<p> The study of the different sleep stages of a patient using his/her recorded EEG signals falls in the area of signal classification. In general, this involves extracting from the EEG signals, a signal feature on which the classification is performed. In this thesis, we apply the techniques of signal classification to the analysis of the sleep of a patient. The feature we use is the power spectral density (PSD) matrices of a multi-channel EEG signal. This not only allows us to examine the power spectrum contents of each signal which complies with what clinical experts use in their visual judgement of EEG signals, but also allows the correlation between the multi-channel signals to be studied. To establish a metric facilitating the classification, we analyze the structure as well as exploit the specific geometric properties of the space of PSD matrices. Specifically, we study this space from the viewpoint of Riemannian manifolds. We apply a Riemannian metric and, with the aid of fibre bundle theory, develop intrinsic (geodesic) distance measures for the PSD matrix manifold. To utilize such new distance measures effectively for EEG signal classification, we need to find a suitable weighting matrix for the PSD matrices so that the distances between similar features are minimized while those between dissimilar features are maximized. A closed form expression for this weighting matrix is obtained by solving an equivalent convex optimization problem. The effectiveness of using these novel weighted distance measures is verified by applying them to the sleep pattern classification of a collection of recorded EEG signals using the k-nearest neighbor decision algorithm with excellent results. </p> / Thesis / Doctor of Philosophy (PhD)

El texto completo de este trabajo no está disponible en el Repositorio Académico UPC por restricciones de la casa editorial donde ha sido publicado. / This paper seeks to study the correspondence between the results of the test of variable of attention (TOVA) and the signals acquired by the Muse electroencephalogram (EEG) in the positions AF7 and AF8 of the cerebral cortex. There are a variety of research papers that estimates an index of attention in which the different characteristics in discrete signals of the brain activity were used. However, many of these results were obtained without contrasting them with standardized tests. Due to this fact, in the present work, the results will be compared with the score of the TOVA, which aims to identify an attention disorder in a person. The indicators obtained from the test are the response time variability, the average response time, and the d′ prime score. During the test, the characteristics of the EEG signals in the alpha, beta, theta, and gamma subbands such as the energy, average power, and standard deviation were extracted. For this purpose, the acquired signals are filtered to reduce the effect of the movement of the muscles near the cerebral cortex and then went through a subband decomposition process by applying transformed wavelet packets. The results show a well-marked correspondence between the parameters of the EEG signal of the indicated subbands and the visual attention indicators provided by TOVA. This correspondence was measured through Pearson’s correlation coefficient which had an average result of 0.8. / Revisión por pares / Revisión por pares

This dissertation aims to develop a formal foundation to design an adaptive output feedback predictor for a class of unknown systems where parameters and order are unknown or high-dimensional. We present a reduced-order adaptive output-predictor scheme based on modal reduction and Lyapunov's method. Moreover, the credibility of the proposed reduced-order adaptive output-predictor scheme is validated by mathematical proof, and numerical and experimental studies, such as single pendulum, double pendulum, six-link pendulum, rope as a high-dimensional rope, and EEG data. Then the dissertation goal is to experimentally validate the proposed reduced-order model parameterization technique for tracking uncertain linear time-invariant (LTI) single-input, single-output (SISO) systems. The proposed theory focuses on parameterizing a high-dimensional, uncertain model and introduces a reduced-order adaptive output predictor capable of forecasting the system's output. This predictor utilizes auto-regressive filtered vectors, incorporating the input and output history. The adaptive output predictor is a simplified and known model, making it suitable for controlling high-dimensional, uncertain SISO systems without access to full-state measurements. Specifically, this work establishes the foundation for parameterizing uncertain models, creating a virtual structure that emulates the actual system, and offering a more manageable model for control when the objective is solely to regulate the system's output. The primary focus of this research is to assess the effectiveness and output-tracking capabilities of the proposed approach. These capabilities are extensively examined across diverse platforms and hardware configurations, relying solely on input and output data from the models without incorporating any additional information on the system dynamics. In the first experiment, the predictor's ability to track the angle of a single pendulum, including additional dynamics, is evaluated using only input-output data. The second experiment targets tracking the endpoint of a rope connected to a single pendulum, where the rope emulates a high-dimensional model. A vision system is designed and employed to acquire the rope endpoint position data. Before the rope experiment, a set of experiments is conducted on single pendulum hardware to ensure the accuracy of the vision system's data collection. Comparative analysis between data from object tracking via vision and data acquired through an encoder demonstrates negligible error. Finally, the input and the endpoint output data from the rope experiment are fed into the predictor to assess its capability to track the rope endpoint position without utilizing specific knowledge of the experimental hardware. Achieving negligible error in tracking implies that the predictor provides a simple and accurate representation of the rope dynamics. Consequently, designing a controller for this known model is equivalent to designing a controller for the actual rope system dynamics. The predictor, by closely emulating the behavior of the rope, becomes a reliable surrogate model for control design, simplifying the task of controller design for the complex and uncertain high-dimensional system. Finally, this study introduces a novel approach to enhance controller design for complex brain dynamics by employing a reduced-order adaptive output predictor proposed in [1], fine-tuned with chirp binaural beats. The proposed technique is promising for developing closed-loop controllers in non-invasive brain stimulation therapies, such as binaural beats stimulation, to improve working memory. The study focuses on parameterizing uncertain models and creates a predictor that utilizes auto-regressive filtered vectors to forecast mean phase lock values generated by binaural beats stimulation. The simplified and known model of the predictor proves effective in tracking brain responses, as demonstrated in experiments evaluating its ability to track mean phase locking values. The results indicate negligible tracking error, suggesting the predictor's reliability in representing brain dynamics and simplifying the task of controller design for the complex and uncertain high-dimensional system. / Doctor of Philosophy / This dissertation explores the development of a reduced-order adaptive output predictor for unknown systems with unknown or high-dimensional parameters and order. A reduced-order adaptive output predictor scheme is introduced, validated through mathematical proof, and tested in diverse scenarios, including pendulum systems and EEG data. The focus is on parameterizing uncertain models and creating a simplified adaptive output predictor capable of forecasting system output, specifically for SISO systems. Experimental validation involves tracking the angle of a single pendulum and the endpoint of a high-dimensional rope, demonstrating the predictor's accuracy without detailed knowledge of system dynamics. The study extends its application to complex brain dynamics, using the predictor fine-tuned with chirp binaural beats. Results show promise for developing closed-loop controllers in non-invasive brain stimulation therapies, offering a novel approach to improve working memory via helping to design closed-loop controllers.

Physiological signals, electroencephalogram (EEG) and heart rate variability (HRV), are generated by complex self-regulating systems. These signals are extremely inhomogeneous and nonstationary, and fluctuate in an irregular and highly complex manner. These fluctuations are due to underlying dynamics of the system. The synchronous neural activity measured as scalp EEG indicates underlying neural dynamics of the brain. Hence, quantitative EEG analysis has become a very useful tool in interpreting results from physiological experiments. The analysis of HRV provides valuable information to assess the autonomous nervous system (ANS). The HRV can be significantly affected by physiological state changes and many disease states. Hence, HRV analysis is becoming a major experimental and diagnostic tool. In this thesis, we focus on the study of EEG and HRV time series using tools from nonlinear time series analysis with special emphasis on its implications in detecting physiological state changes such as, in diseases like epileptic seizure and schizophrenia, and in altered states of consciousness as in sleep and meditation. The proposed nonlinear techniques are used in discriminating different physiological states from control states. Artifact processing of EEG signal Interferences (artifacts) from various sources unavoidably contaminate EEG recordings. In quantitative analysis, results can differ significantly by these artifacts, which may lead to wrong interpretation of the results. In this part of the thesis, we have devised methods to minimize ocular and muscle artifacts in EEG. The artifact correction methods are based on blind source separation (BSS) techniques such as singular value decomposition (SVD), algorithm for multiple signal extraction (AMUSE), canonical correlation analysis (CCA), information maximization (INFOMAX) independent component analysis (ICA) and joint approximate diagonalization of eigen-matrices (JADE) ICA. We have proposed a method to simulate clean and artifact corrupted EEG data based on the BSS methods. In order to enhance the performance of BSS methods, a technique called wavelet-filtered component inclusion method has been introduced. In addition, second-order statistics (SOS) and higher-order statistics (HOS) based BSS methods have been studied considering less number of EEG channels; and performance comparison of these methods has also been made. We have also addressed the problem of simultaneous correction of ocular and muscle artifacts in EEG recordings using the BSS methods. Irrespective of the BSS methods, the component elimination method has introduced high spectral error in all the bands after reconstruction of clean EEG. However, the wavelet filtered component inclusion method has retained almost all spectral powers of EEG channels in theta, alpha, and beta bands after ocular artifact minimization. When the number of EEG channels is very less, the enhanced CCA (SOS BSS) has given superior artifact minimization results than HOS BSS methods, especially in delta band. The component elimination method is used in muscle artifact minimization, and hence the SVD method cannot be used for this purpose since it leads to large spectral distortion of reconstructed EEG. The AMUSE and CCA methods have given comparable performance in muscle artifact minimization. In addition, the JADE method has introduced less mean spectral error compared to other methods. The CCA method has shown superior performance in simultaneous minimization of ocular and muscle artifacts, and AMUSE and JADE methods have given comparable results. Furthermore, the less computation time of wavelet enhanced SOS BSS methods make them very useful in real clinical environments. Fractal characterization of time series In biomedical signal analysis, fractal dimension (FD) is used as a quantitative measure to estimate complexity of physiological signals. Such analysis helps to study physiological processes of underlying systems. The FD can also be used to study dynamics of transitions between different states of systems like brain and ANS, in various physiological and pathological states. In this part, we have proposed a method to estimate FD of time series, called multiresolution box-counting (MRBC) method. A modification of this method resulted in multiresolution length (MRL) method. The estimation performance of the proposed methods is compared with that of Katz, Sevcik, and Higuchi methods, by simulating mathematically defined fractal signals, and also the computation time is compared between the methods. The MRBC and MRL methods have given comparable performance to that of Higuchi method, in estimating FD of waveforms, with the advantage of less computational time. In addition, various properties of the FD are studied and discussed in connection with classical signal processing concepts such as amplitude, frequency, sampling frequency, effect of noise, band width, correlation, etc. The FD value of signals has increased with number of harmonics, noise variance, band-width, and mid-band frequency, and decreased with degree of correlation in AR signal. An analogy between Katz FD and smoothed Teager energy operator has also been made. Application of fractal analysis to EEG and HRV time series The fluctuation of EEG potentials normally depends upon degree of alertness, and varies in amplitude and frequency. Hence, the EEG is an important clinical tool for studying sleep and sleep related disorders, epileptic seizures, schizophrenia, and meditation. In this part of the thesis, we have used FD which gives signal complexity, and detrended fluctuation analysis (DFA) which gives multiscale exponent of time series to quantify EEG. We have extended the concept of FD to multiscale FD to compute complexity of time series at multiple scales. The main applications of the proposed method are epileptic seizure detection, sleep stage detection, schizophrenia EEG analysis, and analysis of heart rate variability during meditation. For seizure detection, we have used intracranial EEG recordings with seizure-free and seizure intervals. In sleep EEG analysis, whole-night sleep EEG is used and results are compared with the manually scored hypnogram. The schizophrenia symptom is further categorized into positive and negative symptoms and complexity is estimated using FD and DFA. We have also analyzed HRV data of Chi and Kundalini meditation using FD and DFA techniques. In all the applications considered, we have tested for statistical significance of the computed parameters, between the case of interest and corresponding control cases, to discriminate between the physiological states. The ocular artifact has reduced FD while muscle artifact increased FD of EEG. The FD of seizure EEG has shown high value compared to that of seizure-free EEG. In addition, the seizure-free EEG has more DFA exponent-1 than seizure EEG. The value of FD of EEG is decreased with deepening of sleep, wake state having high FD value. The FD of REM state sleep EEG showed value between that of wake and state-1. The DFA exponent-1 has increased with deepening of sleep state, having small value for wake state. The REM state has given exponent-1 value between wake and state-1. The schizophrenia subjects have shown lower FD value than healthy controls in all the EEG channels except the bilateral temporal and occipital regions. The positive symptom sub-group has shown comparatively high FD values than healthy controls as well as overall schizophrenia sample in the bilateral tempero-parietal-occipital region. In addition, the positive symptom sub-group has shown significantly higher regional FD values than negative symptom sub-group especially in right temporal region. The overall schizophrenia samples as well as the positive and negative subgroup have shown least FD values in the bilateral frontal region. The values of DFA exponent-2 have shown significant high value in schizophrenia samples. In addition, the schizophrenia group has shown less DFA exponent-1 in bilateral temporal region than healthy control. The FD, multiscale FD, DFA exponents have shown significant performance in discriminating different physiological states from control states. The FD value of HRV time series during meditation is less compared to pre-meditation state in both Chi and Kundalini meditation. Irrespective of the type of meditation, meditation state has shown significantly high DFA exponent-1 than pre-meditation state, and significantly high DFA exponent-2 in pre-meditation state compared to meditation state. Functional connectivity analysis of brain during meditation In functionally related regions of the brain, even in those regions separated by substantial distances, the EEG fluctuations are synchronous, which is termed as functional connectivity. In this part, a novel application of functional connectivity analysis of brain using graph theoretic approach has been made on the EEG recorded from meditation practitioners. We have used 16 channel EEG data from subjects while performing Raja Yoga meditation. The pre-meditation condition is used as control state, against which meditation state is compared. For finding connectivity between EEG of various channels, we have computed pair-wise linear correlation and mutual information between the EEG channels, to form a connection matrix of size 16x16. Then, various graph parameters, such as average connection density, degree of nodes, characteristic path length, and cluster index, are computed from the connection matrix. The computed parameters are projected on to the scalp to get topographic head maps that give spatial variation of the parameter, and results are compared between meditation and pre-meditation states. The meditation state has shown low average connection density, less characteristic path length, and high average degree in fronto-central and central regions. Furthermore, high cluster index is shown in frontal and central regions than pre-meditation state. The parameters such as complexity, characteristic path length and average connection density are used as features in quadratic discriminant classifier to classify meditation and pre-meditation state, and have given good accuracy performance. Connectivity analysis using mutual information has given high average connection density in meditation state in theta, alpha and beta bands compared to pre-meditation state. The characteristic path length is high in delta, alpha and beta bands in meditation state. In addition, the meditation state has shown high degree and cluster index in theta and beta bands compared to pre-meditation state. Nonlinear dynamical characterization of HRV during meditation The cardiovascular system is influenced by internal dynamics as well as from various external factors, which makes the system more dynamic and nonlinear. In this part of the thesis, a novel application of using HRV data for studying Chi and Kundalini meditation has been made. The HRV time series are embedded into higher dimensional phase-space using Takens’ embedding theorem to reconstruct the attractor. After estimating the minimum embedding dimension to unfold the attractor dynamics, the complexity of the attractor is computed using correlation dimension, Lyapunov exponent, and nonlinearity scores. In all the analyses, the pre-meditation state is used as control state against which meditation state is compared. The statistical significance of the parameters estimated is tested to discriminate meditation state from control state. The HRV time series of both pre-meditation and meditation have shown similar minimum embedding dimensions in both Chi and Kundalini meditation. Irrespective of the type of meditation, the meditation state has shown high correlation dimension, largest Lyapunov exponent, and low nonlinearity score compared to pre-meditation state. Recurrent quantification analysis of HRV during meditation In this part, a novel application of recurrent quantification analysis (RQA) to HRV during meditation is studied. Here, the time series is embedded into a higher dimensional phase-space and Euclidean distance between the embedded vectors is calculated to form a distance matrix. The matrix is converted into binary matrix by applying a suitable threshold, and plotted as image to get recurrence plot. Various parameters are extracted from the recurrence plot such as percent recurrence rate, diagonal parameters (determinism, divergence, entropy, ratio), and vertical or horizontal parameters (laminarity, trapping time, maximal vertical line length). The procedure is applied to HRV data during meditation and pre-meditation (control) to discriminate between the states. The HRV of meditation state has shown more diagonal line structure whereas more black patches are observed in pre-meditation state. In addition, at low embedding dimensions, the meditation state has shown low recurrence rate, high determinism, low divergence, low entropy, high ratio, high laminarity, high trapping time, and less maximal vertical line length compared to pre-meditation state. These RQA parameters have shown superior performance in discriminating meditation state from control state.

<p style="text-align: justify">Cilj ove doktorske disertacije je da predstavi model i računarski sistem koji re&scaron;ava problem osoba sa invaliditetom koja nisu u mogućnosti da koriste ruke ili funkciju govora u ostvarivanju komunikacije. Disertacija se bavi problematikom mogućnosti primene ekonomski pristupačnih asistivnih tehnologija u domenu primene servisa elektronske uprave. Od asistivnih tehnologija disertacija istražuje mogućnosti primene neinvazivne BCI tehnologije u poređenju sa sistemima baziranih na HD kamerama. Razvijen je računarski sistem koji omogućava integraciju u operativni sistem i upotrebu računara za unos komandi upotrebom detekcije moždanih talasa.</p> / <p>The main objective of this doctoral thesis is to present the model and a computer system that solves the communication problem of people with disabilities (people who cannot use their hands or the function of speech communication). The dissertation researches the possibility to apply economic affordable assistive technologies in the field of application of e-government services. Thesis explores the possibilities of application of non-invasive BCI technology compared to systems based on HD<br />cameras. Has been developed a computer system that allows the integration into the<br />operating system that allow to enter commands by the detection of brain waves.</p>

En raison de la grande quantité d’EEG acquise sur plusieurs journées, une technique de compression efficace est nécessaire. Le manque des experts et la courte durée des crises encouragent la détection automatique des convulsions. Un affichage uniforme est obligatoire pour assurer l’interopérabilité et la lecture des examens EEG transmis. Le codeur certifié médical WAAVES fournit des CR élevés et assure une qualité de diagnostic d’image. Durant nos travaux, trois défis sont révélés : adapter WAAVES à la compression des signaux, détecter automatiquement les crises épileptiques et assurer l’interopérabilité des afficheurs EEG. L’étude du codeur montre qu’il est incapable de supprimer la corrélation spatiale et de compresser des signaux monodimensionnels. Par conséquent, nous avons appliqué l’ICA pour décorréler les signaux, la mise en échelle pour redimensionner les valeurs décimales et la construction d’image. Pour garder une qualité de diagnostic avec un PDR inférieur à 7%, nous avons codé le résidu. L’algorithme de compression EEGWaaves proposé a atteint des CR de l’ordre de 56. Ensuite, nous avons proposé une méthode d’extraction des caractéristiques des signaux EEG basée sur un nouveau modèle de calcul de la prédiction énergétique (EAM) des signaux. Ensuite, des paramètres statistiques ont été calculés et les Réseaux de Neurones ont été appliqués pour détecter les crises épileptiques. Cette méthode nous a permis d’atteindre de meilleure sensibilité allant jusqu’à 100% et une précision de 99.44%. Le dernier chapitre détaille le déploiement de notre afficheur multi-plateforme des signaux physiologiques. Il assure l’interopérabilité des examens EEG entre les hôpitaux. / Due to the large amount of EEG acquired over several days, an efficient compression technique is necessary. The lack of experts and the short duration of epileptic seizures require the automatic detection of these seizures. Furthermore, a uniform viewer is mandatory to ensure interoperability and a correct reading of transmitted EEG exams. The certified medical image WAAVES coder provides high compression ratios CR while ensuring image quality. During our thesis, three challenges are revealed : adapting WAAVES coder to the compression of the EEG signals, detecting automatically epileptic seizures in an EEG signal and ensure the interoperability of the displays of EEG exams. The study of WAAVES shows that this coder is unable to remove spatial correlation and to compress directly monodimensional signals. Therefore, we applied ICA to decorrelate signals, a scaling to resize decimal values, and image construction. To keep a diagnostic quality with a PDR less than 7%, we coded the residue. The proposed compression algorithm EEGWaaves has achieved CR equal to 56. Subsequently, we proposed a new method of EEG feature extraction based on a new calculation model of the energy expected measurement (EAM) of EEG signals. Then, statistical parameters were calculated and Neural Networks were applied to classify and detect epileptic seizures. Our method allowed to achieve a better sensitivity up to 100% and an accuracy of 99.44%. The last chapter details the deployment of our multiplatform display of physiological signals by meeting the specifications established by doctors. The main role of this software is to ensure the interoperability of EEG exams between healthcare centers.

El texto completo de este trabajo no está disponible en el Repositorio Académico UPC por restricciones de la casa editorial donde ha sido publicado. / The present work proposes an algorithm to detect and identify the artifact signals produced by the concrete gestural actions of jaw clench and eyebrows raising in the electroencephalography (EEG) signal. Artifacts are signals that manifest in the EEG signal but do not come from the brain but from other sources such as flickering, electrical noise, muscle movements, breathing, and heartbeat. The proposed algorithm makes use of concepts and knowledge in the field of signal processing, such as signal energy, zero crossings, and block processing, to correctly classify the aforementioned artifact signals. The algorithm showed a 90% detection accuracy when evaluated in independent ten-second registers in which the gestural events of interest were induced, then the samples were processed, and the detection was performed. The detection and identification of these devices can be used as commands in a brain–computer interface (BCI) of various applications, such as games, control systems of some type of hardware of special benefit for disabled people, such as a chair wheel, a robot or mechanical arm, a computer pointer control interface, an Internet of things (IoT) control or some communication system. / Revisión por pares

Hjärndatorgränssnitt är enheter som fångar upp hjärnsignaler via elektroder på huvudet och översätter dem till datamängder och instruktioner mot externa enheter och applikationer. Gränssnitten har främst använts inom den medicinska domänen för att hjälpa personer med neurofysiologiska åkommor, men har även på senare tid börjat användas av ickemedicinska skäl av privatpersoner. I takt med att gränssnitten ökar i popularitet och når en bredare massa kommer det att innebära ett större informationsflöde av användardata som i sin tur kan bära på väldigt känslig information. Information såsom hälsodata och autentiseringsmetoder är några av flera informationstillgångar som ligger i farozonen enligt flera artiklar och kan råka ut för ett eller flera hot. För få en tydligare bild av de olika hoten samt dess konsekvens och sannolikhet har det genomförts en riskanalys gällande hemanvändares informationssäkerhet. För att få fram sårbarheter, hot och åtgärder som förekommer i riskanalysen har det utförts en tematisk analys. Genom den tematiska analysen visade det sig att det fanns flera hot mot hemanvändarnas konfidentialitet där användares PIN-koder, autentiseringsmetoder och hälsodata låg i farozonen. För att få en bättre förståelse kring hur gränssnitten fungerar samt hur stor sannolikhet det är för olika hot har det även genomförts en intervju med en lektor i kognitiv neurovetenskap, följande tillsammans med artiklarna från den tematiska analysen utgjorde därmed grunden för riskanalysen. Genom riskanalysen visade det sig att hoten mot hemanvändarnas möjlighet att använda gränsssnitten hade en ännu större sannolikhet att inträffa än hot mot användares konfidentialitet. / Brain- Computer Interfaces are devices that capture brain signals via electrodes on the head and then translates them into data sets and instructions to external devices and applications. The interfaces have mainly been used in the medical domain to help people with neurophysiological disorders but have also recently begun to be used for non-medical reasons by private persons. As the interfaces increase in popularity and reach a wider mass, it will mean a greater flow of information of user data that in turn can carry very sensitive information. Information such as health data and authentication methods are some of several information assets that are at risk according to multiple articles and may face one or more threats. To get a clearer picture of the various threats, their consequences and probabilities, a risk analysis has been carried out. In order to identify vulnerabilities, threats and measures that appear in the risk analysis, a thematic analysis has been performed. The thematic coding showed that there were several threats to the home user’s confidentiality where user’s PIN-codes and health data were at risk. In order to gain a better understanding of how the interfaces work and how likely it is for various threats to succeed, an interview was conducted with a senior lectrurer in cognitive neuroscience, the following together with the articles from the thematic analysis thus formed the basis for the risk analysis. The risk analysis showed that threats to home users' ability to use the interfaces were even more likely to occur than threats to user confidentiality.

For most of the naturally occurring signals, especially biomedical signals, the underlying physical process generating the signal is often not fully known, making it difficult to obtain a parametric model. Therefore, signal processing techniques are used to analyze the signal for non-parametrically characterizing the underlying system from which the signals are produced. Most of the real life systems are nonlinear and time varying, which poses a challenge while characterizing them. Additionally, multiple sensors are used to extract signals from such systems, resulting in multichannel signals which are inherently coupled. In this thesis, we counter this challenge by using Recurrence Plot based techniques for characterizing biomedical systems such as heart or brain, using signals such as heart rate variability (HRV), electroencephalogram(EEG) or functional magnetic resonance imaging (fMRI), respectively, extracted from them. In time series analysis, it is well known that a system can be represented by a trajectory in an N-dimensional state space, which completely represents an instance of the system behavior. Such a system characterization has been done using dynamical invariants such as correlation dimension, Lyapunov exponent etc. Takens has shown that when the state variables of the underlying system are not known, one can obtain a trajectory in ‘phase space’ using only the signals obtained from such a system. The phase space trajectory is topologically equivalent to the state space trajectory. This enables us to characterize the system behavior from only the signals sensed from them. However, estimation of correlation dimension, Lyapunov exponent, etc, are vulnerable to non-stationarities in the signal and require large number of sample points for accurate computation, both of which are important in the case of biomedical signals. Alternatively, a technique called Recurrence Plots (RP) has been proposed, which addresses these concerns, apart from providing additional insights. Measures to characterize RPs of single and two channel data are called Recurrence Quantification Analysis (RQA) and cross RQA (CRQA), respectively. These methods have been applied with a good measure of success in diverse areas. However, they have not been studied extensively in the context of experimental biomedical signals, especially multichannel data. In this thesis, the RP technique and its associated measures are briefly reviewed. Using the computational tools developed for this thesis, RP technique has been applied on select single channel, multichannel and multimodal (i.e. multiple channels derived from different modalities) biomedical signals. Connectivity analysis is demonstrated as post-processing of RP analysis on multichannel signals such as EEG and fMRI. Finally, a novel metric, based on the modification of a CRQA measure is proposed, which shows improved results. For the case of single channel signal, we have considered a large database of HRV signals of 112 subjects recorded for both normal and abnormal (anxiety disorder and depression disorder) subjects, in both supine and standing positions. Existing RQA measures, Recurrence Rate and Determinism, were used to distinguish between normal and abnormal subjects with an accuracy of 58.93%. A new measure, MLV has been introduced, using which a classification accuracy of 98.2% is obtained. Correlation between probabilities of recurrence (CPR) is a CRQA measure used to characterize phase synchronization between two signals. In this work, we demonstrate its utility with application to multimodal and multichannel biomedical signals. First, for the multimodal case, we have computed running CPR (rCPR), a modification proposed by us, which allows dynamic estimation of CPR as a function of time, on multimodal cardiac signals (electrocardiogram and arterial blood pressure) and demonstrated that the method can clearly detect abnormalities (premature ventricular contractions); this has potential applications in cardiac care such as assisted automated diagnosis. Second, for the multichannel case, we have used 16 channel EEG signals recorded under various physiological states such as (i) global epileptic seizure and pre-seizure and (ii) focal epilepsy. CPR was computed pair-wise between the channels and a CPR matrix of all pairs was formed. Contour plot of the CPR matrix was obtained to illustrate synchronization. Statistical analysis of CPR matrix for 16 subjects of global epilepsy showed clear differences between pre-seizure and seizure conditions, and a linear discriminant classifier was used in distinguishing between the two conditions with 100% accuracy. Connectivity analysis of multichannel EEG signals was performed by post-processing of the CPR matrix to understand global network-level characterization of the brain. Brain connectivity using thresholded CPR matrix of multichannel EEG signals showed clear differences in the number and pattern of connections in brain connectivity graph between epileptic seizure and pre-seizure. Corresponding brain headmaps provide meaningful insights about synchronization in the brain in those states. K-means clustering of connectivity parameters of CPR and linear correlation obtained from global epileptic seizure and pre-seizure showed significantly larger cluster centroid distances for CPR as opposed to linear correlation, thereby demonstrating the efficacy of CPR. The headmap in the case of focal epilepsy clearly enables us to identify the focus of the epilepsy which provides certain diagnostic value. Connectivity analysis on multichannel fMRI signals was performed using CPR matrix and graph theoretic analysis. Adjacency matrix was obtained from CPR matrices after thresholding it using statistical significance tests. Graph theoretic analysis based on communicability was performed to obtain community structures for awake resting and anesthetic sedation states. Concurrent behavioral data showed memory impairment due to anesthesia. Given the fact that previous studies have implicated the hippocampus in memory function, the CPR results showing the hippocampus within the community in awake state and out of it in anesthesia state, demonstrated the biological plausibility of the CPR results. On the other hand, results from linear correlation were less biologically plausible. In biological systems, highly synchronized and desynchronized systems are of interest rather than moderately synchronized ones. However, CPR is approximately a monotonic function of synchronization and hence can assume values which indicate moderate synchronization. In order to emphasize high synchronization/ desynchronization and de-emphasize moderate synchronization, a new method of Correlation Synchronization Convergence Time (CSCT) is proposed. It is obtained using an iterative procedure involving the evaluation of CPR for successive autocorrelations until CPR converges to a chosen threshold. CSCT was evaluated for 16 channel EEG data and corresponding contour plots and histograms were obtained, which shows better discrimination between synchronized and asynchronized states compared to the conventional CPR. This thesis has demonstrated the efficacy of RP technique and associated measures in characterizing various classes of biomedical signals. The results obtained are corroborated by well known physiological facts, and they provide physiologically meaningful insights into the functioning of the underlying biological systems, with potential diagnostic value in healthcare.

[pt] Para o diagnóstico de distúrbios do sono, um dos exames mais usado é a polissonografia (PSG), na qual é registrada uma variedade de sinais fisiológicos. O exame de PSG é observado por um especialista do sono, processo que pode levar muito tempo e incorrer em erros de interpretação. O presente trabalho desenvolve e compara o desempenho de quatro sistemas baseados em arquiteturas de redes neurais de aprendizado profundo, mais especificamente, redes convolutivas (CNN) e redes recorrentes Long-Short Term Memory (LSTM), para a identificação de despertares relacionados ao esforço respiratório (Respiratory Effort-Related Arousal-RERA) e a eventos de despertar relacionados à apneia/hipopneia. Para o desenvolvimento desta pesquisa, foram usadas as informações de apenas seis canais eletroencefalográficos (EEG) provenientes de 994 registros de PSG noturna da base de dados PhysioNet CinC Challenge2018, além disso, foi considerado o uso de class weight e Focal Loss para lidar com o desbalanceamento de classes. Para a avaliação de cada um dos sistemas foram usadas a Accuracy, AUROC e AUPRC como métricas de desempenho. Os melhores resultados para o conjunto de teste foram obtidos com os modelos CNN1 obtendo-se uma Accuracy, AUROC e AUPRC de 0,8404, 0,8885 e 0,8141 respetivamente, e CNN2 obtendo-se uma Accuracy, AUROC e AUPRC de 0,8214, 0,8915 e 0,8097 respetivamente. Os resultados restantes confirmaram que as redes neurais de aprendizado profundo permitem lidar com dados temporais de EEG melhor que os algoritmos de aprendizado de máquina tradicional, e o uso de técnicas como class weight e Focal Loss melhoram o desempenho dos sistemas. / [en] For the diagnosis of sleep disorders, one of the most commonly used tests is polysomnography (PSG), in which a variety of physiological signs are recorded. The study of PSG is observed by a sleep therapist, This process may take a long time and may incur misinterpretation. This work develops and compares the performance of four classification systems based on deep learning neural networks, more specifically, convolutional neural networks (CNN) and recurrent networks Long-Short Term Memory (LSTM), for the identification of Respiratory Effort-Related Arousal (RERA) and to events related to apnea/hypopnea. For the development of this research, it was used the Electroencephalogram (EEG) data of six channels from 994 night polysomnography records from the database PhysioNet CinC Challenge2018, the use of class weight and Focal Loss was considered to deal with class unbalance. Accuracy, AUROC, and AUPRC were used as performance metrics for evaluating each system. The best results for the test set were obtained with the CNN1 models obtaining an accuracy, AUROC and AUPRC of 0.8404, 0.8885 and 0.8141 respectively, and RCNN2 obtaining an accuracy, AUROC and AUPRC of 0.8214, 0.8915 and 0.8097 respectively. The remaining results confirmed that deep learning neural networks allow dealing with EEG time data better than traditional machine learning algorithms, and the use of techniques such as class weight and Focal Loss improve system performance.