divided by number of cache misses is a main performance characteristi c of a cache Cache misses are categorized by their three causes compulsory misses, conflict misses, and capacity misses. A compulsory miss is caused by the first access to a block th at has never been in the cache. Capacity misses are those due to insufficient cache capacity (i.e., the reuse distance is too large). Conflict misses are those misses in a direct mapped or set associative cache that are eliminated in a fully associative ca che of the same size [62]. A fully associative cache with the Least Recently Used (LRU) policy indicates that a reference is a cache hit if its reuse distance is smaller than the number of cache lines. If the reuse distance is larger than the number of ca che lines, it results in a miss. Heart Beat Cache (HBC)

PAGE 60

60 Figure 4 3. An illustration of the definition of reuse distance in cache behavior. can be considered as a special kind of fully associative cache such that a data block (heart beats clusters) can b e stored in any cache line, while the mapping is determined by Euclidean distance ranges instead of data addresses. In HBC, the hit/miss ratio reflects the ability to identify and retain the real heart beats. Moreover, this ration reflects the accuracy of the classification process. Similarly, the reuse distance can be used to characterize the behavior of HBC. For each record in the MIT BIH Arrhythmia Database we performed the following procedure to evaluate the reuse distance distribution. First, the enti re distance span is divided equally into 50 segments. The number 50 was determined based on the close examination of the distance distribution for all 48 records. Most of the clusters can be covered completely in the range of 1, which is exactly the length of one segment. Second, each detected heart beat is assigned an ID which is the segment index where its distance (with the template) falls into. Third, the reuse distance of each detected heart beat is calculated by counting the number of beats between t he current beat and the previous beat with the same ID. Finally, all of the reuse distances are sorted into bins for analysis. Figure 4 4 shows the reuse distance distribution for the 48 records in the MIT BIH Arrhythmia Database From Figure 4 4, we made the following observations. (1) The records

PAGE 61

61 that consist of normal beats without the presence of other abnormal beats have the densest concentrations in short reuse distances. This is because of the strong recurrence and similar shape properties of normal beats. (2) The abnormal beats with larger shape variations (such as PVCs) occupy more Euclidean distance segments and therefore cause the distribution curve to expand into the region of long reuse distance. (3) For almost all records, the majority of the b eats are found within the reuse distance of 1. This is a strong indication of the spatial and temporal locality in the proposed classification architecture. A B C D Figure 4 4. The reuse distance plots from record #100 to #124 in the MIT BIH Arrhythmia Database A) Record #100. B) Record #101. C) Record #102. D) Record #103. E) Record #104. F) Record #105. G) Record #106. H) Record #107. I) Record #108. J) Record #109. K) Record #111. L) Record #112. M) Record #113. N) Record #114. O) Record #115. P) Rec ord #116. Q) Record #117. R) Record #118. S) Record #119. T) Record #121. U) Record #122. V) Record #123. W) Record #124.

PAGE 62

62 E F G H I J K L Figure 4 4. Continued.

PAGE 63

63 M N O P Q R S T Figure 4 4. Continued.

PAGE 64

64 U V W Figure 4 4. Con tinued.

PAGE 65

65 A B C D E F Figure 4 5. The reuse distance plots simulated using records #200 to #234 in the MIT BIH Arrhythmia Database A) Record #200. B) Record #201. C) Record #202. D) Record #203. E) Record #205. F) Record #207. G) Record #208. H) R ecord #209. I) Record #210. J) Record #212. K) Record #213. L) Record #214. M) Record #215. N) Record #217. O) Record #219. P) Record #220. Q) Record #221. R) Record #222. S) Record #223. T) Record #228. U) Record #230. V) Record #231. W) Record #232. X) R ecord #233. Y) Record #234.

PAGE 66

66 G H I J K L M N Figure 4 5 Continued.

PAGE 67

67 O P Q R S T U V Figure 4 5 Continued.

PAGE 68

68 W X Y Figure 4 5 Continued. After comparing the 48 reuse distance plots against the simulated used cache en tries in Table 4 7 we found that most of the reuse distance plot s project the simulated number of used cache entries quite accurately, except for record #208, record #221, and record #228 To understand the underlying causes of the behaviors of these t hre e records, we further investigat ed their reuse d istance distribution by changing the bin size s Rather than using the original bin size that is calculated by dividing the entire distance span equally into 50 bins, we in turn divided the distance span into 40, 30, 20, and 10 segments The simulations were performed for all of these configurations. Figure 4 6 shows the reuse distance plots with all the different configurations and the distance distribution plot for the record #208 Clearly the plots show that the wider the bin size (fewer segments for the entire distance span) the narrower the reuse distances distribution It is

PAGE 69

69 because the PVC beat types in record #208 have larger shape variations compared to t he distance range we normally define d for other beat types. Therefore more than one bin w as need ed to cover the entire distance range resulting in the longer reuse distance for PVC beats F igure 4 7 conveys similar information as Figure 4 6. In Figure 4 8, t he large reuse distance distribution is cau sed by the PVCs that constitute with two major ranges. A B C D Figure 4 6 The reuse distance plots simulated using record # 208 in the MIT BIH Arrhythmia Database A) 50 bins. B) 40 bins. C) 30 bins. D) 20 bins. E) 10 bins. F) Distance distribution

PAGE 70

70 E F Figure 4 6. Continued

PAGE 71

71 A B C D E F Figure 4 7 The reuse distance plots simulated using record # 221 in the MIT BIH Arrhythmia Database A) 50 bins. B) 40 bins. C) 30 bins. D) 20 bins. E) 10 bins. F) Distance distribution

PAGE 72

72 A B C D E F Figure 4 8 The reuse distance plots simulated using record # 228 in the MIT BIH Arrhythmia Database A) 50 bins. B) 40 bins. C) 30 bins. D) 20 bins. E) 10 bins. F) Distance distribution

PAGE 73

73 Figure 4 9 shows a summed reuse distance distribution of al l the records in the MIT BIH Arrhythmia Database The x axis represents the reuse distance and the y axis represents the number of beats expressed in log scale. Clearly most similarly shape heart beats occur at short reuse distance, and the number of heart beats decreases strongly with the increase of reuse distance: there is an approximately 100 times reduction at a distance of 8, and 1000 times reduction at a distance of 32. Alternatively, Figure 4 9 can be viewed as the probability of hitting the cache at a certain reuse distance. As an example, we can specify the number of beats that can be stored in cache and use the distribution curve to find the corresponding reuse distance. This chosen reuse distance, considered to be equivalent to the number of cac he lines under ideal SNR, is then padded with a number of extra lines for storing the noise clusters. Figure 4 9 The distribution of the summed reuse distance for all the detected heart beats in the MIT BIH Arrhythmia Database

PAGE 74

74 4.3 Heart Beat Cache 4 3 .1 Cache Organization The heart beat classification architecture presented in this study is the Heart Beat Cache, a term inspired by the Computer Architecture field. The online beat classification system is illustrated in Figure 4 10 and consists of a cus tom designed static random access memory (SRAM), a memory controller, and a Euclidean distance calculator. The key id ea is to maintain a small table which provides the necessary information to determine the cluster attributes. On the right bottom side of F igure 4 10 is the Euclidean distance distribution of these three beat types that appear in record #207. Figure 4 11 shows the overlapped beat waveforms of these three beat types. The cache consists of six columns as described below, Valid: if the bit is se save power. Mean [7:0]: the mean distance value of all the beat s belonging to this cluster. Max [7:0]: the upper bound value of the cluster. Min [7:0]: the lower bound value of the cluster. Age [5:0]: the total number of beat s detected after the last b e at assigned to this cluster. In other words, it represents how often this cluster is updated. Count [5:0]: the total number of beat s that belong to this cl uster, with the maximum valu e being 63. 4 3 .2 Cache Operations Cache operations begin b y storing a newly detected beat in a first i n, first out (FIFO) in the QRS complex de tector. The first detected beat is stored as a temp orary template inside the Heart Beat Cache. All the subse quen t beats are compared using their Euclidean distance against the template. Those beat s whose Euclidean distance falls within a specified range belong to the same

PAGE 75

75 Figure 4 10 System block diagram and detailed Heart Beat Cache organization. Figure 4 11 Heart beats samples taken from record #207 in MIT BIH Arrhythmia Database

PAGE 76

76 cluster. The extracted cluster attributes are s tored in one entry of the Heart Beat Cache and the counter value for that entry i s updated whenever a new beat gets added to the c luster. When the counter reaches the chos en maximum value, a valid beat type cluster is identif ied in the cache. The next beat that belongs to this cluster is then chosen as the new template and replaces the current template. Afterwa rds, for each new incom ing beat, the Heart Beat Cache only needs to perform the following three routine operations: 1) Insertio n: For each newly detected beat the Euclidean distance from the template beat is computed. The memory controller then searches the cache for a v alid cluster that this new beat belongs to. If there is no valid cluster found, the first available free entry will be turned on and filled in. 2) Merge: A newly detected beat is compared against all the clusters. If a beat can be classified as belonging to mo re than one of the clusters, these clusters are merged into one. The merging process involves a recalculation of the cluster attributes and a gradual expansion of the upper and lower cluster boundaries. 3) hes the specified maximum value, the ers can prevent overlapped beats and noise beats from forming a valid beat type cluster. It also frees the cache space for future beat clustering. 4 3 .3 Power Management The leakage power becomes increasingly significant in deep sub micron technologies. For ECG processing applications, the operating frequency is typically low, and thus the dynamic power is not as important as the leakage power. The best wa y to control leak age is to turn off power to the memory with power gating. Simulation r esults reveal that 14 entries are sufficient to store the cluster informati on at all times. 4.4 Matlab Simulation Results 4.4.1 Simulation Setup First, a program reads in the selected EC G data sets and their attributes from MIT BIH Arrhythmia Database to the Matlab workspace. The data is then re sampled from 320 Hz to

PAGE 77

77 250Hz and band pass filtered. A Matlab function implemented for the Heart Beat Cache algorithm is then executed. Finally, the performance evaluation for the classification results is calculated by comparing against the ground truth (annotations). The illustration of the simulation Figure 4 12 Simulation setup illustration using the record #106 from MIT BIH Arrhythmia Dat abase setup is shown in Figure 4 12. In this example, the MIT BIH #106 data set is used, and two dominant beat types (Normal and PVC) are successfully clustered into different cache entries. 4.4.2 Evaluation of Classification Performance We are essentiall y interested in how well the algorithm performs in terms of discriminating normal beats from abnormal beats, and also in recognizing different abnormal beats. Here we define several terms to serve this purpose. TP (true positives) refers to the correct cla ssification, i.e., the beats with one annotation being clustered into one cache entry. FP (false positives), sometimes called false alarms is an incorrect classification which means the beats with other annotated types fall into a cache entry occupied by beats that belong to a specific annotated type.

PAGE 78

78 FN (false negatives), sometimes called misses is an incorrect misclassification which means the annotated beats fall out of their designated cache entry. For each beat type, the performance is measured by th e following parameters: Sensitivity is defined as the amount of correctly clustered beats for each annotated type. Positive Predictivity is defined as the amount of correctly clustered beats for all the detected beats for each annotated type. Note that the total number of annotated beats is the sum of TP and FN and the total number of detected beats is the sum of TP and FP 4.4.3 Normal Beats Normal beats account for the majority of the detected beat s in the database. Since we choose the template from the majority of the beats, the Euclidean distance between the detected normal beats and the template is small. The majority of the beats occupy cache entries with high beat counts and high occurrence fre cache. The simulated clustering result for normal beats for the 48 records in database is summarized in Table 4 1 and Figure 4 13 4.4.4 Premature Ventricular Contraction Table 4 2 and Figure 4 14 sh ow the clustering results for PVC. PVCs yield more inconsistent results due to their larger shape variation and the possibility of PVCs in some records containing more than one shape type. 4.4.5 Other Abnormal Beats Tables 4 3, 4 4, 4 5, and 4 6 show the c lustering results for other abnormal beat types. Other abnormal beat types have better performance than PVCs.

PAGE 79

79 4.4.6 Cache Size Usage Table 4 7 records the number of cache entries used during the simulation. Figure 4 10 shows the number of records versus th e number of used entries. The number of cache entries is strongly dependent on the number of detected beat types, the SNR of the signals, and the frequency of the y of the pre existent entries, more cache entries may need to be inserted to accommodate these newly formed clusters, which are spread out to begin with. Later these spread out clusters can get merged, the more unqualified clusters can get deleted and fina lly the number of valid entries decreases and comes back to a stable value. The simulation shows that, the maximum cache size is 14 entries for the worst case, while on an average, 6 entries are used. 4.5 Discussions Based on the results in Table 4 2, the classification of normal beats shows consistent performance throughout 48 data sets, with good average sensitivity (95.83%) and predictivity (89.92%). This consistency is due to less shape variation among normal beats and results in better clustering. PVCs yield more inconsistent results due to their larger shape variation and the possibility of PVCs in some records containing more than one shape type. Several records had poor performance caused by a larger FP or FN FP is large when the specified range for an initial cluster is too large, causing beats that do not belong to get assigned to that cluster. On the other hand, FN is large when the initial range for a cluster is too small, causing beats to miss the range. Further studies to improve the classifica tion performance for PVCs are needed. Table 4 8 and Figure 4 11 summarize the performance comparison between [2] and Heart Beat Cache Heart Beat Cache on Paced, LBBB and Fusion, but not on RBBB. However, the Heart Beat Cache requires much less hardware complexity.

PAGE 80

80 Table 4 1. Clustering results for Normal beats. Record Record Patient # QRS TP FP FN Sensitivity Predictivity 100 Male, age 69 2238 2232 33 6 99.73% 98.54% 101 Female, age 75 1860 1853 7 7 99.62% 99.62% 102 Femal e, age 84 99 99 36 0 100.00% 73.33% 103 Male, age N/A 2081 2077 8 4 99.81% 99.62% 104 Female, age 66 163 114 584 49 69.94% 16.33% 105 Female, age 73 2526 2382 128 144 94.30% 94.90% 106 Female, age 24 1507 1475 93 32 97.88% 94.07% 108 Female, age 87 17 39 1733 84 6 99.65% 95.38% 112 Male, age 54 2536 2532 12 4 99.84% 99.52% 113 Female, age 24 1788 1775 0 13 99.27% 100.00% 114 Female, age 72 1820 1815 43 5 99.73% 97.69% 115 Female, age 39 1952 1947 8 5 99.74% 99.59% 116 Male, age 68 2301 1687 15 614 73.32% 99.12% 117 Male, age 69 1533 1476 2 57 96.28% 99.86% 119 Female, age 51 1543 1424 54 119 92.29% 96.35% 121 Female, age 83 1860 1852 13 8 99.57% 99.30% 122 Male, age 51 2475 2470 2 5 99.80% 99.92% 123 Female, age 63 1515 1481 3 34 97.76% 99.80% 200 Male, age 64 1742 1693 604 49 97.19% 73.70% 201 Male, age 68 1624 1468 58 156 90.39% 96.20% 202 Male, age 68 2060 2034 70 26 98.74% 96.67% 203 Male, age 43 2529 2280 411 249 90.15% 84.73% 205 Male, age 59 2570 2551 30 19 99.26% 98.84% 208 Female, age 23 1585 1304 377 281 82.27% 77.57% 209 Male, age 62 2620 2615 428 5 99.81% 85.93% 210 Male, age 89 2422 2401 236 21 99.13% 91.05% 212 Female, age 32 922 909 23 13 98.59% 97.53% 213 Male, age 61 2640 2604 411 36 98.64% 86.37% 215 Male, age 81 3194 3165 42 29 99.09% 98.69% 217 Male, age 65 244 197 134 47 80.74% 59.52% 219 Male, age N/A 2082 2052 244 30 98.56% 89.37% 220 Female, age 87 1953 1941 114 12 99.39% 94.45% 221 Male, age 83 2030 1873 41 157 92.27% 97.86% 222 Female, age 84 2061 2053 571 8 99.61% 78.24% 223 Male, age 73 2028 1995 468 33 98.37% 81.00% 228 Female, age 80 1688 1636 176 52 96.92% 90.29% 230 Male, age 32 2254 2247 210 7 99.69% 91.45% 231 Female, age 72 314 311 7 3 99.04% 97.80% 233 Male, age 57 2229 2223 683 6 99.73% 76.5 0% 234 Female, age 56 2700 2697 62 3 99.89% 97.75%

PAGE 81

81 A B Figure 4 13 Clustering performance results for n ormal beats A) R ecord s # 100 to #123. B) R ecord s #200 to # 234 0.00% 10.00% 20.00% 30.00% 40.00% 50.00% 60.00% 70.00% 80.00% 90.00% 100.00% #100 #101 #102 #103 #104 #105 #106 #108 #112 #113 #114 #115 #116 #117 #119 #121 #122 #123 Sensitivity Predictivity 0.00% 10.00% 20.00% 30.00% 40.00% 50.00% 60.00% 70.00% 80.00% 90.00% 100.00% #200 #201 #202 #203 #205 #208 #209 #210 #212 #213 #215 #217 #219 #220 #221 #222 #223 #228 #230 #231 #233 #234 Sensitivity Predictivity

PAGE 82

82 Table 4 2. Clustering results for Premature Ventricular Contraction (PVC) beats. R ecord Record Patient # QRS TP FP FN Sensitivity Predictivity 100 Male, age 69 1 1 0 0 100.00% 100.00% 102 Female, age 84 4 0 0 4 --104 Female, age 66 2 1 0 1 50.00% 100.00% 105 Female, age 73 41 24 103 17 58.54% 18.90% 106 Female, age 24 520 446 4 5 74 85.77% 90.84% 107 Male, age 63 59 8 121 51 13.56% 6.20% 108 Female, age 87 17 0 0 17 --109 Male, age 64 38 12 1 26 31.58% 92.31% 111 Female, age 47 1 1 0 1 100.00% 100.00% 114 Female, age 72 43 21 4 22 48.84% 84.00% 116 Male, age 68 109 103 25 6 94.50% 80.47% 118 Male, age 69 16 9 19 7 56.25% 32.14% 119 Female, age 51 444 436 163 8 98.20% 72.79% 121 Female, age 83 1 0 0 1 --123 Female, age 63 3 0 0 3 --124 Male, age 77 47 40 15 7 11.26% 72.73% 200 Male, age 64 826 365 98 461 44. 19% 78.83% 201 Male, age 68 198 196 313 2 98.99% 38.51% 202 Male, age 68 19 19 17 0 100.00% 52.78% 203 Male, age 43 444 50 32 394 11.26% 60.98% 205 Male, age 59 71 48 21 23 67.61% 69.57% 207 Female, age 89 105 50 39 55 47.62% 56.18% 208 Female, age 2 3 992 720 304 272 72.58% 70.31% 209 Male, age 62 1 1 1 0 100.00% 100.00% 210 Male, age 89 194 18 22 176 9.28% 45.00% 213 Male, age 61 220 82 22 138 37.27% 78.85% 214 Male, age 53 256 204 36 52 79.69% 85.00% 215 Male, age 81 164 156 31 8 95.12% 83.42% 217 Male, age 65 162 72 812 90 44.44% 8.14% 219 Male, age N/A 64 22 194 34 34.38% 10.19% 221 Male, age 83 396 365 163 31 66.92% 69.13% 223 Male, age 73 473 127 44 346 26.85% 74.27% 228 Female, age 80 362 178 50 184 49.17% 78.07% 230 Male, age 32 1 0 0 1 --231 Female, age 72 2 1 1 1 50.00% 50.00% 233 Male, age 57 831 177 8 654 21.30% 95.68% 234 Female, age 56 3 0 0 3 --

PAGE 83

83 A B Figure 4 1 4 Clustering performance results for PVCs A) Records #100 to #1 19 B) Records #200 to #23 3 0.00% 10.00% 20.00% 30.00% 40.00% 50.00% 60.00% 70.00% 80.00% 90.00% 100.00% #100 #104 #105 #106 #107 #109 #111 #114 #116 #118 #119 Sensitivity Predictivity 0.00% 10.00% 20.00% 30.00% 40.00% 50.00% 60.00% 70.00% 80.00% 90.00% 100.00% #200 #201 #202 #203 #205 #207 #208 #209 #210 #213 #214 #215 #217 #219 #221 #223 #228 #231 #233 Sensitivity Predictivity

PAGE 84

84 Table 4 3 Clustering results for Paced b eat s Record Record Patient # QRS TP FP FN Sensitivity Predictivity 102 Female, age 84 2027 2024 163 3 99.85% 92.55% 104 Female, age 66 1379 1360 145 19 98.62% 90.37% 107 Male, age 63 2078 1864 18 214 89.70% 99.04% 217 M ale, age 65 1542 1021 160 521 78.60% 86.45% Table 4 4. Clustering results for Left Bundle Branch Block (LBBB) beats. Record Record Patient # QRS TP FP FN Sensitivity Predictivity 109 Male, age 64 2492 2143 13 49 86.00% 99.40% 111 Female, age 47 2123 21 19 8 4 99.81% 99.62% 207 Female, age 89 1457 1444 315 9 99.11% 82.09% 214 Male, age 53 2002 1970 43 32 98.40% 97.86% Table 4 5. Clustering results for Right Bundle Branch Block (RBBB) beats. Record Record Patient # QRS TP FP FN Sensitivity Predictivity 118 Male, age 69 2166 2156 100 10 99.54% 95.57% 124 Male, age 77 1530 1515 43 15 99.02% 97.24% 207 Female, age 89 86 23 24 63 26.74% 48.94% 212 Female, age 32 1825 1808 11 17 99.07% 99.40% 231 Female, age 72 1253 1250 9 3 99.76% 99.29% 232 Female, a ge 76 397 318 29 79 80.10% 91.64% T able 4 6. Clustering results for f usion of Paced and Normal beats. Record Record Patient # QRS TP FP FN Sensitivity Predictivity 104 Female, age 66 1379 1360 145 19 98.62% 90.37% 217 Male, age 65 1542 1021 160 521 66. 21% 86.45%

PAGE 85

85 Table 4 7. Number of cache entries used during the simulation. Record Patient Number of Used Entries Record Patient Number of Entries 100 Male, age 69 3 200 Male, age 64 11 101 Female, age 75 5 201 Male, age 68 5 102 Female, age 84 1 202 Male, age 68 5 103 Male, age N/A 1 203 Male, age 43 13 104 Female, age 66 4 205 Male, age 59 5 105 Female, age 73 14 207 Female, age 89 9 106 Female, age 24 6 208 Female, age 23 9 107 Male, age 63 13 209 Male, age 62 3 108 Female, age 87 2 210 Male, age 89 4 109 Male, age 64 4 212 Female, age 32 6 111 Female, age 47 2 213 Male, age 61 14 112 Male, age 54 1 214 Male, age 53 14 113 Female, age 24 2 215 Male, age 81 9 114 Female, age 72 6 217 Male, age 65 14 115 Female, age 39 2 219 Male, age N/A 13 116 Male, age 68 12 220 Female, age 87 3 117 Male, age 69 2 221 Male, age 83 7 118 Male, age 69 6 222 Female, age 84 3 119 Female, age 51 7 223 Male, age 73 9 121 Female, age 83 2 228 Female, age 80 14 122 Male, age 51 2 230 Male, age 32 5 123 Fe male, age 63 5 231 Female, age 72 4 124 Male, age 77 8 232 Female, age 76 7 233 Male, age 57 12 234 Female, age 56 2

PAGE 86

86 Figure 4 1 5 Number of used cache entries for 48 data sets. Table 4 8 Performance comparison results for two beat classificat ion algorithms. Armato et al. Heart Beat Cache Heartbeat Record Sensitivity +P Heartbeat Sensitivity +P Normal 119, 200, 209, 212, 217, 221, 231, 233 --Normal 95.83% 89.92% PVC 119, 200, 214, 221, 233 94.10% 95.43% PVC 55.75% 61.74% Paced 107, 217 99.69% 92.17% Paced 91.69% 92.10% LBBB 109,111, 214 96.05% 94.74% LBBB 95.83% 94.74% RBBB 118,124, 212,231, 232 92.45% 95.16% RBBB 95.50% 96.63% Fusion 217 91.53% 22.71% Fusion 82.42% 88.41%

PAGE 87

87 Figure 4 1 6 Performance comparison for two beat classi fication algorithms.

PAGE 88

88 CHAPTER 5 HARDWARE IMPLEMENTAT ION 5 .1 ECG System ECG systems should ideally be able to operate in a continuous and autonomous fashion without human intervention Several hardware char acteristics needed to achieve a fully autonom ous and self contained system are : (1) Autonomous beat template generation: The template matching algorithm needs known beat waveforms before execution begins In current practice, a template is often formed by averaging the manually selected beat s from a set o f test data. For autonomous template gener ation, our system generates beat templates by continuously learning from the detected beats without user input Even with drastic extreme shape changes, a new template can be quickly regenerated. ( 2) Real time onl ine beat classification : For a vital signs monitoring system, fast and autonomous beat classification is required for both bandwidth reduction and online diagnosis. Most current beat classification algorithms require a manual definition of each cluster bef ore classification begins, and thus these algorithms cannot be initialized when a number of beat clusters and templates are unknown. To avoid this disadvantage, we propose a real time online beat classification architecture that can dynamically determine t he number of beat clusters and store the relevant characteristics Figure 5 1. Single ECG processor block diagram.

PAGE 89

89 5 2 RTL Implementation In this chapter a hardware implementation for a self contained ECG pr ocessing architecture is describ ed i n register transfer level (RTL) The datapath consists of application specific functional units for arith metic and conditional logic as well as the memory space f or beat templates and the Heart Beat Cache As shown in Figure 5 3, there are four stages in the Fig ure 5 2. Detailed ECG processor block diagram. datapath the QRS detector, beat templates, Euclidean dis tance calculator, and the Heart Beat Cache wi th memory controller. A n equal number of clock cycles is required for the execution of each stage offering the opportunity for a pip eline operation to accommodate beat s from other leads in close temporal proximity. 5.2.1 QRS Detector T he QRS detector, which uses threshold crossing method, has a FIFO which is 64 deep with an 8 bits in width. The QRS detection i s performed by comparing the amplitude of the 38th sample in the FIFO with the threshold. If the amplitude crosses the threshold, a 64 cycle wide

PAGE 90

90 pulse will be generated by the pulse generator. During which the entire beat wavefo rm will be received by the beat classifier At other times the beat classifier ignores the incoming samples. 5 2 .2 Beat Classifier Upon recei ving notification from the QRS detector, one of t he template memory locations will be filled with the complete detected beat waveform, while t he other is kept fixed with a known beat shape as reference. The Eucli dean Distance calculator determine s the Euclidean distance between two beat templates and is implemented using the arithmetic IPs in the Synopsys DesignWare Library. Figure 5 3. Sin gle channel processor datapath block diagram. The memory controller oversees the cache read/write operations and beat classification alg orithm executions. U pon notification from the Euclidean calculator that the newly calculated distance is ready, the cont roller scans the cache and compares the distance with the store d max min distance range for each cluster entry. Th e appropriate actions to be taken depend on the cluster s metadata. The cycle accurate timing diagram of the memory controller is given in Fig ure 5 5.

PAGE 91

91 Figure 5 4. Timing diagram for the processor pipeline datapath and the detailed Heart Beat Cache read/write operations of the memory controller. 5 2 .3 Function al Simulations Functionality is verified using Cadence NCSim with previously generate d ECG data sets. The data set is first converted from the Matlab file format to the NCSim compatible binary format, and then t he cycle based simulation results are closely examined to ensure that th e fixed point RTL design is functionally equivalent to the floating point algorithms in Matlab.

PAGE 92

92 Screenshots, from the NCSim program using MIT BIH ECG data are given in Figure 5 6 and 5 7. Figure 5 6 shows the function simulation waveforms for the QRS detector, and Figure 5 7 shows the waveforms for beat classifie r Figure 5 5. Function simulation waveforms for QRS detector in NCSim. Figure 5 6. Function simulation waveforms for beat classifier in NCSim.

PAGE 93

93 5 3 FPGA Implementation 5 3 .1 Synthesis The implementation was synthesized using Xilinx ISE 10.1 and mappe d on to the Virte x IV (4VSX35FF668 10) FPGA chip The synthesis results are listed in Table 5 1. Table 5 1. FPGA synthesis report. Function Block Slices 4 input LUTs Flip Flops QRS Detector 254 (69.8%) 550 (88.0%) 158 (52.1%) Euclidean Calculator 57 (15. 7%) 44 (7.0%) 72 (23.8%) Cache Controller 53 (14.6%) 31 (5.0%) 73 (24.1%) Total 364 625 303 4VSX35FF668 10 15360 30720 30720 5 3 .2 Memory Two ZBT SRAMs are placed on the FPGA board and used to store the beat templates (128 bytes) and the Hea rt Beat Cache (80 bytes). Figure 5 7. ZBT SRAM interface.

PAGE 94

94 5.4 Optimizations To further improve the performance and reduce the hardware cost of a single channel ECG processor, there are some architecture level optimizations that may be considered. For e xample, the study on how to determine the optimum number of rows in cache has been described in Chapter 2. Nonetheless, there are some parameter settings that need to be investigated. 5.4.1 FIFO size FIFOs in the QRS detector are used to store detected bea t waveforms. The size of FIFO is mainly determined by typical beat width and ADC digitization sampling rate. The beat widths vary between 100 ms (QRS interval) to 600 ms (PR interval + QT interval), which are equivalent to 25 samples to 150 samples using 2 50 Hz sampling rate. A typical beat waveform is shown in Figure 5 1. Larger FIFO contains more noise than the targeted waveform, while smaller FIFO might lose some important features in a complete beat waveform such as P wave and U wave. From a hardware pe rspective, smaller FIFO consumes less power. There is a need for a quantitative study on the relationship between FIFO size and the beat classification accuracy. 5.4.2 Cluster Size The cluster size largely affects the accuracy of beat classification. Large r cluster size results in unwanted mix of two different beat shapes. Smaller cluster size produces more clusters than necessary. A fixed size cluster scheme is used in Chapter 3. However, an adaptive and optimized cluster size could potentially achieve bet ter accuracy. Quantitative study on the Euclidean distance variation will be conducted to get the optimal initial cluster size. 5.4.3 Number of Templates A beat shape is located in a feature space by calculating its Euclidean distances with pre selected t emplates. The more templates are used to locate a beat, the more dimensions a feature

PAGE 95

95 space has, thus the more variability for a beat shape to differentiate itself For example, in the case of two templates, the coordinate of each detected beat is consid ered as two Euclidean distances with templates or values along two axes of a two dimension plane. Figure 5 8. Typical heart beat waveform.

PAGE 96

96 CHAPTER 6 CONCLUSION 6 .1 Conclusive Comments Wearable devices allow the monitoring of vital signs at home, and enable wirelessly transmitting of physiological data to a remote diagnostic center. In this work, we focus on monitoring the ECG signal, which is a cruc ial tool for cardiologists to diagnose the condition of a heart The electrocardiogram (ECG) measure s the electrical activity of the heart, and is probably the most representative and noninvasive test to diagnose cardiac disorders. The identification of an arrhythmia event involves detecting heart beats automatically, discriminating abnormal beats from n ormal beats, and identifying the type of arrhythmia. For an automatic ECG diagnosis algorithm, retrieving characteristic points is essential. The QRS complex is the most significant characteristic waveform set of the cardiac cycle and corresponds to the de polarization of the ventricles in the heart. We evaluated the performance of different QRS detection algorithms, and selected one of the algorithms for our system. We proposed a heart beat classification architecture, Heart Beat Cache, that benefits from t he cache concept in Computer Architecture. The idea is to maintain a small table with low power such that it can fit into the small medical devices. The information recorded in the cache provides necessary matching attributes to clustering the he art beats Finally, the complete system of a n a utonomous single channel processor for ECG monitoring is presented. The VLSI implementation with the pipeline datapath is described and a cycle based timing diagram is given. The RTL implementation was functionally veri fied and was used for the generation of the FPGA synthesis report.

PAGE 97

97 6.2 Future Lines of Research 6.2.1 Explore More Databases To further test the performance of the architecture, we plan to use more arrhythmia databases from PhysioBank, such as MIT BIH Supr aventricular Arrhythmia Database MIT BIH Malignant Ventricular Arrhythmia Database and MIT BIH Atrial Fibrillation Database 6.2.2 Intervals and Waves To further distinguish different types of abnormal beats, characteristics that are not mentioned previo usly need be considered, including P wave, PR interval, QRS interval, QT interval, and QTc interval. For example, APB ( Atrial Premature Beat ) has slightly prolonged PR interval, and the RR interval between two QRS enclosing APB i s less than twice the norma l RR interval. 6.2.3 Heart Rate Variability The rhythm of a healthy heart is actually surprising irregular even under resting conditions. Heart rate variability (HRV) has been widely studied as an indicator of the entrainment between brain and heart, as we ll as of cardiac performance after severe cardiac diseases and operations. Figure 6 1. Sample plot for APB.

PAGE 98

98 LIST OF REFERENCES [1] B. U. Kohler, C. Hennig, and R. Orglmeister, "The principles of software QRS detection," Engineering in Medicine and Biolog y Magazine, IEEE vol. 21, no. 1, pp. 42 57, 2002 [2] A. Armato, E. Nardini, Antonio Lanata, Gaetano Valenza, C. Mancuso, E P. Scilingo, and D. De Rossi, ISDA pp. 660 664 2009. [3] L. S IEEE Eng. Med. Biol. Mag. vol. 14, no. 2, pp. 167 173, Mar./Apr. 1995. [4] Y. H. Hu, W. J. Tompkins, J. L. Urrusti, and V. X. Afonso, J. Electrocardiol. vol. 26, pp. 66 73, 1993. [5] adaptable ECG beat classifier using a mixture of experts approach IEEE Trans. Biomed. Eng. vol. 44, no. 9, pp. 891 900, Sep. 1997. [6] E. Patterson and B. J. Scherlag, "Decremental conduction in the posterior and anterior AV nodal inputs," Journal of Interventional Cardiac Electrophysiology vol. 7, pp. 137 148, Oct. 20 02. [7] S. S. Barold, "Willem Einthoven and the birth of clinical electrocardiography a hundred years ago," Cardiac Electrophysiology Review vol. 7, pp. 99 104, J an 2003 [8] E. Besterman and R. Creese, "Waller: Pioneer of electrocardiography," British Heart Jo urnal no. 42, pp. 61 64, 1979 [9] M. Rivera Ruiz, C. Cajavilca, and J. Varon, "Einthoven's string galvanometer: The first electrocardiograph," Texas Heart Institute Journal vol. 35, no. 2, pp. 174 178, 2008 [10] A. R. P erez Riera, C. Ferreira, C. F. Filho, M Ferreira, A. Meneghini, A. H. Uchida, E. Schapachnik, S. Dubner, and L. Zhang, "The enigmatic sixth wave of the electrocardiogram: The U wave," Cardiology Journal no. 5, pp. 408 421, 2008 [11] R. C. Bourge, W. T. Abraham, P. B. Ada mson, M. F. Aaron, A. Mag alski, M. R. Zile, A. L. Smith, F. W. Smart, M. A. O'Shaughnessy, M. L. Jessup, B. Sparks, D. L. Naftel, and L. W. Stevenson, "Randomized controlled trial of an implantable continuous hemodynamic monitor in patients with advanced heart failure: The COMPASS HF Study," Journal of the American College of Cardiology vol. 51, no. 11, pp. 1073 1079, 2008

PAGE 99

99 [12] F. Braunschweig, B. Kjellstor m, M. S odeh all, N. Clyne, and C. Linde, "Dynamic changes in right ventricular pressures during haemodialysis recorded with an imp lantable haemodynamic monitor," Nephrology Dialysis Transplantation vol. 21, no. 1, pp. 176 183, 2006 [13] D. W. Reynolds, N. Bartelt, R. Taepke, and T. D. Bennett, "Measurement of pulmonary artery diastolic pressure from the right ventricle," Journal of the American College of Cardiology vol. 25, no. 5, pp. 1176 1182, 1995 [14] S. C. Malpas, "Neural infl uences on cardiovascular variability: possibilities and pitfalls," American Journal of Physiology Heart and Circulatory Physiology vol. 282, no. 1, pp. 16 2 0, 2002 [15] C. Guger, G. Edlinger, R. Leeb, G. Pfurtscheller, A. Antley, M. Garau, A. Brogni, D. Friedman, and M. Slater, "Heart rate variability and eventrelated ECG in virtual environments," in Presence 2004: The 7th Annual International Workshop on Presen ce pp. 240 245, 2004 [16] N. Thakor, J. Webster, and W. Tompkins, "Estimation of QRS complex power spectrum for design of a QRS filter," IEEE Trans. Biomed. Eng. vol. 31, no. 11, pp. 702 706, 1984 [17] S. Ding, X. Zhu, W. Chen, and D. Wei, "Derivation of respi ratory signal from single channel ECGs based on source statistics," International Journal of Bioelectromagnetism vol. 6, no. 1, 2004 [18] G. B. Moody, R. G. Mark, A. Zoccola, and S. Mantero, "Derivation of respiratory signals from multilead ECGs," Computers in Cardiology vol. 12, pp. 113 116, 1985 [19] J. O. Wisbeck, A. K. Barros, and R. G. Ojeda, "Application of ICA in the separation of breathing artifacts in ECG signals," IEEE Trans. Biomed. Eng. 1998 [20] M. Costa, A. L. Goldberger, and P. C. K., "Multiscale en tropy analysis of biological signals," Physical Review E vol. 71, 2005 [21] L. A. Nunes Amaral, A. L. Goldberger, P. C. Ivanov, and H. E. Stanley, "Modeling heart rate variability by stochastic feedback," Computers in Cardiology 1999 [22] H. X. Zhang, Y. S. Zh u, and Z. M. Wang, "Complexity measure and complexity rate information based detection of ventricular tachycardia and fibrillation," Medical and Biological Engineering and Computing vol. 38, pp. 553 557, 2000 [23] F. Gritzali, "Towards a generalized scheme f or QRS detection in ECG waveforms," Signal Processing vol. 15, pp. 183 192, Sept. 1988

PAGE 100

100 [24] E. Skordalakis, "Syntactic ECG processing: A review," Pattern Recognition vol. 19, no. 4, pp. 305 313, 1986 [25] R. Poli, S. Cagnoni, and G. Valli, "Genetic design of o ptimum linear and non linear QRS detectors," IEEE Transactions on Biomedical Engineering vol. 42, pp. 1137 1141, Nov 1995 [26] L. Sornmo, O. Pahlm, and M. E. Nygards, "Adaptive QRS detection: A study of performance," IEEE Transactions on Biomedical Engineeri ng vol. BME 32, pp. 392 401, Jun 1985 [27] R. Murray, S. Kadambe, and G. F. Boudreaux Bartels, "Extensive analysis of a QRS detector based on the dyadic wavelet transform," in Time Frequency and Time Scale Analysis, 1994., Proceedings of the IE EE SP Internat ional Symposium pp. 540 543, Oct 1994 [28] N. Arzeno, Z. D. Deng, and C. S. Poon, "Analysis of first derivative based QRS detection algorithms," IEEE Transactions on Biomedical Engineering vol. 55, pp. 478 484, Feb. 2008 [29] M. Aboy, J. McNames, T. Thong, D. Tsunami, M. S. Ellenby, B. Goldstein, and A. Member, "An automatic beat detection algorithm for pressure signals," IEEE Transactions on Biomedical Engineering October 2005. [30] M. Okada, "A digital filter for the QRS complex detection," IEEE Transactions on Biomedical Engineering vol. BME 26, pp. 700 703, Dec 1979 [31] F. Chiarugi, V. Sakkalis, D. Emmanouilidou, T. Krontiris, M. Varanini, and I. Tollis, "Adaptive threshold QRS detector with best channel selection based on a noise rating system," in Computers in Cardiology pp. 157 160, October 2007 [32] I. I. Christov, "Real time electrocardiogram QRS detection using combined adaptive threshold," BioMedical Engineering OnLine vol. 3, no. 1, p. 28, 2004 [33] W. A. H. Engelse and C. Zeelenberg, "A single scan algorithm for QRS detection and feature extraction," Computers in Cardiology vol. 6, pp. 37 42, 1979 [34] J. Pan and W. J. Tompkins, "A re al time QRS detection algorithm ," IEEE Transactions on Biomedical Engineering vol. 32, pp. 230 236, March 1985 [35] P. Baldi, S. Br unak, Y. Chauvin, C. A. F. Andersen, and H. Nielsen, "Assessing the accuracy of prediction algorithms for classification: an overview," Bioinformatics vol. 16, no. 5, pp. 412 424, 2000

PAGE 101

101 [36] J. Thomas, J. Lewis K. W. Clark, C. N. Mead, K. L. Ripley, B. F. Sp enner, J. Oliver, and G. Charles, "Automated cardiac dysrhythmia analysis," Proceedings of the IEEE vol. 67, pp. 1322 1337, Sep 1979 [37] M. Botsivaly, C. Koutsourakis, and B. Spyropoulos, "Evaluation of a new technique for the detection of ventricular fibri llation and ventricular tachycardia," in 22nd Annual EMBS International Conference pp. 938 941, 2000 [38] J. Millet Roig, J. Lopez Soriano, A. Mochol, R. Ruiz Granell, and R. Garcia Civera, "An additional and easy algorithm based on morphological analysis of unipolar electrogram: M issclassification in implantable cardioverters defibrillators," Computers in Cardiology vol. 25, pp. 561 564, 1998 [39] M. Engin, "ECG beat classification using neuro fuzzy network," Pattern Recognition Letters vol. 25, pp. 1715 1722 2004 [40] M. Bahmanyar and W. Balachandran, "Beat to beat classification of long electrocardiograms using entropies and hierarchical clustering," in IEEE Engineering in Medicine and Biology 27th Annual Conference pp. 5579 5581, 200 5. [41] J. Nolan, P. D. Batin R. Andrews, S. J. Lindsay, P. Brooksby, M. Mullen, W. Baig, A. D. Flapan, A. Cowley, R. J. Prescott, J. M. Neilson, and K. A. Fox, "Prospective study of heart rate variability and mortality in chronic heart failure: Results of the united kingdom heart fa ilure evaluation and assessment of risk trial (UK Heart)," Circulation vol. 98, no. 15, pp. 1510 1516, 1998. [42] G. Zuanetti, J. M. Neilson, R. Latini, E. Santoro, A. P. Maggioni, and D. J. Ewing, "Prognostic significance of heart rate variability in post my ocardial infarction patients in the fibrinolytic era: The GISSI 2 results," Circulation vol. 94, no. 3, pp. 432 436, 1996. [43] H. V. Huikuri, T. Seppanen, M. J. Koistinen, K. J. Airaksinen, M. Ikaheimo, A. Castellanos, and R. J. Myerburg, "Abnormalities in b eat to beat dynamics of heart rate before the spontaneous onset of life threatening ventricular tachyarrhythmias in patients with prior myocardial infarction," Circulation vol. 93, no. 10, pp. 1836 1844, 1996. [44] J. Bigger, J. Thomas, J. L. Fleiss, R. C. St einman, L. M. Rolnitzky, R. Kleiger, and J. Rottman, "Frequency domain measures of heart period variability and mortality after myocardial infarction," Circulation vo l. 85, no. 1, pp. 164 171, 1992 [45] J. Bigger, J. Thomas, R. C. Steinman, L. M. Rolnitzky, J. L. Fleiss, P. Albrecht, and R. J. Cohen, "Power law behavior of RR interval variability in healthy middle aged persons, patients with recent acute myocardial infarction, and patients with heart transplants," Circulation vol. 93, no. 12, pp. 2142 2151, 1996.

PAGE 102

102 [46] H. V. Huikuri, T. H. Makikallio, K. E. J. Airaksinen, T. Seppanen, P. Puukka, I. J. Raiha, and L. B. Sourander, "Power law relationship of heart rate variability as a predictor of mortality in the elderly," Circulation vol. 97, no. 20, pp. 2031 203 6, 1998. [47] K. K. L. Ho, G. B. Moody, C. K. Peng, J. E. Mietus, M. G. Larson, D. Levy, and A. L. Goldberger, "Predicting survival in heart failure case and control subjects by use of fully automated methods for deriving nonlinear and conventional indices of heart rate dynamics," Circulation vol. 96, no. 3, pp. 842 848, 1997. [48] H. V. Huikuri, T. H. Makikallio, C. K. Peng, A. L. Goldberger, U. Hintze, and M. Moller, "Fractal correlation properties of R R interval dynamics and mortality in patients with depresse d left ventricular function after an acute myocardial infarction," Circulation vol. 101, no. 1, pp. 47 53, 2000. [49] R. W. Glenny, H. T. Robertson, S. Yamashiro, and J. B. Bassingthwaighte, "Applications of fractal analysis to physiology," J Appl Physiol vo l. 70, no. 6, pp. 2351 2367, 1991. [50] P. C. Ivanov, L. A. Nunes Amaral, A. L. Goldberger, S. Havlin, M. G. Rosenblum, H. E. Stanley, and Z. R. Struzik, "From 1/f noise to multifractal cascades in heartbeat dynamics," Chaos vol. 11, no. 3, pp. 641 652, 2001. [51] U. Rajendra Acharya, K. Paul Joseph, N. Kannathal, C. M. Lim, and J. S. Suri, "Heart rate variability: a review ," Med Biol Eng Comput vol. 44, pp. 1031 1051, December 2006. [52] A. Am ann, R. Tratnig, and K. Unterko er, "Reliability of old and new ventricular fibrillation detection algorithms for automated external defibrillators," BioMedical Engineering OnLine vol. 4, no. 1, p. 60, 2005. [53] L. J. T. Jr., K. W. Clark, C. N. Mead, K. L. Ripley, B. F. Spenner, and G. C. O. Jr., "Automated cardiac dysrhythmia anal ysis," Proceedings of the IEEE vol. 67, pp. 1322 1337, september 1979. [54] T. Dahan and A. Sade, "Application of ICA in removing a rtifacts from the ECG," Master's thesis, Israel Institute of Technology, 2004 2005. [55] N. Thakor, Y. Zhu, and K. Pan, "Ventricular tachycardia and fibrillation detection by a sequential hypothesis testing algorithm," IEEE Trans Biomed Eng vol. 37, no. 9, pp. 837 843, 1990. [56] G. Ersoy and M. Holewijn, "Development of a QRS detector algorithm for ECG's measured during centrifuge traini ng," in Proceedings of the 18th Annual International Conference of the IEEE vol. 4, pp. 1390 1391 vol.4, Nov 1996.

PAGE 103

103 [57] C. M. van Ravenswaaij Arts, L. A. Kollee, J. C. Hopman, G. B. Stoelinga, and H. P. van Geijn, "Heart rate variability ," Ann Intern Med vol 118, pp. 436 447, March 1993. [58] C. K. Peng, S. Havlin, J. Hausdor, J. Mietus, H. Stanley, and A. Goldberger, "Fractal mechanisms and heart rate dynamics: Longrange correlations and their breakdown with disease," Journal of Electrocardiology vol. 28, no. Supplement 1, pp. 59 65, 1995. Research and Technology Transfer in Computerized Electrocardiology. [59] Fast and Accurate FPGA Based QRS Detection System 30th Annual EMBS International Conference pp. 4828 4831, 2008. [60] M. based System for ECG Beat Detection and Classification IFMBE Proceedings 16, Medicaon 2007 pp. 66 69, 2007. [61] Elmore, Kimberly L., Michae Euclidean distance as a s imilarity metric for principal component a nalysis Monthly Weather Review 129, 540 549 [62] J. L. Hennessy and D. A. Patterson, Computer a rchitecture: A q uantita tive approach (third edition ) Morgan Kaufmann Publishers San Franc isco, CA, USA. [63] F. Castells, P. Laguna, L. Sornmo, A. Bollmann and J. M. Roig URASIP Journal on Advances in Signal Processing Vol. 2007. [64] Proceedings of PDCS'Ol pages 617 662, Aug 2001. [65] M Ze ng, J. G. Lee, G. S. Choi and J. A IEICE Electron. Express Vo l. 6, No. 9, pp.560 565, 2009 [66] Al A dynamic Fourier series for the compression of ECG using FFT an d adaptive c Medical Engineering and Physics 17, 197 203.

PAGE 104

104 BIOGRAPHICAL SKETCH Chung Ching P eng was born in Taipei, Taiwan, ROC The older of two children, h e grew up mostly in Taipei, Taiwan graduating f rom Affiliated Senior High School of N ational Taiwan N ormal University ( HSNU) in 1997 Chung Chi ng received the B.S. degree in Electrical and Control E ngineering from National Chiao Tung University (NCTU) Hsinchu, Taiwan, in June 2001. In July 2001 Chung Ching enlisted in the Navy as an Ens ign for his mandatory military service. During his service in the Navy ( 1 year and 8 months ) Chung Ching was a b attalion level human resource and computer system officer in charge Upon completion of his military duty in March 2003 Chung Ching was then e nrolled in the Computer and Information Science and Engineering Department (CISE) at University of Florida as a teaching assistant in August 2003, an d Electrical and Computer Engineering Departme nt (ECE) as a n Alumni Fellowship r ecipient in August 2006 His research interests are computer architecture, low power VLSI and biomedical signal processing. His research advisor is Dr. Jih Kwon Peir. Chung Ching received the M.S. degree in Computer and Information Science and Engineering (CISE) and the Ph.D. deg ree in E lectrical and Computer Engineering (ECE) from University of Florida in August 2006 and April 2011 respectively He will work at MaxLinear Inc., in Carlsbad, CA with their ASIC group as a Member of Technical Staff.