Citation
Multivariate Pattern Decoding of fMRI Signals in Disorders of Consciousness

Material Information

Title:
Multivariate Pattern Decoding of fMRI Signals in Disorders of Consciousness
Creator:
Opri, Enrico
Publisher:
University of Florida
Publication Date:
Language:
English

Thesis/Dissertation Information

Degree:
Master's ( M.S.)
Degree Grantor:
University of Florida
Degree Disciplines:
Biomedical Engineering
Committee Chair:
SITARAM,RANGANATHA
Committee Co-Chair:
GUNDUZ,AYSEGUL
Committee Members:
LANG,PETER J

Subjects

Subjects / Keywords:
Datasets ( jstor )
Magnetic resonance imaging ( jstor )
Mental imagery ( jstor )
Pain perception ( jstor )
Paradigms ( jstor )
Persistent vegetative states ( jstor )
Preprocessing ( jstor )
Searchlights ( jstor )
Signals ( jstor )
Toolboxes ( jstor )
classification
doc
fmri

Notes

General Note:
Evaluation of patients with disorder of consciousness (DOC) is conducted prevalently with bedside assessment. However, this approach has been known to produce misdiagnosis. Recent studies propose neuroimaging as a new diagnostic tool, potentially leading to the implementation of a more robust methodology to classify the patients depending on their brain activations, as minimally conscious status (MCS) or persistent vegetative state (PVS). A hierarchical protocol has been suggested for functional neuroimaging studies (Owen et al., 2005), progressing sequentially from the simplest form of brain processing to more complex cognitive functions. It was also demonstrated that pattern classification of brain signals in different behavioral tasks could enable DOC patients to use a binary fMRI-BCI (Boly et al., 2007). In light of the above, our study aimed to take this approach to a next step, by trying to assess new methodologies to classify the patients in the different clinical DOC categories, as minimally conscious status (MCS) or persistent vegetative state (PVS), by using functional Magnetic Resonance Imaging (fMRI) in a battery of experiments on intentional control, language competence, working memory, emotions and pain sensation. During this project, the MANAS 4 toolbox was implemented based on a revised version of a previously developed toolbox (Rana et al., 2013) for fMRI data classification. The toolbox used multivariate Support Vector Machine (SVM), and conducted a sub-study to assess the performances of different preprocessing steps (Tanabe et al., 2002) and feature selection algorithms, namely, Fisher Scoring (Gu, Li, & Han, 2012b), Fisher Scoring with Searchlight and Effect Mapping (Lee et al., 2010) using a motor task fMRI dataset (Rana et al., 2013). The toolbox allows a high degree of customization through configuration files, letting the researcher to focus on the analysis of the data, rather than the building of the processing. In the subsequent stages of this project, the MANAS 4 toolbox was used to analyze activations in a Classical Conditioning Paradigm with Emotional Sounds, to evaluate the possibility of a future implementation of an fMRI-BCI for communication in Alzheimer patients, by showing an overall good prediction of the conditioned stimuli.

Record Information

Source Institution:
UFRGP
Rights Management:
All applicable rights reserved by the source institution and holding location.
Embargo Date:
8/31/2016

Downloads

This item has the following downloads:


Full Text

PAGE 1

MULTIVARIATE PATTERN DECODING OF FMRI SIGNALS IN DISORDERS OF CONSCIOUSNESS By ENRICO OPRI A THESIS PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE O F MASTER OF SCIENCE UNIVERSITY OF FLORIDA 2014

PAGE 2

© 2014 Enrico Opri

PAGE 3

To my father and mother, who always support ed me in my dreams

PAGE 4

4 ACKNOWLEDGMENTS I thank my advisor, Dr. Sitaram, for guiding me during this year as an exchange student, by proposing new methodologies and involving every member of our lab with never ending energy. I thank Josue and Mohit for keeping company in the lab even for two days straight without sleep and Aniruddh who helped me in v arious thorny situations. Special thanks also go to my friends here in Gainesville, to all the people that I encountered here and the Atlantis students who came to Gainesville together with me .

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 7 LIST OF FIGURES ................................ ................................ ................................ .......... 8 ABSTRACT ................................ ................................ ................................ ................... 10 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .... 12 2 BACKGROUND ................................ ................................ ................................ ...... 17 Real time Functional Magnetic Resonance Imaging ................................ ............... 17 Decoding Mental State ................................ ................................ ............................ 18 Pattern Classification ................................ ................................ .............................. 19 Support Vector Machines ................................ ................................ ........................ 21 Disorders of Consciousness ................................ ................................ ................... 24 Bedside Assessment ................................ ................................ .............................. 26 Classical Conditioning ................................ ................................ ............................. 2 6 3 MATERIALS AND METHODS ................................ ................................ ................ 31 Overview ................................ ................................ ................................ ................. 31 Preprocessing of fMRI D atasets ................................ ................................ ............. 31 Feature Selection Algorithms ................................ ................................ .................. 33 Fisher Scoring ................................ ................................ ................................ .. 33 Fisher Scoring with Searchlight (FSSL) ................................ ............................ 34 Effect Mapping ................................ ................................ ................................ . 34 4 CONDITIONING WITH EMOTIONAL SOUNDS: TOWARDS AN FMRI BCI .......... 38 Overview ................................ ................................ ................................ ................. 38 Methods ................................ ................................ ................................ .................. 38 Parti cipants and Acquisition ................................ ................................ ............. 38 St imuli ................................ ................................ ................................ ............... 39 Experimental Paradigm ................................ ................................ .................... 39 Behavioral Measures ................................ ................................ ........................ 40 Analysi s ................................ ................................ ................................ ............ 41 Results ................................ ................................ ................................ .................... 42 Behavioral Data SAM Ratings ................................ ................................ ........ 42 Neuroimaging C l assification ................................ ................................ ............. 43 Discussion ................................ ................................ ................................ .............. 43

PAGE 6

6 5 FMRI PATTERN CLASSIFICATION TO ASSESS DISORDERS OF CONSCIOUSNESS ................................ ................................ ................................ 49 Overview ................................ ................................ ................................ ................. 49 Methods ................................ ................................ ................................ .................. 49 Parti cipants and Acquisition ................................ ................................ ............. 49 Experimental Paradigms and Stimuli ................................ ................................ 50 Pain Perception Task ................................ ................................ ....................... 51 Empathy for Pain Task ................................ ................................ ..................... 51 Mental Imagery Task ................................ ................................ ........................ 51 Analysis ................................ ................................ ................................ ............ 52 Results ................................ ................................ ................................ .................... 55 Discussion ................................ ................................ ................................ .............. 56 6 MANAS 4 TOOLBOX ................................ ................................ .............................. 69 Overview ................................ ................................ ................................ ................. 69 Features of the Toolbox ................................ ................................ .......................... 69 How to Use ................................ ................................ ................................ ............. 70 7 TESTING MANAS 4 TOOLBOX WITH A MOTOR TASK DATASET ...................... 73 Overview ................................ ................................ ................................ ................. 73 Methods ................................ ................................ ................................ .................. 73 Parti cipants and Acquisition ................................ ................................ ............. 73 Experimental Paradigms and Stimuli ................................ ................................ 73 Analysis ................................ ................................ ................................ ............ 74 Results ................................ ................................ ................................ .................... 74 Discussion ................................ ................................ ................................ .............. 75 8 CONCLUSIONS ................................ ................................ ................................ ..... 78 Overview ................................ ................................ ................................ ................. 78 Emotion Based fMRI BCI ................................ ................................ ........................ 78 Analysis to Assess Disorders of Consciousness ................................ .................... 78 APPENDIX: MANAS 4 USAGE EXAMPLE ................................ ............................ 80 LIST OF REFERENCES ................................ ................................ ............................... 82 BIOGRAPHICAL SKETCH ................................ ................................ ............................ 90

PAGE 7

7 LIST OF TABLES Table page 4 1 ROI analysis scores based on AAL. ................................ ................................ ... 45

PAGE 8

8 LIST OF FIGURES Figure page 2 1 Linea rly separable data, binary case ................................ ................................ .. 29 2 2 Hierarchy of the different state s in Disorders of Consciousness ........................ 29 2 3 The 4 phases of Classical Conditioning in the Pavlov ian dog experiment .......... 30 3 1 Example of autocorrelation of the averaged BOLD time series before detrending without z normalization ................................ ................................ ..... 36 3 2 Example of autocorrelation of the avera ged BOLD time series after detrending without z normalization ................................ ................................ ..... 36 3 3 Time series of averaged BOLD before detrending (blue) and after detrending (red) without applying z normalization ................................ ................................ 37 4 1 Block design p a radigm used in the c lassical c onditioni n g experiment ................ 46 4 2 Classification acc uraci es using Effect Mapping as feature selectio n .................. 46 4 3 Classification accuracies using Fisher Scoring with Searchlight as feature selection ................................ ................................ ................................ ............. 47 4 4 Functional Brain Map of the features selecte d with Effec t Mapping in the early acquisition phase ................................ ................................ ....................... 47 4 5 Functional Brain Map of the features selected with Effec t Mapping in the late acquisition phase ................................ ................................ ................................ 48 4 6 Functional Brain Map of the features selected wit h Effect Mapping in the extinction phase ................................ ................................ ................................ .. 48 5 1 Pain perception paradigm ................................ ................................ ................... 58 5 2 Empathy for Pain paradigm ................................ ................................ ................ 59 5 3 Mental Imagery paradigm ................................ ................................ ................... 60 5 4 Series of DOC patient images ................................ ................................ ............ 61 5 5 Series of Anatomical T1 Images ................................ ................................ ......... 62 5 6 Classification accuracies of Pain Perception Task, using Fisher Scoring with Searchlight for feature selecti on with 10 fold cross validation ............................ 63

PAGE 9

9 5 7 Classification accuracies of Empathy for Pain Task, using Fisher Scoring with Searchlight for feature selection with 10 fold crossvalidation ............................. 63 5 8 Classification accuracies of Mental Imagery Task, using Fisher Scoring with Searchlight for feature selection with 10 fold cross validation ............................ 63 5 9 Functional brain map of the activations during the Pain P erception Task in a MCS patient ................................ ................................ ................................ ........ 64 5 10 Functional brain map of the activations during the Pain Perce ption Task in a healt hy control ................................ ................................ ................................ .... 64 5 11 Functional brain map of the activations during the Empathy for Pain Task in a PVS patient ................................ ................................ ................................ ......... 65 5 12 Functional brain map of the activations during the Empathy for Pain Task in a healthy control ................................ ................................ ................................ .... 65 5 13 Functional brain map of the activations during the Mental Imagery Task ersu ................................ 66 5 14 Functional brain map of the activations during the Mental Imagery Task ............................... 66 5 15 Functional brain map of the activations during the Mental Imagery Task ................................ 67 5 16 Functional brain map of the activations during the Mental Imagery Task ........................... 68 6 1 MANAS 4 Results Manager ................................ ................................ ................ 72 6 2 MANAS 4 Slice Viewer ................................ ................................ ....................... 72 7 1 Averaged c lassification accuracies for each method of feature selection used in the analysis (FS, FSSL, EM). ................................ ................................ .......... 76 7 2 Functional Brain Map of the featur es selected with Fisher Scoring .................... 76 7 3 Functional Brain Map of the features selected with Fisher Scoring with Searchlight ................................ ................................ ................................ .......... 77 7 4 Functional Brain Map of the features selected with Effect Mapping. .................. 77 A 1 Example of run script with common parameters for analysis. ............................. 80 A 2 Example of run script with common parameters for preprocessing. ................... 81

PAGE 10

10 Abstract of Thesis Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Master of Science MULTIVARIATE PATTERN DECODING OF FMRI SIGNA LS IN DISORDERS OF CONSCIOUSN ESS By Enrico Opri August 2014 Chair: Ranganatha Sitaram Major: Biomedical Engineering Evaluation of patients with disorder of consciousness (DOC) is conducted prevalently with bedside assessment. However, this approach has been known to produce misdiagnosis. Recent studies propose neuroimaging as a new diagnostic too l, potentially leading to the implementation of a more robust methodology to classify the patients depending on their brain activations , as minimally conscious status (MCS) or per sistent vegetative state (PVS). A hierarchical protocol has been suggested fo r functional neuroimaging studies (Owen et al., 2005) , progressing sequentially from the simplest form of brain processing to more complex cognitive functions. It was also demonstrated that pattern classification of brain signals in different behavioral tasks could enable DOC patie nts to use a binary fMRI BCI (Boly et al., 2007) . In light of the above, our study aimed to take this approach to a next step, by trying to assess new methodologies to classify the patients in the different clinical DOC categories, as minimally conscious stat us (MCS) or persistent vegetative state (PVS), by using functional Magnetic Resonance Imaging (fMRI) in a battery of experiments on

PAGE 11

11 intentional control, language competence, working memory, emotions and pain sensation. During this project, the MANAS 4 tool box was implemented based on a revised version of a previously developed toolbox (Rana et al., 2013) for fMRI data classification. The toolbox used mul tivariate Support Vector Machine (SVM), and conducted a sub study to assess the performances of different preprocessing steps (Tanabe et al., 2002) and feature selection algorithms, namely, Fisher Scoring (Gu, Li, & Han, 2012b) , Fisher Scoring with Searchlight and Effect Mapping (Lee et al., 2010) using a motor task fMRI dataset (Rana et al., 2013) . The toolbox allows a high degree of customization through configuration files, letting the researcher to focus on the analysis of the data, rather than the building of the processing. In the subsequent stages of this project, the MANAS 4 toolbox was used to analyze activations in a Classical Conditioning Paradigm with Emotional Sounds, to evalu ate the possibility of a future implementation of an fMRI BCI for communication in Alzheimer patients, by showing an overall good prediction of the conditioned stimuli .

PAGE 12

12 CHAPTER 1 INTRODUCTION Men ought to know that from nothing e lse but the brain come joys, delights, laughter and sports, and sorrows, griefs, despondency, and lamentations. And by this, in an especial manner, we acquire wisdom and knowledge, and see and hear, and know what are foul and what are fair, what are bad an d what are good, what are sweet, and what unsavory; some we discriminate by habit, and some we perceive by their utility. By this we distinguish objects of relish and disrelish, according to the seasons; and the same things do not always please us. And by the same organ we become mad and delirious, and fears and terrors assail us, some by night, and some by day, and dreams and untimely wanderings, and cares that are not suitable, and ignorance of present circumstances, desuetude, and unskillfulness . [ ] Sin ce, then, the brain, as being the primary seat of sense and of the spirits, perceives whatever occurs in the body, if any change more powerful than usual take place in the air, owing to the seasons, the brain becomes changed by the state of the air. For, o n this account, the brain first perceives, because, I say, all the most acute, most powerful, and most deadly diseases, and those which are most difficult to be understood by the inexperienced, fall upon the brain. (Hippocrates, 400 BC ) . the concern of philosophical and scientific studies during a large part of human history. It may seem surpri sing that there h as rather been slow progres s in the development of methods to assess it or to evaluate its implications. Although t obvious answer, we still find ourselves not being able to give a tan gible one. It may be beyond our conceptual grasp , even in the future (McGinn, 1993) . Th e Cartesian division of body and mind still survives in our culture and in the Scientific World, including Neurology and Psychiatry, while there is the need of a bridge between these two fields though they are apparently focused towards the same goal. From the point of view of a Neurologist it is intriguing how the number of parallel processors available in our brain, which are the neurons, are able to give ri se to self consciousness, and that is a similar

PAGE 13

13 wonder to all the Artificial Intelligence Field. It is beautifully expressed in this passage from Monadolog y : S upposing that there were a mechanism so constructed as to think, feel and have perception, we mi ght enter it as into a mill. And this granted, we should only find on visiting it, pieces which push one against another, but never anything by which to explain a perception . (Leibniz, 1714) . While Leibnitz associated consciousness with perception, he believed that mechanistic theories by themselves could not explain how the system gave ri se to consciousness. Many humans share the same intuition (Young & Wijdicks, 2008) , but they cannot infer its source . themselves. Only later it took the current meaning, being profoundly influenced by culture, religi on and philosophical history. consciousness as the waking state , the ability to perceive and interact with the external world, and consciousness as awareness , the feeling of being consc ious of something, aware of being awake and being able to process both external inputs and internal self reflection, leading to a sense of self, being self conscious of our existence. While it seems evident and easy to demonstrate when consciousness is pre sent, it is not easy to demonstrate the opposite, the concept of the unconscious process and where resides a threshold to define when we are truly conscious of a perception. Studies shows that conscious and unconscious processes cannot be dissociated (Peremen & Lamy, 2014) by their time course but there are different theories on the

PAGE 14

14 implication of the vision are the most apparent, as the Unconscious Thought Theory and The Rule Principle (Dijksterhuis & Nordgren, 2006; Claxton, 2000; Chrabaszcz & Dougherty, 2012) . The state of consciousness can be affected by different causes, traumatic or not, but they all lead to Disorders of Consciousness (DOC). One of the most known DOC is the Vegetative State, a possible state after coma, a condition where the patient becomes extremely unresponsive, not manifesting voluntary movement, behavior and even normal reflexes. The evaluation of robust markers for DOC in severely brain damaged pati ents is challenging. The current methodology is relying heavily on bedside assessment, an approach that has been known to produce misdiagnosis up to 40% (Andrews et al., 1996; Schnakers et al., 2009; Chi lds & Mercer, 1996) . Various studies (Ste nder et al., 2014; Sleigh & Warnaby, 2014) aimed to use neuroimaging to diagnose the patient status. It has been suggested that functional imaging studies should be conducted hierarchically (Owen et al., 2005) , beginning with the simplest form of processing within a particular dom ain and increase gradually the grade of cognitive function requested to process the proposed tasks. It has been shown (Boly et al., 2007) that pattern classification of brain activations could allow DOC patients to communicate with binary responses (yes no) . In light of the above, our study aimed to take this approach to a next step, by trying to assess new methods to classify the patients in the different clinical DOC categories, as minimally conscious status (MCS) or pe rsistent vegetative state (PVS). We ap ply functional Magnetic Resonance Imaging (fMRI) in a battery of experiments

PAGE 15

15 testing intentional control, language competence, working memory, emotions and pain sensation. Another condition that can be classified as moderate DOC is dementia, that is charac terized by a cognitive impairment (Liberati et al., 2012) . The use of Brain Computer Interface (BCI) in patients in this condition is seriously limited because of need for an active cognitive task required for the control of a BCI, as in,A disease (AD) which is a form of dementia. However, AD Patients were shown to be still able to process emotions and perform emotion recognition (Luzzi et al., 2007) . Hence Liberati et al . proposed a novel method to avoid the issue of cognitive impairment in AD patients, by using classical condition within an emotion based fMRI BCI. The study first tested the performance of the method in healthy subjects, by applying a se mantic double conditioning paradigm (congruent and incongruent word pairs) using emotional sounds (positive and negative) to condition a binary response (affirmative or negative) to allow a basic BCI communication. The main aim in this paradigm was to asse ss the discrimination performance between congruent and incongruent word pairs after conditioning. During the analysis of the above dataset in this, I have implemented MANAS 4, a revised version of a previously developed toolbox (Rana et al., 2013) for fMRI data classification with multivariate Support Vector Machine (SVM), and conducted a sub study to assess the performances of different preprocessing steps (Tanabe et al., 2002) and different feature selection algorithms, namely, Fisher Scoring (Gu, Li, & Han, 2012a) , Fisher Scoring with Searchlight (Pereira, Mitchell, & Botvinick, 2009) and Effect Mapping (Lee et al., 2010) .

PAGE 16

16 MANAS 4 is a tool to perform batch processing of data, needing only the preparation of JavaS cript Object Notation and allowing a high degree of customization, letting the researcher to focus on the analysis of the data, rather than the building of the processing. The robustness of the tools implem ented in the toolbox were tested on a previously published fMRI dataset (Rana et al., 2013) , which was collected during a motor execution paradigm.

PAGE 17

17 CHAPTER 2 BACKGROUND Real time Functional Magnetic Resonance Imaging Conventional Magnetic Resonance Imaging (MRI) has been a slow imaging modality with an underlying tradeoff between acquisition speed and signal to noise ratio (SNR) (Cohen, 2001) . The physical reason is that MR signal d erives from the translation of the magnetization of a sample to a radio signal during a phase defined as recovery. The recovery has a consistent delay being necessary to return the samples to a low energy state, limiting the possibility of a fast acq uisition. The signal detected is the Blood Oxygenation Dependent (BOLD) response, representing the change in blood flow and oxygenation that correlates with neural activity. It is based on the notion that an active brain region consumes more oxygen, hence there should be an increased blood flow to satisfy the consumption (Kwong et al., 1992; Ogawa et al., 1992) . It has an unavoidable delay in the presentation of the effect of the activation. This methodology allows to non invasively record brain activ ity. Numerous technical advancements allowed a substantial re duction in acquisition time, enabling a real time functional MRI (rtfMRI) approach. The most important speed advancement was brought by the development of the Echo planar Imaging (EPI) technique, capable of imaging the entire brain within 1 2 s. With this t emporal resolution, fMRI can accurately follow the time course of brain activation. In conventional fMRI experiments the image reconstruction can be executed after the experiment has been completed (offline). The real time fMRI instead, on which fMRI Brain Computer Interface (fMRI BCI) is based, requires the reconstruction of a

PAGE 18

18 volume after every acquisition of the MR signal (online) within the experiment, every Repetition Time (TR). Many factors have influence on both signal acquisition and real time perfor mance of fMRI: static magnetic field (B0) intensity, spatial resolution, temporal resolution, echo time and magnetic field inhomogeneities. It has to be considered that the increase of spatial resolution decreases the SNR and increases the acquisition tim e, therefore there will be a compromise between spatial and temporal resolution. Commonly in an fMRI BCI experiment a 64x64 image matrix, for each slice in a volume, is acquired, resulting in an in plane resolution of 3 4 mm, while it is used a slice thickness of 5 m m , resulting typically in 36 40 slices for each volume. Considering that online reconstruction and processing are performed, the use of spatial filter ing improves SNR and compensat in g for head motion and improving analysis of inter subject variability (Weiskopf et al., 2004) . Decoding Mental State Recent improvements in neuroimaging and analysis techniques (e.g. multivariate of a single individual based only on non invasive measurement of brain activity with functional MRI (Haynes & Rees, 2006) . A mental state can be decoded from fMRI signals with a multivariate pattern classification, as it was used in this work, by discriminating between patterns of activation clusters in the brain or by using a predefined Regions of Interest (ROIs) (e.g. in sula for emotional experiments) . The aim to decode mental states is to obtain an understanding of the human brain mechanisms

PAGE 19

19 in connection with different behaviors, and to enable more complex setup as the implementation of fMRI BCIs. These method ologie s have been applied in studies of visual perception showing that perceptual variation can be dynamically decoded during binocular rivalry in focused regions of the early visual cortex, using trained pattern of r esponses related to each monocular perception. The same approach was extended to different categories of brain states as emotions (Sitaram et al., 2011) , distinguishing between 4 different emotions elici ted, during the presentation of emotional pictures from the (IAPS) (Lang et al., 2008) , with 75% average accuracy. Hen ce, it seems feasible that the implementation of an fMRI BCI approach would allow people with impairments to learn volitional self regulation of the BOLD signal. Classical conditioning approaches can even be used to allow classification without having a co nscious response (Liberati et al., 2012) . Pattern Classification Prediction of mental states from brain activity is one of the major aims in fMRI studies, by find ing patterns of activations to classify between different conditions, allowin g discrimination between a set of expected behaviors . Such as approach would enable clinical treatment of problems such as the assessm ent of affect in people who can not communicate well, such as patients of dementia and DOC. Pattern based methods use mach ine learning techniques such as Multilayer Artificial Neural Network (MANN) or Support Vector Machines (SVM) to discriminate spatial, temporal and spectral patterns in a system with a multivariate approach, aiming to find discriminative patterns in a group of features in each instance of the considered dataset. These methodologies have been successfully used in different Computer

PAGE 20

20 Science problems such as face recognition, image recognition, and speech recognition (Jain, Duin, & Mao, 2000) . Mac hine learning techniques used in pattern classification can be grouped mainly as supervised techniques, when the class of each instance is known and available to train the classifier, and unsupervised techniques, when the grouping and the class of each ins tance is not available and they have to be inferred from the data. The growth of computational power and the advancements in machine learning applications have enabled the use of such methodologies with neuroimaging data, characterized by large datasets of high resolution brain scans at several time points. The classification can be conducted with a univariate approach, which considers each voxel as an independe nt channel/feature even if brain activity was recorded from a consistent number of locations simultan eously , or a multivariate approach, wh ich considers voxels as features , allowing to access an inherently higher information content. Because u nivariate approach es are tailored to detect statistically significant brain activations, they however ignore subt hreshold activations and hence are unable to detect physiologically relevant changes in the brain. On the otherhand, a multivariate approach is based on the insight that multiple spatially distributed locations have a synergic role during a task (Lee et al., 2010) . For these reasons the multivariate approach is growing in popularity in the scientific community (Haynes & Rees, 2005; Jimura & Poldrack, 2012). Interestingly a multivariate pattern classification approach can also be used as an objective cr iterion to evaluate the importance of different Regions of Interest (ROI) in a given task. This analysis can be simply executed by

PAGE 21

21 considering different groups of ROIs and comparing the classification performances in decoding the desired task (Haynes et al., 2007) . Pattern Classification can be used not only for discrimination s between different task related activations, but also to assess the consistency of brain activation across tasks, sessions or subjects and track temporal transitions of mental states. Multivariate Pattern Classification has been successfully used in sever al fMRI studies, including unconscious determinants of free decisions (Soon, B rass, Heinze, & Haynes, 2008) , visual processing (Kamitani & Tong, 2005) , and emotions (Mourão Miranda, Friston, & Brammer, 2007) , and fMRI BCI (Sitaram et al., 2011) . This work focused on the use of SVM, one of the most widel y used methods for classification of fMRI signals, due to its ability to deal with a high dimensional dataset with its relative robustness against the curse of dimensionality (LaConte, 2005; Mahmoudi , 2012) . Support Vector Machines SVM is a supervised learning method used for binary classification and regression. It is preferable to use linear SVM to provide a par simonious mapping between the brain locations (voxels) and the feature weights obtained after training the SVM. A linear SVM will find a separating hyperplane in an M dimensional space, where M is the number of features/voxels, using the information contai ned in the labeling (supervised method). Optimal separation of the instances is achieved by maximization of the margin solving a quadratic optimization problem. The margin is the distance between the separating hyperplane and the input vectors (i.e. suppor t vectors) closest to it (Schlkopf & Smola, 2001; Schölkopf, Bur ges, & Smola, 1998; Vapnik, 2000) .

PAGE 22

22 A common approach to evaluate the classification training results is to display the weight values at each voxel, and thus generating functional maps that show the most discriminating voxels related to a task. Lee and col leagues (2010) proposed an improved method defined as Effect Mapping that includes the information content of the Support Vectors, using the Mutual Information between the raw values and the related predicted values, into the computation of a functional ma p. Commonly SVM analysis requires each volume to be remapped to a 1 dimensional vector by considering it as a single instance, while the BOLD signal in each voxel is considered as a feature . Classification is implemented as follows , considering instance s (input vectors) with determined labels : ( 2 1 ) (possible values of the labels) (N: number of input vectors) Where the weight vector and the constant value , estimated by the SVM training algorithm from the training dataset, define a linear decision boundary , is a sign function, respectively. The single label can assume value 1 or 1, depending on which condition of interest they belong (e.g. +1 for hand movement, 1 for resting condition). The training of a classifier involves finding the hyperplane t hat best separates the training in stances (input vectors) in the feature space. The classical method is the hard margin SVM (Vapnik & Lerner, 1963) , which assumes that the dataset is linearly separable ( Figure 2 1 ) , therefore every instance must be correc tly separated (reassigned to its real class). A les s strict alternative is the

PAGE 23

23 soft margin SVM that allows some points to misclassified with the introduction of a slack variable, although it penalizes these points accordingly (Cortes & Vapnik, 1995) . This method is preferable in neuroimaging studies, where the data is unlikely to be perfectly separable given the underlying noise and behaviors. The weight vector , in the linear case, is obtained by maximizing the margin in Eq .( 2 2 ): ( 2 2 ) w ith constraint Eq .( 2 3 ) ( 2 3 ) This equals to solve the minimization problem i nverting Eq. ( 2 2 ) , described by the Eq.( 2 4 ) introducing also the soft margin : ( 2 4 ) with constraints Eqs . ( 2 5 ) and ( 2 6 ): , ( 2 5 ) ( 2 6 ) Where is the slack variable , (cost parameter) is the weight on the slack variable, that denotes the extent to which misclassification is allowed. Eqs. ( 2 4 ), ( 2 5 ), ( 2 6 ) can be rewritten in the Euler Lagrange form using Lagrange multipliers, described in Eq.( 2 7 ): ( 2 7 )

PAGE 24

24 Where is the Lagrange multiplier assuming non zero values when the corresponding input vector (instance) is a Support Vector, zero otherwise. Disorder s o f Consciousness Disorders of Consciousness (DOC) are medical condi tions resulting from interference with either or both cyclical wakefulness and awareness of self (Bernat, 2006) . There are different conditions in DOC, derived from dif ferent causes, traumatic or not. A Moderate DOC example is delirium and dementia, characterized by a cognitive impairment (Liberati et al., 2012) , but commonly they are not included in the DOC categories. The first phase of DOC is the coma ( Figure 2 2 ), characterized by the absence of arousal and consciousness (Laureys et al. , 2004) . It is a state where the patient is unresponsive and stimulation cannot elicit spontaneous periods of wakefulness and eye opening. Patients who survive usually recover within 2 4 weeks but in some cases the recovery does not progress further from the vegetative state or minimally conscious state. The Vegetative State (VS) is a condition that can be chronic and irreversible, defined as Persistent Vegetative State (PVS), or acute and reversible, defined as Minimally Conscious State (MCS). In the VS the patient has no awareness of self and the environment (B. Jennett & Plum, 1972; von Wild, Laureys, Gerstenbrand, Dolce, & Onose, 2012) and is incapable of interacting with others, even when the state characterized by the presence of sleep wake cycles. Patients continue to have preserve d autonomic functions including autonomous breathing and intact reflexes. The state can mislead diagnosis by the wakefuln ess and non purposeful m ovement that

PAGE 25

25 might lead doctors to think that the patient is conscious even if they are not. Patients can occa sionally perform complex behaviors such as shedding tears, grunting, screaming, smiling without a discernible reason (Zeman, 1997) . The problem underlying a possible misdiagnosis is the evaluation of the presence of unambiguou s sign of conscious perception or voluntary action. This evidence need to be assessed with a prolonged observation of the patient behavior. An acute case is Minimally Conscious State where the patient shows evidence of awareness with poor but existent resp onsiveness to external stimuli. Even if in both VS and MCS cases the patient is partially responsive, the MCS state maintains measurable evidence of awareness whereas the VS state does not . MCS patients also can present purposeful movements (relevant to th e context) and simplified gestural or verbal communication to cognitive task (Bernat, 2006; McQuillen, 1991) . A Vegetative State is defined as Persistent Vegetative State (PVS) w hen the patient remains in the VS condition more than 1 month after acute traumatic or non traumatic brain damage. The patient can still revert from that state to a less chronic state. PVS can lead to Permanent Vegetative State which is considered as an ir reversible state. A VS is defined as permanent in non traumatic cases after 3 month, in traumatic cases after 12 months (The Multi Society Task Force on Persistent Vegetative State, 1994a, 1994b) . Guidelines for such cases are less reliable with non traumatic origins, in cases when it is hard to predict a possible recovery (Menon et al., 1998) . The inability of co mmunication in most of these medical conditions is also cause of ethical concerns for treatment decisions that may include removal of life support (Bryan Jennett, 2002) .

PAGE 26

26 Another permanent stage that need to be mentioned is Brain Death, which can be th e final phase of the DOC conditions. Its clinical assessment is commonly based on the evaluation of the loss of all the brainstem reflexes and continuous loss of autonomous functions, such as breathing (Medical Consultants on the Diagnosis of Death, 1981) . Brain Death is usually considered equal to the death of the individual, although there are differences in the diagnosis criteria in each country. Studies are cur rently conducted to make such classification and diagnosis more uniform (Haupt & Rudolf, 1999; M. Smith, 2012) . Bedside Assessment In the diagnosis of DOC cases, one n eeds to carefully assess any evidence of awareness and consciousness to avoid any misclassification, as treatment for different classification will differ substantially. Different standardized scales are recommended in clinical practice (American Congress of Rehabilitation Medicine et al., 2010) , such as the largely used Glasgow Coma Scale and Coma Recovery Scale Revised (CRS R), where the latter is cons idered to be more robust and discriminative. All these methods require systematic examinations to discriminate responses to stimuli between voluntary or contextual and pure reflexes. Repeated assessment is important given the swift changes in arousal level in each DOC subject. This is also important to decrease the high rate of misdiagnosis (averagely 40%) caused by the complex condition and lack of clear models (Andrews et al., 1996; Childs et al., 1993) . Classical Conditioning Classica l Conditioning is a learning mechanism elicited through association between a conditioned stimulus (CS) and an unconditioned stimulus (US). It was discovered by the Russian physiologist Ivan Pavlov (Pavlov & Anrep, 2003) leading to

PAGE 27

27 its great influence in the field of psychology, specifically known by the name behaviorism. Behaviorism supp oses that learning occurs through interaction with the environment and consequently the latter influences the behavior. In classical conditioning, the US is a stimulus that evokes automatically a response, defined as Unconditioned Response (UR), while the CS is a stimulus that is considered neutral. The CS does not evoke a response before the pairing. After being associated with an US, it will trigger a Conditioned Response (CR) if the conditioning took place in the subject. Obtaining CR is the aim of the methodology. CR is a learned response that will be unlearned with time or other techniques. A classic example is that of the Pavlovian dog which was administered food as US and light flashes was used as CS, bringing the dog to associate the UR (salivation in presence of food) with the presentation of a light stimulus (Medin et al., 2004) . Different procedures were implemented to reflect different protocol designs s uch as Forward Conditioning, where CS precedes US and usually is considered faster, and Backward Conditioning, where US precedes CS (Chang et al., 2004) . In this work it was used Forward Conditioning, specifically Trace Conditioning, where the US is presented after the CS ending such that they do not overlap. A Classical Conditioning procedure can be divided into three phases. The Habituation Phase is when CS and US are presented unpaired to avoid eliciting any shift of attention to a particular stimulus caused by the novelty of the sound in the subject environment, leading to diminished response over time ( Figure 2 3 A, B). Acquisition Phase is when the learning takes root in the subject over time (CR acquired) coming to evoke the Conditioned Response ( Figure 2 3 C). The procedure commonly

PAGE 28

28 consist in repeated pairings of CS and US until the CS evokes the CR. Presenting CS without US will gradually elicit CR to a lesser extent. This phase is defined as Extinction, underlying the gradual loss of CR. Over time with an unpaired CS, the subject will revert to a pre learning behavior ( Figure 2 3 D ) .

PAGE 29

29 Figure 2 1 . Linearly separable data, binary case. Figure 2 2 . Hierarchy of the different states in Disorders of Consciousness. The Minimally Conscious State have higher probability of recovery, com pared to Vegetative State and Permanent Vegetative State.

PAGE 30

30 Figure 2 3 . The 4 phases of Classical Conditioning in the Pavlov ian dog experiment . B efore the conditioning (habituation phase) , A) an uncond itioned stimulus (the food) is presented , B) an conditioned stimulus (the ringing bell) is presented. C) during conditioning (acquisition) both conditioned and unconditioned stimulus are presented. D) after conditioning (extinction) only the co nditioned st imulus is presented that should elicit the same arousal as the unconditioned stimulus.

PAGE 31

31 CHAPTER 3 MATERIALS AND METHODS Overview This chapter describes the methods applied for the analysis and implementation of pattern classifica tion (explained in CHAPTER 6 ) of the fMRI datasets (described in CHAPTER 4 and CHAPTER 5 ). CHAPTER 4 will fo cus on assessing the feasibility of an fMRI BCI based on emotional cues. CHAPTER 5 will focus on hierarchical studies to assess Disorders of Consciousness. CHAPTER 6 will provide det ails of the implementation of the MANAS 4 toolbox used in all the analyses. CHAPTER 7 will assess the performance of our approach on a pre existing and validated fMRI dataset of overt motor execution . Preprocessing of fMRI D ata sets An important step to consider before statistical analysis of fMRI images is the preprocessing of the raw data composed of raw volumes of fMRI data. The preprocessing in this work was performed using SPM8 (Wellcome Department of Imaging Neuroscience, London) a nd MANAS 4 toolbox. The preprocessing pipeline of the MANAS 4 toolbox is composed of standard procedures of fMRI signal preprocessing, namely, Realignment, Slice Timing correction, Skull stripping, Coregistration, Segmentation of brain tissue, Normalizatio n on a Talairich Brain (Talairach & Tournoux, 1988) or a Montreal Neurological Institute (MNI) brain template (Evans, Collins, & Milner, 1992; Evans et al., 1993) . There are 2 main causes for a general temporal trend in an fMRI task, subject BOLD dependent and machine dependent. The keyword detrending represents the set of methods applied to compensate for these effects. Although temporal filtering is

PAGE 32

32 advised, there is discrepancy in the literature regarding which filter ing method would be the best to not alter irreversibly the information content, considering that the data has usually a low temporal resolution (TR ~2.38 3.14). A first evaluation of the effect of an underlying trend in the signal can be performed by the l ooking at the shape of the autocorrelation function. If the fMRI time series has an underlying linear trend, it will be easily seen with a visual inspection of the autocorrelation plot ( Figure 3 1 , Figure 3 2 ). T here are mainly 2 methods for detrending the fMRI signal: estimation of the trend with polynomial fitting (used in this work) or high pass filtering (above 0.01 Hz) (Friman et al., 2004) . In t he case of polynomial fitting , used in this work ( Figure 3 3 ), often the detrending step is reduced to the removal of the mean or a linear estimation of the signal, not being able to estimate properly the distortions introduced by non linear detrending. In this work time series outliers were reduced in amplitude using a moving average approach , considering a predefined number of nearby points . This approach was selected to remove temporal distortion in the BOLD signal due to movement a rtifact that is usually not completely removed and to avoid decreased sensitivity in z normalization. Z normalization is considered a necessary step in machine learning analysis to avoid dimensional bias, caused by non physiologically relevant variation in the BOLD amplitude, such as local blood flow variations and signal amplitude differences due to the distance from the receive r coil of the MRI . It is described by: ( 3 1 )

PAGE 33

33 where is the normalized signal, is the raw signal, and are respectively the mean and the standard deviation of the raw signal. Feature Selection Algorithms Machine l earning approaches, in general, and SVM, in particular, suffer from the (Morgenthaler, 1962) . However, SVM is usually quite robust against this problem. Although, parameter (Mahmoudi, Brovelli, 2012) it is always preferable to reduce the number of features/voxels while increasing the number of folds in the cross validation step. Feature Selection algorithms are used to reduce the dimensionality, but their use has to be compliant with the experiment requirements. A common need in Neuroscience experiments is the possibility to remap the brain activations on the brain space (patient or normalized brain). Therefore, any transformation of the feature set in another space has to take into consideration the availability of the inverse procedure to correctly remap the extracted features. In this work different approaches for feature selection algorithms were evaluated, n amely, Fisher Scoring ( Gu et al., 2012b) , Fisher Scoring with Searchlight (Pereira et al., 2009) and Effect Mapping (Lee et al., 2010) , to assess their robustness and possibility to be used in other fMRI analys is. F isher S coring Fisher Scoring is a univariate method which aims to find features that have the maximum distance between samples of different classes ( but minimum distance between samples of the same class ( (R.O. D uda, P.E. Hart, & D.G.

PAGE 34

34 Stork, 2007) . Being a univariate approach, each voxel is considered as an individual channel. The Fisher score is defined as (for a single channel in a binary case ) : ( 3 2 ) F isher Scoring with Sear chlight (FSSL) Fisher Scoring with Searchlight is a method that extends Fisher Scoring, partially avoiding its intrinsic univariate characteristic. It was developed specifically for fMRI analysis (Kriegeskorte & Bandettini, 2007; Etzel, Zacks, & Braver, 2013) , with a multivariate approach that measures information content in a small subset of the 3D volume (typically with a spherical kernel) using classical Fisher Scoring. Its common implementat ion is as a wrapper for Fisher Scoring by considering the sum of nearby Fisher scores within a spatial kernel. Effect Mapping Developed in the laboratory of Dr. Sitaram (Lee et al., 2010 , 2011; Sitaram et al 2010 ) , it is a method to interpret in a more consistent manner the output of a trained SVM classifier . Effect value in each voxel is obtained by multiplying the SVM weight value in the voxel with the Mutual Information (MI) between the BOLD time series in the voxel and the overall output o f the trained SVM, as described in Eq. ( 3 3 ) :

PAGE 35

35 ( 3 3 ) : the voxel, (with number of voxels) where is the Effect Map value for th e voxel , is the weight vector estimated with SVM training, is the normalized MI.

PAGE 36

36 Figure 3 1 . Example of a utocorrelation of the averaged BOLD time series before detrending without z normalization . Figure 3 2 . Example of autocorrelation of the averaged BOLD time series after detrending without z normalization .

PAGE 37

37 Figure 3 3 . Time series of averaged BOLD before detrending (blue) and after detrending (red) without applying z normalization.The trend is superimposed on the raw time series data (black ) .

PAGE 38

38 CHAPTER 4 CONDITIONING WITH EMOTIONAL SOUNDS: T OWARDS AN FMRI BCI Overview As a first step to enable basic communication through a brain computer interface (BCI) in cognitively impaired patients , the possibility to condition two distinct brain responses, namely affirmative and negative, using a semanti c classical conditioning paradigm in an fMRI setting was here investigated . Subjects were presented with congruent and incongruent word pairs as CS , which were paired with an emotionally pleasant (baby laughter) or unpleasant (scream) sound as US , in order to elicit emotional CR. For the feature selection two different approaches were used: Fisher Scoring and Effect Mapping. In all the three different phases (early acquisition, late acquisition, extinction) higher activations in the Insula , Amygdala and Hip pocampus were found. D ifference in affirmative and negative brain responses was detected as an effect of conditioning for the incongruent word pairs, opening the possibility for using this paradigm in a binary choice BCI for communication in patients of br ain disorders . Methods Part i cipants and Acquisition Ten right handed, native German speaking, healthy individuals ( 6 males, 6 females), ranging in age from 21 to 28 (mean age = 25.3, SD = 1.77 years), participated in this study. All participants gave writt en informed consent prior to participation in the fMRI experiment. The study was approved by the Ethics Committee of the Medical Faculty of the University of Tübingen and was performed in compliance with the Code of Ethics of the World Medical Association (Declaration of Helsinki).

PAGE 39

39 The experiment was performed using a Siemens AG (Erlangen, Germany) 3T Trio MRI scanner. Functional T2* weighted images were acquired with a standard 12 channels head coil, in transversal orientation (TR = 1.5 s, TE = 30 ms, flip angle = 70°, matrix 64 x 64, voxel size = 3.3 x 3. 3 x 5.0 mm 3 , 16 slices ) covering the whole brain. The first 10 volumes of every b lock were discarded to permit equilibrium. Stimuli The stimuli consisted of 300 German word pairs, half of them congruent (e .g. pair was formed by a superordinate category and a subordinate object, and congruence was given by the belonging of the object to the category. The ca tegories that were used were Animals, Countries, Fruit, Furniture, Sports, Clothing, Instruments, Drinks, Crockery, and Jobs. The word pairs were recorded from a native German speaker. The duration of each word pair was 1.5 s. The negative and affirmative responses to the word pairs constituted the conditioned stimuli (CS). The unconditioned stimuli (US) were two standardized emotional sounds drawn from the Internatio nal Affective Digitized Sounds (Bradley & Lang, 2007) : a pleasant emotional stimulus (a baby laughter) and an unpleasant emotional stimulus (a scream). The duration of each US was also 1.5 s. Participants heard all auditory stimuli through MRI compatible headphones with efficient gr adient noise suppression (up to 45 dB) and a filter system with more than 90 dB RF suppression (MR Confon Audio System, Leibniz Institute for Neurobiology at Magdeburg, Germany). Experimental P aradigm The paradigm consisted of a single session divided into six blocks, as shown in Figure 4 1 . In the first block, defined as habituation phase, 25 incongruent word pairs

PAGE 40

40 (CS1), 25 congruent word pairs (CS2), 25 unpleasant emotional stimuli (scream, US1) and 25 pleasant emotional stimuli (lau ghter, US2) were presented to the subject in a random order. The inter stimulus interval (ISI) was also randomized to optimize design efficiency, and could last for 6, 7.5, or 9 s. The habituation phase served to evaluate the activations relative to each type of stimulus individually, before their association in the conditioning process. In the second and third blocks, which were identically structured and constituted the early acquisition phase, 25 incongruent word pairs (CS1) and 25 congruent word pairs (CS2) were randomly presented. Each word pair was immediately followed by an US: scream (US1) after incongruent word pairs and laughter (US2) after congruent word pairs. Also in the fourth and fifth block (late acquisition phase), 25 CS1 and 25 CS2 were pre sented. In the fourth block, only 10 CS1 and 10 CS2 were paired (40% of the word pairs were followed by the appropriate US) with US1 and US2, respectively. In the fifth block, only 5 CS1 and 5 CS2 were paired (20%). In the sixth and final block (extinction block), none of the 50 word pairs were followed by an US. The gradual diminution of the percentage of word pairs followed by emotional stimulation served to verify whether the conditioning had taken place, meaning that congruent and incongruent word pairs could be discriminated thanks to their association to the emotional sounds, even when the emotional sounds were not present anymore. Behavioral M easures At the end of the first block (habituation) and fifth block (end of acquisition) the subjects were ins tructed to use the Self Assessment Manikin (Bradley & Lang, 1994) to rate the valence (pleasantn ess/unpleasantness) and the arousal related to the two emotional US (scream and laughter . The aim was to evaluate the habituation rate

PAGE 41

41 throughout the measurement. The SAM comprises two 9 point scales, ranging from Analysis Data were preprocessed using Statistical Parametric Mapping (SPM8, Wellcome Department of Imaging Neuroscience, London, UK) and analyzed with MANAS 4 run on Matlab R2013b (Mathworks, Inc., Sherborn , MA, USA). Images of each subject were realigned and unwarped to correct for head movement, and were normalized to a standard Echo Planar Imaging (EPI) template in Montreal Neurological Institute (MNI) space. Spatial smoothing was applied using a Gaussian kernel with full width at half maximum of 9 mm. Before proceeding with statistical analyses , the data were detrended with a linear polynomial regressor. Noisy features (voxels) were excluded based on variance, with outlier removal. The data were then z nor malized with respect to time. The task s were grouped by type and were analyzed separately, (2nd and 3rd Blocks, 4th and 5th Blocks, 6 Block). The 3rd and the 4th image w ere included for all the sessions. A whole brain mask from the digital atlas of the Automatic Anatomical Labeling (AAL) (Tzourio Mazoyer et al., 2002) was used to discard voxels outside from an MNI brain. To investigate brain activations and select the most important features for the classification , two supervised methods for feature selection were used (separately) , namely, Fisher Scoring with Search Light (Kriegeskorte & Bandettini, 2007) and Effect Mapping (Lee et al. , 2010) . Both were executed with a 10 fold cross validation. The feature selection with FSSL wa s executed considering a 3x3x3 spherical kernel , choosing the 2000 most important features (highest score).

PAGE 42

42 The feature selection with Effect Mapping was executed by choosing t he most important features with a cutoff threshold of 20% of the norm alize d range of all the scoring. Threshold was chosen after an optimization step based on classification performance and visual inspection (noise in the functional mapping). Both feature selection methods resulted in feature map s that w ere quantitatively and qualitatively analyzed, by inspecting the most active and discriminating areas. The featu re maps were used as masks during classification, to include only the most important features (voxels) from the images in the dataset. Thereafter, classification was executed with a linear SVM, using LibSVM (C. C. Chang & Lin, 2011) , with a 10 fold cross validation . Each group was analyzed separately . Res ults Behavioral Data SAM R atings A two sample t test indicated that participants rated the scream as significantly more unpleasant (block 1: t(20) = 10.62, block 5: t(20) = 3.09, p<0.01) and more arousing (block 1: t(20) = 5.87, block 5: t(20) = 2.66, p<0.02) than the laughter, both at the beginning and the end of the measurement. The arousal associated with the scream was significantly smaller at the end of block 5 compared to block 1 (t(20) = 2.96, p<0.01), although no significant difference was found for the laughter (p=0.3). T he valence associated to the scream and laughter did not change significantly during the experiment (p=0.2 in both cases). One of the subject s had to be excluded for not responding correctly to the stimuli limiting the analysis of the data on 10 subjects.

PAGE 43

43 Neuroimaging C lassification The functional brain maps obtained with both feature selection steps showed the common activations. Feature Selection with Effect Mapping (block 2, 3 = 90%, block 4, 5 = 88%, block 6 = 97%) improve the overal l accuracy of the classification ( Figure 4 2 , Figure 4 3 ) compared to FSSL Mapping (block 2, 3 = 84%, block 4, 5 = 82%, block 6 = 89%). Th e number of voxels selected in each brain region by different feature select ion approach with the Effect Mapping method ( Table 4 1 ) was computed and averaged for all the 11 subjects and all the different stages (blocks) of the conditioning paradigm . In the Early Acquisition phase (block 2 and 3) , the comparison between congruent and incongruent word pairs (pairing 100%) showed differences in the functional brain map of the features ( Figure 4 4 ) in the following regions: Insula, known to be involved in emotional cues ; Superior Temporal Gyrus containing the auditory cortex, elicited by the auditory cue s; and Hippocampus, known to be involved in memory and emotional tasks. In the Late Acquisition phase (block 4 and 5) the comparison between congruent and incongruent word (pairing 40% and 20%) showed activations in the following regions ( Figure 4 5 ): Superior Temporal Gyrus , Insula , and Putamen, known to be involved in processing of learning and pain. In the Extinction phase (block 6) the comparison between congruent and incongruent word (no pairing) elicited activations in the followi ng regions ( Figure 4 6 ) : Insula , Anterior Cingulate , and Angular Gyrus, known to be involved in word processing . Discussion This experiment al paradigm assessed the feasibility of apply ing classical conditioning to two different responses (affirmative and negative) using positive and

PAGE 44

44 negative emotional US, in order to help in the discrimination of the two res ponses . The results show ed that aversive semantic classical conditioning took place, demonstrated by the activations found in the insula, known to be related to emotion processing (Phan, Wager, Taylor, & Liberzon, 2002; Phillips et al., 1998; Ruiz, Buyukturkoglu, Rana, Birbaumer, & Sitaram, 2014) , present throughout all the phases, even when emotional cues (US) were not presented anymore. T his confirm s th e feasibility of using the activation patterns in the insula and emotion related areas to interpret brain responses after a preliminary conditioning. The Angular Gyrus was also activated , showing that an evaluation of the meaning and the relation of the word pairs was executed during their presentation (Lacey, Stil la, & Sathian, 2012) . The Amygdala, Anterior Cingulate and Insular Cortex are crucial structures in the acquisition of aversive delay conditioning, independent of the kind of the conditioning paradigm (Sehlmeyer et al., 2009) . The Amygdala was not detected as it wa s excluded from the fMRI pulse sequence due to technical difficulties . The sequence was optimized for a future real time fMRI BCI implementatio n, forcing us to limit the volume dimension and spatial resolution for a faster acquisition. The Anterior Cingulate was better activated in the Extinction phase. The activations found in the P utamen, a region related to implicit learning (Packard & Knowlton, 2002) and category learnin g (Ell, Marchant, & Ivry, 2006) , is an additional indication that the learning took place .

PAGE 45

45 Table 4 1 . ROI analysis scores . Average of the included features selected in each AAL ROI. Brain Region Early Acquisition Late Acquisition Extinction Anterior Cingulate 9.8% 6.3% 4.8% Insula 8.6% 7.8% 9.0% Hippocampus 6.8% 5.9% 6.4%

PAGE 46

46 Figure 4 1 . Block design p a radigm used in the c lassical c onditioni n g experiment . It i s composed of 6 blocks divided in 4 phases (habituation, early acquisition, late acquisition, extinction) . Figure 4 2 . Classification accuracies using E ffect Mapping method as feature selection. Accuracies are reported for each phase (early acquisition, late acquisition, extinction) in all the subjects.

PAGE 47

47 Figure 4 3 . Classification accuracies using Fisher Scoring with Searchlight as feature selection. Accuracies are reported for each phase (early acquisition, late acquisition, extinction) in all the subjects. Figure 4 4 . Functional Brain Map of the features selected with Effect Mapping in the early acquisition phase ( blocks 2 and 3 , 10 fold cross validation, accuracy=84 .04 %).

PAGE 48

48 Figure 4 5 . Functional Brain Map of the features selected with Effec t Mapping in the late acquisition phase ( blocks 4 and 5 , 10 fold cross validation, accuracy =82.93 %). Figure 4 6 . Functional Brain Map of the features selected with Effec t Mapping in the extinction phase ( block 6 , 10 fold cross validation, accuracy=89 .4 %).

PAGE 49

49 C HAPTER 5 FMRI PATTERN CLASSIFICATION TO ASSESS DISORDERS OF CONSCIOUSNESS Overview As it has been suggested by the work of Owen (Owen et al., 2005) , neuroimaging studies on DOC would benefit if they are conducted hierarchically . The current methodology used to diagnose the state of patients with Disorders of Consciousness (DOC) is mainly based on bedside assessment , which has been proved to be inaccurate in 40% of the cases (Andrews et al., 1996; Schnakers et al., 2009; Childs & Mercer, 1996) . Functional neuroimaging could facilitate improve ments in the accuracy of the diagnosis which is in turn important for increas ing the probability of recovery by com paring the brain activations of the patient with that of healthy individuals . This study aimed to investigate the possibility of using pattern classification to distinguish b etween DOC states (VS, PVS, MCS) with a battery of experiments , including, pain stimulation , working memory, emotions and intentional contro l . Methods Parti cipants and Acquisition Five right handed, native German speaking, healthy i ndividuals participated in this study as controls . Forty one right handed, native German speaking, DOC patients participated in the study. The patients were pre labeled based on bedside assessment diagnosis (19 as MCS , 16 as PVS, 6 as VS). The study was ap proved by the Ethics Committee of the Medical Faculty of the University of Tübingen and was performed in compliance with the Code of Ethics of the World Medical Association (Declaration of Helsinki).

PAGE 50

50 The experiment was performed in the German cities, Tübing en and Vogtareuth , using the facilities in both locations as it was possible to move the patients too far from their hospital for their medical condition . In Tübingen , the study used a Siemens AG (Erlangen, Germany) 3T Trio MRI scanner. Functional T2* weighted images were acquired with a standard 12 channels head coil, in transversal orientation (TR = 2.38 s, TE = 30 ms, flip angle = 70°, matrix 64 x 64, voxel size = 3.0 x 3.0 x 3 .0 mm 3 , 36 slices) covering the whole brain. The first 10 volumes of every b lock were discarded to permit equilibrium. In Vogtareuth , the study used a Siemens AG (Erlangen, Germany) 3T Symphony Tim MRI scanner. Functional T2* weighted images were acquired with a standard 12 channels head coil, in transversal orientation (TR = 3.41 s, TE = 50 ms, flip angle = 9 0°, matrix 64 x 64, voxel size = 3.0 x 3.0 x 3 .0 mm 3 , 36 slices) covering the whole brain. The first 6 volumes of every b lock were discarded to permit equilibrium . Experimental Paradigm s and Stimuli Three different experimental paradigm s were conducted hierarchically (Owen et al., 2005) . Considering the difficulties in the recording of DOC patients, each paradigm ha d only one session. DOC patients need to be monitored and connected to their life support systems, limiting the duration off the protocol to avoid any repercussion on the health of the patients. The life support systems caused artifacts in the recording of both T1 structural images and T2 EPI functional images. The experimental paradigms wer e, in order of execution, Pain Perception task, Empathy for P ain task and Mental Imagery task .

PAGE 51

51 Pain Perception Task The stimuli consisted of Electric painful stimuli presented to the controls and patients in a block design. There were two blocks, the bas eline block where the patient had to rest and no stimulus was presented, and stimulation block w h ere 60 painful stimuli were presented. The duration of each block was approximately 60 s and there was an interval of 5 s between them . Each s timulus , in the s timulation block, last ed 2 ms and the interstimulus interval (ISI) wa s 1 s ( Figure 5 1 ). A one second delay was also added at the beginning of the e xperiment to avoid conflicts in the presentation software . Empathy for Pain Task The stimuli consisting of 40 emotional sounds (Bradley & Lang, 2007) , 20 painful, 20 neutral were presented to the controls and patients in a block design. The sound samples showing pain of other s are not considered painful stimuli. There were two block types. The emotional block (painful sound) and Neutral block (neutral sound) included 4 emotional/neutral sounds lasting 6 s each with an interstimulus interval of 1.5 s, 5 blocks for each condit ion for a total of 10 blocks . The duration of each block was approximately 33 s and there was an interval of 15 s between them ( Figure 5 2 ) . Participants heard all the emotional sounds through MRI compatible headphones with efficient gradient noise suppression (up to 45 dB) and a filter system with more than 90 dB RF suppression (MR Confon Audio System, Leibniz Institute for Neurobiology at Magdeburg, Germany) . Mental Imagery Task The mental imagery task ( Figure 5 3 ) was organized in a randomized design paradigm. At the beginning of the paradigm standard summarized auditory instruction s

PAGE 52

52 (in German) to perform a specific task (instruction long) were presented with a duration correlated with the length of the instruction. After presenting the standard instructions, short and concise auditory instructions were presented (instruction short) with a fixed duration of 2 s for 5 times each. Each instruction was presented for 6 times (long and short), for a total of 18 block s , with a 30 s of interval between each block. Participants heard a ll the auditory instructions through MRI compatible headphones with efficient gradient noise suppression (up to 45 dB) and a filter system with more than 90 dB RF suppression (MR Confon Audio System, Leibniz Institute for Neurobiology at Magdeburg, Germany ). Analysis Data were preprocessed using Statistical Parametric Mapping (SPM8, Wellcome Department of Imaging Neuros cience, London, UK) and analyzed with MANAS 4 run on Matlab R2013b (Mathworks, Inc., Sherborn, MA, USA). The T1 anatomical images of each subject were segmented with SPM8 after skull strip with BET in the FSL package (Jenkinson, Beckmann, Behrens, Woolrich, & Smith, 2012; S. M. Smith et al., 2004) , obtaining the CSF, White Matter and Gray Matter masks. Skull stripping was performed with default parameters as recommended by the authors, and the p erformance of this step was evaluated with visual inspection. This step was executed to improve the segmentation performances, deleting any external artifact that would have caused mismatch in the estimation of the boundaries. Segmentation in SPM8 is based on a priori spatial distribution of brain tissues. Images of each subject were realigned to correct for head movement using SPM8 . Images and masks of each subject were normalized to a standard T1 template

PAGE 53

53 in Montreal Neurological Institute (MNI) space , co regi stering them with the T1 anatomical image of each subject after the skull strip step . Spatial smoothing was applied using a Gaussian kernel with full width at half maximum of 6 mm. All the tasks were processed with MANAS 4 toolbox. T he data were detrended with a linear polynomial regressor and n oisy features (voxels) were excluded based on variance, with outlier removal. T he data were then z normalized with respect to time. The tasks were analyzed separately, and all were masked with a Gray Matter mask. White Matter was excluded from the analysis considering that it is commonly damaged in traumatic injuries with disruption in fibers. Considering a subset of the brain, especially in traumatic brain injuries, helps the classification of the functional data, decr easing the number of features and excluding the voxels that are known to be less significant , containing known damaged tissues. One of the challenge s of the analysis with traumatic patients is the consistent alteration of the brain shape . T herefore th e segmentation and normalization processes need to be executed using the patients anatomical images . Considering the limited amount of samples in the f MRI sessions, and to select the best features for the classification a feature se lecti on algorithm was executed . The feature selection approach of Fisher Scoring with Searchlight (Kriegeskorte & Bandettini, 2007) , a supervised multivariate method considered more robust in cases with low number of instances , was used with 10 fold cross validation during classification training . Fisher Scoring with Searchlight was execu ted considering a 3x3x3 spherical kernel, choosing the 1 000 most important features (highest score). The number of

PAGE 54

54 voxels to consider was chosen after an optimization step based on classification performance and visual ins pection (noise in the functional mapping). The feature maps , obtained after feature selection , were quantitatively and qualitatively analyzed, by inspecting the most active and included areas. The feature maps were used as masks during classification, to i nclude only the interesting features (voxels) from the images in the dataset. Different sample selection (number and position of image volumes) were applied to each task, taking also into account the different TR s for different fMRI recording of th e experiments . Volumes at the beginning of the block were discarded to allow the BOLD signal to reach the peak amplitude. In the Pain Perception Task , the scans in the first 8 seconds of each block were discarded keeping all the rest . With a TR of 3.41 a t otal of 14 volumes per block were selected, discarding the 1 st and 2 nd image. With a TR of 2.38 a total of 22 volumes per block were selected, discarding the 1st, 2nd and 3rd image. In the Empathy for Pain Task the scans in the first 6 seconds of each bloc k were discarded keeping all the rest . With a TR of 3.41 a total of 7 volumes per block were selected, discarding the 1st image . With a TR of 2.38 a total of 10 volumes per block were selected, discarding t he 1st and 2nd image. In the Mental Imagery Task t he scans in the first 7 seconds of each block were discarded keeping all the rest. With a TR of 3.41 a total of 8 volumes per block were selected, discarding the 1 st and 2 nd image. With a TR of 2.38 a total of 11 volumes per block were selected, discarding t he 1st, 2nd image.

PAGE 55

55 The c lassification training was executed with a multivariate linear SVM, using LibSVM (C. C. Chang & Lin, 2011) , with a 10 fold cross validation. In the case of the Mental Imaginery Task , each condition was analyzed in pairs. All the analys e s were conducted with binary classification. Classification accuracies were compared between the different classes of DOC patients, to evaluate if there were any significant difference in the performances. Furthermore the functional brain maps obtained from the feature selection step were compared between patients and controls . Results Analysis on DOC patients proved to be challenging because of artifacts and traumatic injuries, which caused deep alteration s in the brain structures. Brain Segmentation with FSL and SPM failed in some cases where the subject brain was too damaged to apply any of the current algorithm s for tissue segmentation ( Figure 5 4 A). Ventricle enlargement is a typical consequence of a traumatic injury (Poca et al., 2005) and cause the compression against the skull causing a considerable shift of brain tissues and decreased performance in segmentation. Segmentation based on probabilistic distribution produced sufficient ly good performance with structurally altered normalized brains after skull stripping ( Figure 5 4 D, Figur e 5 5 ). Overall, the accuracies were lower in DOC patients compared to healthy patients ( Figure 5 6 , Figure 5 7 , Figure 5 8 ) in eac h t ask. In the Pain Perception Task , the stimulation block elicited activations in the Putamen ( Figure 5 9 ) , known to be involved in the processing of pain perceptions and somatosensory stimuli, and Inferior Temporal Gyrus (only in he althy controls as shown in Figure 5 10 ).

PAGE 56

56 In the Empathy for Pain Task , the negative emotional auditory cue (screaming) elicited activations ( Figure 5 11 , Figure 5 12 ) in the Putamen, know n to be involved in the processing of pain perceptions and lear ning, Middle Temporal Gyrus, and mostly only in healthy controls the Parahippocampal Gyrus, Hippocampus and Amygdala, known to be all involved in emotional processing. In the Mental Imagery Tas k , only the case of Instruction B ( rooms The comparison between the two conditions showed activations ( Figure 5 13 , Figure 5 14 , Figure 5 15 , Figure 5 16 ) in the Postcentral Gyrus and Precentral Gyrus in healthy controls and also in a subset of MCS patients. Activations in Cuneus and Cerebellum were also found. A PVS patient presented activ ations in the Lingual Gyrus and Middle Occupital Gyrus, showing low classification performances, while another PVS presented performances and activations comparable to a normal subject ( Figure 5 15 ). Discussion The results showed that it is possible to discriminat e between patients of different DOC categories. Differences in the classification performances and in the activation pattern between different state s of DOC and healthy subjects were s hown ( Figure 5 6 , Figure 5 7 , Figure 5 8 ) . The results in the Pain Perception Task showed that the patients reacted to the painful stimuli, eliciting activations in the Putamen, known to be involved in pain related motor responses (Starr et al., 2011) . Analysis of the Empathy for Pain Task sh owed that negative emotional auditory cues elicited areas known to be related to emotional stimuli, as Putamen, Parahippocampal Gyrus, Hippocampus and Amygdala.

PAGE 57

57 Hippocampus and Amygdala are usually involved in fear conditioning (Zelikowsky, Hersman, Chawla, Barnes, & Fanselow, 2014) and therefore habituation could be taking place. The binary classification of Instruct ion B and C in the Mental Imag ery Task el icited expected activations in the Primary Motor Cortex and Primary Somatosensory Cortex. The Posterior Lobe of the Cerebellum has been known to be involved in sensory experiences (k inesthetic sensations ) lacking sensory feedback from the environment such as motor imagery (Naito et al., 2002) . Its presence represents that a complex coordinated system between motor areas is sti ll mostly intact. The Cuneus is involved in motor imagery of coordinated exercises, and it should be more active if the task to be imagined involved more than one person (Mochizuki, Sudo, Kirino, & Itoh, 2014) prevent to consider interactions in the imagined environment. We do not have any mean to know which situation the PVS patient had imagined .

PAGE 58

58 Figure 5 1 . Pain perception paradi gm . It consists of block s of electric painful stimuli (6 0 stimuli; ISI=1s; 20ms duration) alternate d with rest (60s). Between each block there is a 5s interval.

PAGE 59

59 Figure 5 2 . Empathy for Pain paradigm . It consists of blocks of emotional sounds (International Affective Digitized Sounds; 5 blocks of 4 emotional/neutr al sound; 10 blocks in total; ISI 1.5s; 6 s durati on) alternated with rest (30s).

PAGE 60

60 Figure 5 3 . Mental Imagery paradigm . It consists of randomized auditory instructions to perform specific motor nd the instruction is presente d (duration vary) for each task. Then short instruction are presented every 2 s (short instruction) for 5 times for each task . Between each block there is a 30 s interval.

PAGE 61

61 Figure 5 4 . Series of DOC p atient images. A) B) Anatomical T1 Images. A) presents clear ventricle enlargement . C) is a functional image during a task of the same patient in B). D) shows the brain stripping segmentation superimposed on a T1 image (white border around brain).

PAGE 62

62 Figur e 5 5 . Series of Anatomical T1 Images. In each image is shown the brain stripping segmentation superimposed as a white border around the brain. A) presents ventricle enlargement. B) presents clearly tr aumatic brain tissues in the lower left lobe. C) presents a typical noisy T1 images. D) is a healthy control.

PAGE 63

63 Figure 5 6 . Classification accuracies of Pain Perception Task, using Fisher Scoring with Searchlight for feature selection with 10 fold cross validation . Accuracies are reported for PVS, MCS, and healthy states. Figure 5 7 . Classification accuracies of Empathy for Pain Task, using Fisher Scoring with Searchlight for feature selection with 10 fold cross validation . Accuracies are reported for PVS, MCS, and healthy states. Figure 5 8 . Classification accuracies of Mental Imagery Task, u sing Fisher Scoring with Searchlight for feature selection with 10 fold cross validation . Accuracies are reported for PVS, MCS, and healthy states.

PAGE 64

64 Figure 5 9 . Functional brain map of the activations during the Pain Perception Task in a MCS patient . F eature selection executed with Fisher Scoring with Searchlight (10 fold cross validation, accuracy=84.11%). Figure 5 10 . Functional brain map of the activat ions during the Pain Perception Task in a healthy control . Feature selection executed with Fisher Scoring with Searchlight (10 fold cross validation, accuracy=79.58%).

PAGE 65

65 Figure 5 11 . Functional brain map of th e activations during the Empathy for Pain Task in a PVS patient. Feature selection executed with Fisher Scoring with Searchlight (10 fold cross validation, accuracy=82.64%). Figure 5 12 . Functional brain map of the activations during the Empathy for Pain Task in a healthy control . Feature selection executed with Fisher Scoring with Searchlight (10 fo ld cross validation, accuracy=90 .05 %).

PAGE 66

66 Figure 5 13 . Functional brain map of the activations during the Mental Imagery Task ( ) in a PVS patient. Feature selection executed with Fisher Scoring with Searchlight (10 fold cross validation, accuracy=56.6 0 %). Figure 5 14 . Functional brain map of the activations during the Mental Imagery Task ( ) in a MCS patient . Feature selection executed with Fisher Scoring with Searchlight (10 fold cross validation, accuracy=78.96 %).

PAGE 67

67 Figure 5 15 . Functional brain map of the activations during the Mental Imagery Task ( ) in a PVS patient . Feature selection executed with Fisher Scoring with Searchlight (10 fold cross validation, accuracy=85.04 %). This patient surprisingly demonstrated high performances giv en his diagnosis .

PAGE 68

68 Figure 5 16 . Functional brain map of the activations during the Mental Imagery Task ( ) in a healthy subject. Feature selection executed with Fisher Scoring with Searchlight (10 fold cross validation, accuracy=88 . 00 %).

PAGE 69

69 CHAPTER 6 MANAS 4 TOOLBOX Overview ind f MRI studies to streamline the analysis of fMRI data with multivariate classification based on Support Vector Machines (SVM). It is partially based on the toolbox MANAS 3 (Rana et al., 2013). It allows to batch pr oce ss all the steps necessary in fMRI analysis, from preprocessing of raw images to the classification. Different algorithms and options w ere implemented in the toolbox and hence, i ts robustness had to be tested ( CHAPTER 7 ) . It is fu lly compliant and integrated with MATLAB (Mathworks, Inc., Sherborn, MA, USA). It provides s upport from the version MATLAB 2008b . A GUI for each section of the toolbox is under active development. Features of the Toolbox The analysis pipeline in the toolbo x is composed by the following steps: Preprocessing : detrending: linear, non linear spatial filtering masking z normalization Cross validation (random picking) : k fold leave one out k fold with the same number of labels across folds k fold with the same nu mber of labels across folds and in the fold

PAGE 70

70 Feature Selection : Fisher Scoring Fisher Scoring with Searchlight (with leave one out inside the fold) Effect Mapping Classification : SVM All the steps are implemented as modular packages, therefore the method s used in the toolbox can be used as stand alone functions (satisfying the dependencies). The preprocessing steps before the analysis (realignment, slice timing correction, normalization) are executed providing a wrapper for batch execution of SPM 8 and FSL metho ds. Tools to convert DICOM images in bulk and to read commonly used DICOM parameters such as TR are also provided, giving a platform independent interface. Each step is also cached after execution , allowing the re execution of any step without the need to be recomputed. After each run of the toolbox, different logging .mat files will be generated, containing the parameters and the results of the feature selection, ROI analysis and classification. Optionally is possible to save all the results of feature select ion and classification as function al brain maps in NIFTI format. It is also possible to inspect the results with a Result Manager GUI ( Figure 6 1 ) and evaluate the BOLD activations with a Slice Viewer ( Figure 6 2 ) built in in the toolbox. How to U se This too l box run s ( Figure A 1 , Figure A 2 ). These files contain a JSON formatted configuration variable

PAGE 71

71 that will b e passed to the desired method (e.g. manas_exec()). This approach allow for a high degree of customization and an easy configuratio n of different analysis.

PAGE 72

72 Figure 6 1 . MANAS 4 Results Manager. It is used to inspect the results and the performance of the feature selection and the classification in the analysis. Figure 6 2 . MANAS 4 Slice Viewer. It is used to inspect the activation found in the analysis.

PAGE 73

73 CHAPTER 7 TESTING MANAS 4 TOOLBOX WITH A MOTOR TASK DATASET Overview The assessment of the MANAS 4 toolbox performances required a robust and well known dataset. Different aspect s of the toolbox were inspected, such a s t he ease of us e , performances of different preprocessing steps (Tanabe et al., 2002) and differe nt feature selection algorithms: Fisher Scoring (Gu, Li, & Han, 2012b), Fisher Scoring with Searchlight , Effect Mapping (Lee et al., 2010). A motor t ask fMRI dataset (Rana et al., 2013) was used, which was already tested with the previous toolbox implementation (MANAS 3). In this dataset the participants had to perform right and left hand movements , known to elicit localized activations in the Motor and Somatosensory Cortex. Methods Parti cipants and Acquisition The dataset contains the data from nine right handed healthy college students (age: 26.4 ± 5.2). The experiment was performed using a Siemens Magnetom (Erlangen, Germany) 3T Trio MRI scanner. Functional T2* weighted images were acquired with a standard 12 channels head coil, in axial orientation (TR = q s, TE = 30 ms, flip angle = 7 8 °, matrix 64 x 64, voxel si ze = 3.0 x 3. 0 x 3.75 mm 3 , 32 slices ) covering the whole brain. The first 4 volumes of every session were discarded to permit equilibrium. Experimental Paradigms and Stimuli The participants were instructed to perform a hand movement, moving fingers and palm freely, presenting the stimuli in a block design. There were three blocks: 6 left -

PAGE 74

74 hand movement, 6 right hand movement, resting, each of 30s of duration. Each subject was recorded for two session. Analysis The volumes of each subject were realigned and unwarped to correct for h ead movement, and normalized to a standard Echo Planar Imaging (EPI) template in Montreal Neurological Institute (MNI) using Statistical Parametric Mapping (SPM8, Wellcome Department of Imaging Neuroscience, London, UK). The analysis were conducted with M ANAS 4 run on Matlab R2013b (Mathworks, Inc., Sherborn, MA, USA). Images of each subject were realigned and unwarped to correct for head movement, and normalized to a standard Echo Planar Imaging (EPI) template in Montreal Neurological Institute (MNI) spac e. Spatial smoothing was applied using a Gaussian kernel with full width at half maximum of 9 mm. After an optimization step based on classification performance, no detrending was applied . The data was then z normalized with respect to time. Classification was executed with multivariate SVM considering three different feature selection algorithm s with 10 fold cross validation : Fisher Scoring, Fisher Scoring with Searchlight, Effect Mapping. Functional brain maps were saved for visual inspection of the activati ons. Results The three different feature sele ction methods le d to approx imately the same accuracy (FS=91%, FSSL=89%, EM=91%) using multivariate classification ( Figure 7 1 ) . Using Fisher Scoring there was prominence in the selection of the right Primary Motor Cortex and Somatosensory Cortex over the left side ( Figure 7 2 A) . M otion

PAGE 75

75 a rtifacts ( Figure 7 2 B ) were also selected as features (when the patient moves at the time of the cue). Using FSSL , the left Primary Motor Cortex and Somatosensory Cortex was selected ( Figure 7 3 A) . Instead the right s ide did not show activations . Motion a rtifacts ( Figure 7 3 B) were also selected as features (when the patient m oves at the time of the cue). They were less visible on the left side. Using Effect Ma p ping features in Primary Motor Cortex and Somatosensory Cortex in both hemispheres were selected ( Figure 7 4 ) . Motion Artifacts seemed to not be se lected. Discussion The performance of the feature classification were considered satisfactory with all the algorithms and no substantial difference was found in the accuracy level. Considering the functional brain mapping of the features selected in all th e cases, Effect Mapping seems to be more selective for both conditions (right and left hand movement) compared to the other two methods. Moreover Effect Mapping seemed to be robust against the motion artifact found with the other two methods. The area sele cted included Postcentral Gyrus and Precentral Gyrus, respectively known as Primary Somatosensory Cortex and Primary Motor Cortex. Furthermore in some cases it was more elicited also Brodmann area 6, that includes part of the Pr e moto r Cortex and Supplement ary Area, crucial areas for motor planning and translation of cognition into movement (Nachev, Kennard, & Husain, 2008) . These areas were expected to be elicited thus confirming the robustness of the methods.

PAGE 76

76 Figure 7 1 . Averaged c lassification accuracies for each method of feature selection used in the analysis (FS, FSSL, EM). Figure 7 2 . Functional Brain M ap of the features selected with Fisher Scoring. A ) shows motor related activations, B ) shows motion artifacts that were selected as activations (10 fold cro ss validation, accuracy=90 . 5 6 %).

PAGE 77

77 Figure 7 3 . Functional Brain Map of the features sele cted with Fisher Sc oring with Searchlight. A ) shows motor related activations, B ) shows motion artifacts that were selected as activations (10 fold cross validation, accuracy=89 .17 %) . Figure 7 4 . Functional Brain Map of the features selected w ith Effect Mapping (10 fold cross validation, accuracy=90 .83 %).

PAGE 78

78 CHAPTER 8 CONCLUSIONS Overview All the studies conducted in this work were aimed to use multivariate pattern classification to discriminate between different brain states and the results showed that all the proposed methodologies can be the first step towards automatized brain state discrimination. Moreover the MANAS 4 Toolbox, used in all the analyses demonstrated good perfor mance. Its use facilitated the execution of different experiments simply varying its . Emotion Based fMRI BCI The results in CHAPTER 4 showed that double conditioning of brain responses is possible (negative and positive emotions and respectively congruent incongruent word pair). The insula activation th roughout all the phases of conditioning mean that classification of negative and affirmative responses with fMRI emotional cues is possible, and this can be considered the first step towards an fMRI BCI fo r cognitively impaired patients that lost the abili ty to communicate verbally (Liberati et al., 2012) . Analysis to Assess Disorders of Consciousness The aim of t he vegetative study was to evaluate the possibility of using multivariate pattern classification to discriminate between the differen t DOC ca t egories with a hierarchical paradigm . At present the classifier performance could be considered as a parameter to evaluate the cognitive status of the patient independent from normalization of the images (needing only realignment). It can be consid ered that less cognitive behavior

PAGE 79

79 there is, less performance in the classification will be found. It can still not be considered a sufficient parameter. Hence, there is a need to improve of segmentation algorithms, seeing the limitation of models bas ed on a priori probabilistic distributions. The use of functional brain maps can be of help in the diagnosis, aiming to decrease the misclassification of patients. The patient in Figure 5 15 was classified as PVS with a bedside asses sment, but in the Motor Imagery Task showed high performance in the classification and there were elicited areas typically present in a motor imagery task (Postcentral Gyrus, Precentral Gyrus, Cerebellum, Cuneus). Hence this patient could be an example of misclassification and proof of the necessity of neuroimaging system in DOC diagnosis. To increase the robustness of the methodology an increased number of healthy controls would be preferable. This would allow to obtain the most important features across t he subjects and evaluate if they can be found in the patients or , to evaluate the brain plas ticity. It has to be taken into account that lack of response in neuroimaging does not imply lack of awareness and conscience, but it has to b e considered that a patient could be in a non responsive state (sleeping) (Owen & Coleman, 2008) . Moreover, it has to be taken into account the increased complexity of the tasks in the paradigm. The aim is to take note of the overall performance in all the task. If good performances are shown in the task with higher complexity, the chance that the patient is consci ous is higher. This is of utmost importance to administer the correct treatment to each patient (Young & Wijdicks, 20 08) .

PAGE 80

80 APPENDIX MANAS 4 USAGE EXAMPLE This appendix expands the MANAS 4 TOOLBOX Execution of A nalysis Figure A 1 . Example of run script with common parameters for analysis.

PAGE 81

81 Execut ion of Preprocessing with SPM 8 Batch W rapper Figure A 2 . Example of run script with common parameters for preprocessing.

PAGE 82

82 LIST OF REFERENCES American Congress of Rehabilitati on Medicine, Brain Injury Interdisciplinary Special Interest Group, Disorders of Consciousness Task Force, Seel, R. T., Sherer, M., disorders of consciousness: evidence base d recommendations for clinical practice and research. Archives of Physical Medicine and Rehabilitation , 91 (12), 1795 1813. doi:10.1016/j.apmr.2010.07.218 Andrews, K., Murphy, L., Munday, R., & Littlewood, C. (1996). Misdiagnosis of the vegetative state: re trospective study in a rehabilitation unit. BMJ , 313 (7048), 13 16. doi:10.1136/bmj.313.7048.13 Bernat, J. L. (2006). Chronic disorders of consciousness. The Lancet , 367 (9517), 1181 1192. doi:10.1016/S0140 6736(06)68508 5 Boly, M., Coleman, M. R., Davis, M. H., Hamps hire, A., Bor, D., Moonen, G., Owen, A. M. (2007). When thoughts become action: an fMRI paradigm to study volitional brain activity in non communicative brain injured patients. NeuroImage , 36 (3), 979 992. doi:10.1016/j.neuroimage.2007.02.047 Br adley, M. M., & Lang, P. J. (1994). Measuring emotion: the Self Assessment Manikin and the Semantic Differential. Journal of Behavior Therapy and Experimental Psychiatry , 25 (1), 49 59. Bradley, M. M., & Lang, P. J. (2007). The International Affective Digit ized Sounds (2nd Edition; IADS 2): Affective ratings of sounds and instruction manual. Technical report B 3. University of Florida, Gainesville, FL. Chang, C. C., & Lin, C. J. (2011). LIBSVM: A Library for Support Vector Machines. ACM Trans. Intell. Syst. Technol. , 2 (3), 27:1 27:27. doi:10.1145/1961189.1961199 Chang, R. C., Stout, S., & Miller, R. R. (2004). Comparing excitatory backward and forward conditioning. The Quarterly Journal of Experimental Psychology. B, Comparative and Physiological Psychology , 57 (1), 1 23. doi:10.1080/02724990344000015 Childs, N. L., & Mercer, W. N. (1996). Misdiagnosis certainly occurs. BMJ , 313 (7062), 944. doi:10.1136/bmj.313.7062.944 Childs, N. L., Mercer, W. N., & Childs, H. W. (1993). Accuracy of diagnosis of persistent veg etative state. Neurology , 43 (8), 1465 1467. Chrabaszcz, J., & Dougherty, M. (2012). Deliberations on unconscious thought theory. Frontiers in Psychology , 350. doi:10.3389/fpsyg.2012.00350 Claxton, G. (2000). reases When You Think Less . {Harper Perennial}. Retrieved from

PAGE 83

83 http://www.amazon.de/exec/obidos/redirect?tag=citeulike01 21&path=ASIN/0060955414 Cohen, M. S. (2001). Real time functional magnetic resonance imaging. Methods (San Diego, Calif.) , 25 (2), 201 2 20. doi:10.1006/meth.2001.1235 Cortes, C., & Vapnik, V. (1995). Support Vector Networks. Machine Learning , 20 (3), 273 297. doi:10.1023/A:1022627411411 Dijksterhuis, A., & Nordgren, L. F. (2006). A Theory of Unconscious Thought. Perspectives on Psychologica l Science , 1 (2), 95 109. doi:10.1111/j.1745 6916.2006.00007.x Ell, S. W., Marchant, N. L., & Ivry, R. B. (2006). Focal putamen lesions impair learning in rule based, but not information integration categorization tasks. Neuropsychologia , 44 (10), 1737 1751. doi:10.1016/j.neuropsychologia.2006.03.018 Etzel, J. A., Zacks, J. M., & Braver, T. S. (2013). Searchlight analysis: promise, pitfalls, and potential. NeuroImage , 78 , 261 269. doi:10.1016/j.neuroimage.2013.03.041 Evans, A., Collins, D., Mills, S. R., Brow n, E. D., Kelly, R. L., & Peters, T. M. (1993). 3D statistical neuroanatomical models from 305 MRI volumes. In Nuclear Science Symposium and M edical Imaging Conference, 1993. , 1993 IEEE Conference Record. (pp. 1813 1817 vol.3). doi:10.1109/NSSMIC.1993.3736 02 Evans, Collins, D., & Milner, B. (1992). An MRI based stereotactic atlas from 250 young normal subjects. Soc. Neurosci. Abstr. , 18 , 408 492. Friman, O., Borga, M., Lundberg, P., & Knutsson, H. (2004). Detection and detrending in fMRI data analysis. Neur oImage , 22 (2), 645 655. doi:10.1016/j.neuroimage.2004.01.033 Gu, Q., Li, Z., & Han, J. (2012a). Generalized fisher score for feature selection. arXiv Preprint arXiv:1202.3725 . Gu, Q., Li, Z., & Han, J. (2012b). Generalized Fisher Score for Feature Selectio n. arXiv:1202.3725 [cs, Stat] . Retrieved from http://arxiv.org/abs/1202.3725 Haupt, W. F., & Rudolf, J. (1999). European brain death codes: a comparison of national guidelines. Journal of Neurology , 246 (6), 432 437. Haynes, J. D., & Rees, G. (2006). Decodi ng mental states from brain activity in humans. Nature Reviews Neuroscience , 7 (7), 523 534. doi:10.1038/nrn1931 Haynes, J. D., Sakai, K., Rees, G., Gilbert, S., Frith, C., & Passingham, R. E. (2007). Reading hidden intentions in the human brain. Current Bi ology: CB , 17 (4), 323 328. doi:10.1016/j.cub.2006.11.072

PAGE 84

84 Hippocrates. (400AD). On the Sacred Disease . Jain, A. K., Duin, R. P. W., & Mao, J. (2000). Statistical pattern recognition: A review. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE IN LIGENCE , 22 (1), 4 37. Jenkinson, M., Beckmann, C. F., Behrens, T. E. J., Woolrich, M. W., & Smith, S. M. (2012). FSL. NeuroImage , 62 (2), 782 790. doi:10.1016/j.neuroimage.2011.09.015 Jennett, B. (2002). The Vegetative State: Medical Facts, Ethical and Legal Dilemmas (1 Jennett, B., & Plum, F. (1972). Persistent vegetative state after brain damage. A syndrome in search of a name. Lancet , 1 (7753), 734 737. Kamitani, Y., & Tong, F. (2005). Decoding the visual a nd subjective contents of the human brain. Nature Neuroscience , 8 (5), 679 685. doi:10.1038/nn1444 Kriegeskorte, N., & Bandettini, P. (2007). Analyzing for information, not activation, to exploit high resolution fMRI. NeuroImage , 38 (4), 649 662. doi:10.1016 /j.neuroimage.2007.02.022 Kwong, K. K., Belliveau, J. W., Chesler, D. A., Goldberg, I. E., Weisskoff, R. M., human brain activity during primary sensory stimulation. Proceedings of the National Academy of Sciences , 89 (12), 5675 5679. doi:10.1073/pnas.89.12.5675 Lacey, S., Stilla, R., & Sathian, K. (2012). Metaphorically feeling: Comprehending textural metaphors activates somatosensory cortex. Brain and Language , 120 (3), 416 421. doi :10.1016/j.bandl.2011.12.016 LaConte, S., Strother, S., Cherkassky, V., Anderson, J., & Hu, X. (2005). Support vector machines for temporal classification of block design fMRI data. NeuroImage , 26 (2), 317 329. doi:10.1016/j.neuroimage.2005.01.048 Lang, P., Bradley, M., & Cuthbert, B. (2008). International affective picture system (IAPS): Affective ratings of pictures and instruction manual . Retrieved from http://csea.phhp.ufl.edu/Media.html Laureys, S., Owen, A. M., & Schiff, N. D. (2004). Brain function in coma, vegetative state, and related disorders. The Lancet Neurology , 3 (9), 537 546. doi:10.1016/S1474 4422(04)00852 X Lee, S., Halder, S., Kübler, A., Birbaumer, N., & Sitaram, R. (2010). Effective functional mapping of fMRI data with support vector machi nes. Human Brain Mapping , 31 (10), 1502 1511. doi:10.1002/hbm.20955

PAGE 85

85 Leibniz, G. W. (1714). The Monadology . Forgotten Books. Liberati, G., Dalboni da Rocha, J. L., van der Heiden, L., Raffone, A., Birbaumer, N., Olivetti Belardinelli, M., & Sitaram, R. (2012 ). Toward a brain computer interface state classification. , 31 Suppl 3 , S211 220. doi:10.3233/JAD 2012 112129 Luzzi, S., Piccirilli, M., & Pr ovinciali, L. (2007). Perception of emotions on happy/sad chimeric faces in Alzheimer disease: relationship with cognitive functions. Alzheimer Disease and Associated Disorders , 21 (2), 130 135. doi:10.1097/WAD.0b013e318064f445 Mahmoudi, A., Takerkart, S., Regragui, F., Boussaoud, D., & Brovelli, A. (2012). Multivoxel Pattern Analysis for fMRI Data: A Review. Computational and Mathematical Methods in Medicine , 2012 , e961257. doi:10.1155/2012/961257 McGinn, C. (1993). Problems in Philosophy: The Limits of Inq uiry (1 edition.). Oxford, Blackwell. perceive pain? Issues Law Med , 6 (4), 373 383. Medical Consultants on the Diagnosis of Death. (1981). Guidel ines for the determination of death. Report of the medical consultants on the diagnosis of death to the Biomedical and Behavioral Research. JAMA: The Journal of the American Medical A ssociation , 246 (19), 2184 2186. Medin, D., Ross, B. H., & Markman, A. B. (2004). Cognitive Psychology (4 edition.). Hoboken, NJ: Wiley. Menon, D., Owen, A., Williams, E., Minhas, P., Allen, C., Boniface, S., & Pickard, J. (1998). Cortical processing in per sistent vegetative state. The Lancet , 352 (9123), 200. doi:10.1016/S0140 6736(05)77805 3 Mochizuki, A. A., Sudo, M. M., Kirino, E., & Itoh, K. (2014). Brain activation associated with motor imagery of coordination exercises and social abilities. European Jo urnal of Sport Science . doi:10.1080/17461391.2014.893019 Morgenthaler, G. W. (1962). Review of Adaptive Control Processes, A Guided Tour by Richard Bellman. Operations Research , 10 (1), 143 146. Mourão Miranda, J., Friston, K. J., & Brammer, M. (2007). Dynamic discrimination analysis: a spatial temporal SVM. NeuroImage , 36 (1), 88 99. doi:10.1016/j.neuroimage.2007.02.020

PAGE 86

86 Nachev, P., Kennard, C., & Husain, M. (2008). Functional role of the supplemen tary and pre supplementary motor areas. Nature Reviews Neuroscience , 9 (11), 856 869. doi:10.1038/nrn2478 Naito, E., Kochiyama, T., Kitada, R., Nakamura, S., Matsumura, M., Yonekura, Y., & Sadato, N. (2002). Internally simulated movement sensations during m otor imagery activate cortical motor areas and the cerebellum. The Journal of Neuroscience: The Official Journal of the Society for Neuroscience , 22 (9), 3683 3691. doi:20026282 Ogawa, S., Tank, D. W., Menon, R., Ellermann, J. M., Kim, S. G., Merkle, H., & Ugurbil, K. (1992). Intrinsic signal changes accompanying sensory stimulation: functional brain mapping with magnetic resonance imaging. Proceedings of the National Academy of Sciences of the United States of America , 89 (13), 5951 5955. Owen, A. M., & Cole man, M. R. (2008). Detecting awareness in the vegetative state. Annals of the New York Academy of Sciences , 1129 , 130 138. doi:10.1196/annals.1417.018 Owen, A. M., Coleman, M. R., Menon, D. K., Berry, E. L., Johnsrude, I. S., Rodd, J. M., Pickard, J. D. (2005). Using a hierarchical approach to investigate residual auditory cognition in persistent vegetative state. Progress in Brain Research , 150 , 457 471. doi:10.1016/S0079 6123(05)50032 3 Packard, M. G., & Knowlton, B. J. (2002). Learning and memory funct ions of the Basal Ganglia. Annual Review of Neuroscience , 25 , 563 593. doi:10.1146/annurev.neuro.25.112701.142937 Pavlov, I. P., & Anrep, G. V. (2003). Conditioned Reflexes . Mineola, N.Y: Dover Publications. Pereira, F., Mitchell, T., & Botvinick, M. (2009 ). Machine learning classifiers and fMRI: a tutorial overview. NeuroImage , 45 (1 Suppl), S199 S209. doi:10.1016/j.neuroimage.2008.11.007 Peremen, Z., & Lamy, D. (2014). Do conscious perception and unconscious processing rely on independent mechanisms? A met a contrast study. Consciousness and Cognition , 24 , 22 32. doi:10.1016/j.concog.2013.12.006 Phan, K. L., Wager, T., Taylor, S. F., & Liberzon, I. (2002). Functional neuroanatomy of emotion: a meta analysis of emotion activation studies in PET and fMRI. Neur oImage , 16 (2), 331 348. doi:10.1006/nimg.2002.1087 Gray, J. A. (1998). Neural responses to facial and vocal expressions of fear and disgust. Proceedings. Biological S ciences / The Royal Society , 265 (1408), 1809 1817. doi:10.1098/rspb.1998.0506

PAGE 87

87 Poca, M. A., Sahuquillo, J., Mataró, M., Benejam, B., Arikan, F., & Báguena, M. (2005). Ventricular enlargement after moderate or severe head injury: a frequent and neglected pro blem. Journal of Neurotrauma , 22 (11), 1303 1310. doi:10.1089/neu.2005.22.1303 R.O. Duda, P.E. Hart, & D.G. Stork. (2007). Pattern Classification. J. Classif. , 24 (2), 305 307. doi:10.1007/s00357 007 0015 9 Rana, M., Gupta, N., Dalboni Da Rocha, J. L., Lee, S., & Sitaram, R. (2013). A toolbox for real time subject independent and subject dependent classification of brain states from fMRI signals. Frontiers in Neuroscience , 7 , 170. doi:10.3389/fnins.2013.00170 Ruiz, S., Buyukturkoglu, K., Rana, M., Birbaumer, N., & Sitaram, R. (2014). Real time fMRI brain computer interfaces: self regulation of single brain regions to networks. Biological Psychology , 95 , 4 20. doi:10.1016/j.biopsycho.2013.04.010 Schlkopf, B., & Smola, A. J. (2001). Learning with Kernels: Suppor t Vector Machines, Regularization, Optimization, and Beyond (1st edition.). Cambridge, Mass: The MIT Press. Schnakers, C., Vanhaudenhuyse, A., Giacino, J., Ventura, M., Boly, M., Majerus, S. Laureys, S. (2009). Diagnostic accuracy of the vegetative and minimally conscious state: Clinical consensus versus standardized neurobehavioral assessment. BMC Neurology , 9 (1), 35. doi:10.1186/1471 2377 9 35 Schölkopf, B., Burges, C. J. C., & Smola, A. J. (1998). Advances in Kernel Methods: Support Vector Learning . C ambridge, Mass: The MIT Press. Sehlmeyer, C., Schöning, S., Zwitserlood, P., Pfleiderer, B., Kircher, T., Arolt, V., & Konrad, C. (2009). Human fear conditioning and extinction in neuroimaging: a systematic review. PloS One , 4 (6), e5865. doi:10.1371/journa l.pone.0005865 Sitaram, R., Lee, S., Ruiz, S., Rana, M., Veit, R., & Birbaumer, N. (2011). Real time support vector classification and feedback of multiple emotional brain states. NeuroImage , 56 (2), 753 765. doi:10.1016/j.neuroimage.2010.08.007 Sleigh, J., & Warnaby, C. E. (2014). Functional brain imaging: gatecrashing the clinical party? The Lancet . doi:10.1016/S0140 6736(14)60223 3 Smith, M. (2012). Brain death: time for an international consensus. British Journal of Anaesthesia , 108 (suppl 1), i6 i9. doi: 10.1093/bja/aer355 Smith, S. M., Jenkinson, M., Woolrich, M. W., Beckmann, C. F., Behrens , T. E. J., Johansen Berg, H., Matthews, P. M. (2004). Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage , 23 Suppl 1 , S20 8 219. doi:10.1016/j.neuroimage.2004.07.051

PAGE 88

88 Soon, C. S., Brass, M., Heinze, H. J., & Haynes, J. D. (2008). Unconscious determinants of free decisions in the human brain. Nature Neuroscience , 11 (5), 543 545. doi:10.1038/nn.2112 Starr, C. J., Sawaki, L., Wit Coghill, R. C. (2011). The contribution of the putamen to sensory aspects of pain: insights from structural connectivity and brain lesions. Brain: A Journal of Neurology , 134 (Pt 7), 1987 2004. doi:10.1093/brain/awr117 Stender, J., Gosseries, O., Bruno, M. A., Charland Verville, V., Vanh audenhuyse, A., Demertzi, A., Laureys, S. (2014). Diagnostic precision of PET imaging and functional MRI in disorders of consciousness: a clinical validation st udy. The Lancet . doi:10.1016/S0140 6736(14)60042 8 Talairach, J., & Tournoux, P. (1988). Co Planar Stereotaxic Atlas of the Human Brain: 3 D Proportional System: An Approach to Cerebral Imaging (1st Edition edition.). Tanabe, J., Miller, D., Tregellas, J., Freedman, R., & Meyer, F. G. (2002). Comparison of detrending methods for optimal fMRI preprocessing. NeuroImage , 15 (4), 902 907. doi:10.1006/nimg.2002.1053 The M ulti Society Task Force on Persistent Vegetative State. (1994a). Medical aspects of the persistent vegetative state (1). The Multi Society Task Force on PVS. The New England Journal of Medicine , 330 (21), 1499 1508. doi:10.1056/NEJM199405263302107 The Multi Society Task Force on Persistent Vegetative State. (1994b). Medical aspects of the persistent vegetative state (2). The Multi Society Task Force on PVS. The New England Journal of Medicine , 330 (22), 1572 1579. doi:10.1056/NEJM199406023302206 Tzourio Mazoy er, N., Landeau, B., Papathanassiou, D., Crivello , F., Etard, O., Delcroix, N., Joliot, M. (2002). Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single subject brain. NeuroImage , 15 (1), 273 289. doi:10.1006/nimg.2001.0978 Vapnik, V. (2000). The Nature of Statistical Learning Theory (2nd ed.). Retrieved from http://www.springer.com/statistics/physical+%26+information+science/book/978 0 387 98780 4 Vapnik, V., & Lerner, A. (1963). Pattern Reco gnition using Generalized Portrait Method. Automation and Remote Control , 24 . Von Wild, K., Laureys, S. T., Gerstenbrand, F., Dolce, G., & Onose, G. (2012). The vegetative state a syndrome in search of a name. Journal of Medicine and Life , 5 (1), 3 15.

PAGE 89

89 Weis kopf, N., Scharnowski, F., Veit, R., Goebel, R., Birbaumer, N., & Mathiak, K. (2004). Self regulation of local brain activity using real time functional magnetic resonance imaging (fMRI). Journal of Physiology, Paris , 98 (4 6), 357 373. doi:10.1016/j.jphysp aris.2005.09.019 Young, G. B., & Wijdicks, E. F. M. (Eds.). (2008). Disorders of consciousness . Zelikowsky, M., Hersman, S., Chawla, M. K., Barnes, C. A., & Fanselow, M. S. (2014). Neuronal ensembles in amygdala, hippocampus, and prefrontal cortex track differential components of contextual fear. The Journal of Neuroscience: The Official Journal of the Society for Neuroscience , 34 (25), 8462 8466. doi:10.1523/JNEUROSCI.3624 13.2014 Zeman, A. (1997). Persistent vegetative state. The Lancet , 350 (9080), 795 799. doi:10.1016/S0140 6736(97)06447 7

PAGE 90

90 BIOGRAPHICAL SKETCH Enrico Opri earned a BS in Biomedical Engineering from Politecnico of Milan (Italy) and a double MS Degree between U niversity of F lorida and Politecnico of Milan in Biomedical Engineering , thanks to the Atlantis CRISP program . His interests are decision making and neural interfaces. In his spare time he likes to code and work on R obotics (while bak ing cakes or playing his viola) .