This thesis deals with the problem of newborn seizre detection from the Electroencephalogram (EEG) signals. The ultimate goal is to design an automated seizure detection system to assist the medical personnel in timely seizure detection. Seizure detection is vital as neurological diseases or dysfunctions in newborn infants are often first manifested by seizure and prolonged seizures can result in impaired neuro-development or even fatality. The EEG has proved superior to clinical examination of newborns in early detection and prognostication of brain dysfunctions. However, long-term newborn EEG signals acquisition is considerably more difficult than that of adults and children. This is because, the number of the electrodes attached to the skin is limited by the size of the head, the newborns EEGs vary from day to day, and the newborns are reluctant of being in the recording situation. Also, the movement of the newborn can create artifact in the recording and as a result strongly affect the electrical seizure recognition. Most of the existing methods for neonates are either time or frequency based, and, therefore, do not consider the non-stationarity nature of the EEG signal. Thus, notwithstanding the plethora of existing methods, this thesis applies the discrete wavelet transform (DWT) to account for the non-stationarity of the EEG signals. First, two methods for seizure detection in neonates are proposed. The detection schemes are based on observing the changing behaviour of a number of statistical quantities of the wavelet coefficients (WC) of the EEG signal at different scales. In the first method, the variance and mean of the WC are considered as a feature set to dassify the EEG data into seizure and non-seizure. The test results give an average seizure detection rate (SDR) of 97.4%. In the second method, the number of zero-crossings, and the average distance between adjacent extrema of the WC of certain scales are extracted to form a feature set. The test obtains an average SDR of 95.2%. The proposed feature sets are both simple to implement, have high detection rate and low false alarm rate. Then, in order to reduce the complexity of the proposed schemes, two optimising methods are used to reduce the number of selected features. First, the mutual information feature selection (MIFS) algorithm is applied to select the optimum feature subset. The results show that an optimal subset of 9 features, provides SDR of 94%. Compared to that of the full feature set, it is clear that the optimal feature set can significantly reduce the system complexity. The drawback of the MIFS algorithm is that it ignores the interaction between features. To overcome this drawback, an alternative algorithm, the mutual information evaluation function (MIEF) is then used. The MIEF evaluates a set of candidate features extracted from the WC to select an informative feature subset. This function is based on the measurement of the information gain and takes into consideration the interaction between features. The performance of the proposed features is evaluated and compared to that of the features obtained using the MIFS algorithm. The MIEF algorithm selected the optimal 10 features resulting an average SDR of 96.3%. It is also shown, an average SDR of 93.5% can be obtained with only 4 features when the MIEF algorithm is used. In comparison with results of the first two methods, it is shown that the optimal feature subsets improve the system performance and significantly reduce the system complexity for implementation purpose.

L’électroencéphalogramme (EEG) est un examen incontournable pour le diagnostic, la définition des structures cérébrales responsables de l’origine de crises et la classification des épilepsies. Cependant les enregistrements recueillis à la surface du scalp sont très perturbés par des artefacts et du bruit, ce qui complique considérablement l’interprétation clinique ou l’analyse automatique.Ce travail a pour objectif d’extraire des descripteurs des signaux d’EEG de surface qui peuvent conduire à la caractérisation de la dynamique spatio-temporelle des crises partielles du lobe temporal. Les estimateurs de relations inter-voies appliqués sont les méthodes linéaires paramétriques symétriques et non symétriques telles que l’inter-spectre (S), la cohérence (C), la Directed Transfert Function (DTF) ou la Partial Directed Coherence (PDC). Les relations sont estimées sur des EEG réels contenant une crise. La détection de fortes relations inter-voies est exploitée pour latéraliser puis caractériser la crise. Toutes les méthodes sont appliquées sur des signaux EEG bruts et prétraités. Une étape de prétraitement basée sur la séparation et classification de sources et le débruitage est mise en œuvre afin d’éliminer les artefacts et le bruit avec une perte minimale d’information en diminuant le risque de fausses détections de relations de connectivité inter-signaux. Les résultats obtenus sur 51 crises montrent que le prétraitement améliore la détection et le taux de bonnes latéralisations. Une méthode de couplage entre l’IS et les méthodes paramétriques directives (PDC et DTF) permet d’améliorer la caractérisation des crises / The electroencephalogram (EEG) is the essential clinical examination for the diagnosis, the definition of brain structures responsible of seizures and epilepsy classification. However, the signals collected on the surface of the scalp are very disturbed by artifacts and noise, which complicates the clinical interpretation or the automatic analysis. This work aims to extract descriptors of surface EEG signals that can lead to the spatio-temporal characterization of the temporal lobe seizures. The inter-channel relationship estimators applied are parametric linear methods, such as cross-spectrum (S), coherence (C), Directed Transfer Function (DTF) or Partial Directed Coherence (PDC). Relations are estimated on real EEG recordings containing a crisis. The detection of strong inter-channel relationships is exploited in order to lateralize and to characterize seizures. All methods are applied to raw and preprocessed EEG signals. A preprocessing step, based on the separation and classification of sources and denoising is implemented to remove artifacts and noise with a minimal loss of information by reducing the risk of false detections of inter-signal connectivity relationships. The results on 51 crises show that a signal preprocessing improves the detection and the rate of correct lateralization. A coupling method between S and directivity parametric methods (PDC and DTF) improves the characterization of crises

Problématique: Plus de 50 millions de personnes dans le monde ont de l’épilepsie, une condition neurologique caractérisée par des crises spontanées récidivantes. Malgré les nombreux médicaments anticrises disponibles, plus du tiers des patients continue à présenter des crises invalidantes, souvent accompagnées de blessures (dues à des chutes, fractures, brûlures, etc.) et parfois même de mort subite par arrêt cardio-respiratoire. La détection de crises avec des objets connectés a été un sujet de recherche en croissance au cours des dernières années. Cependant, les études se sont principalement concentrées sur la détection de crises généralisées tonico-cloniques plutôt que sur les crises focales. Question de recherche: Ce projet porte sur la détection de crises épileptiques de types focales avec altération de l’état de conscience (FIAS) via des signaux électrophysiologiques non invasifs captés par un chandail connecté (chandail intelligent Hexoskin) et en utilisant des méthodes d’apprentissage automatique. Méthodologie: Des patients avec épilepsie ont été recrutés à l’unité de monitoring d’épilepsie (UME) du CHUM pour porter le chandail Hexoskin. Celui-ci mesure trois différents signaux, soit l’électrocardiographie (ECG), l’accélérométrie et les signaux respiratoires. De ces signaux, 82 caractéristiques ont été extraites et des classificateurs d’apprentissage automatique ont été entraînés pour détecter le début des crises épileptiques. Résultats obtenus: Vingt-sept patients ont porté le chandail de manière continue pendant plus de 4 750 heures. 113 FIAS ont été enregistrées. Un modèle XGBoost avec une fenêtre de 10 sec et un chevauchement de 5 secondes ont donné les meilleurs résultats. En utilisant un seuil de 0,65, nous avons détecté 71 crises, ce qui a donné une sensibilité de 66 %, un temps en alerte de 15 % et un taux de fausses alertes (FAR) de 30 par 24 heures. Un seuil de 0,85 a permis de réduire les fausses alertes (8 % du temps en alerte, FAR de 21 par 24 heures), mais avec une sensibilité inférieure de 57 %. Les performances ont varié d'un patient à l'autre : la sensibilité était élevée et le FAR était faible pour certains patients, et inversement pour d'autres, indiquant une variabilité dans l'efficacité de l'algorithme selon les patients. Conclusions: Les résultats suggèrent qu’il est possible de détecter les FIAS avec le chandail Hexoskin mais que le taux de fausse alarme reste en général un peu trop élevé. / Problem Statement: More than 50 million people worldwide have epilepsy, a neurological condition characterized by recurrent, spontaneous seizures. Despite the availability of numerous antiseizure drugs, more than a third of patients continue to experience seizures, often accompanied by injuries (such as falls, fractures, burns) and, in some cases, sudden death due to cardiorespiratory arrest. Seizure detection using wearable technology has become a growing research area in recent years. However, most studies have focused on generalized tonic-clonic seizures rather than focal seizures. Research Question: This project investigates the use of machine learning methods applied to non-invasive physiological signals recorded by a connected shirt (the Hexoskin smart shirt) to detect focal impaired awareness seizures (FIAS). Methodology: Patients with epilepsy were recruited from the epilepsy monitoring unit (EMU) at the University of Montreal Hospital Center (CHUM) and wore the Hexoskin shirt, which measures three different signals: electrocardiography (ECG), accelerometry, and respiratory signals. From these signals, 82 features were extracted, and machine learning classifiers were trained to detect the onset of epileptic seizures. Results: A total of 113 FIAS from 27 patients were analyzed in this study. Patients wore the shirt continuously for over 4,750 hours. A XGBoost Model using a window of 10 seconds and overlap of 5 seconds gave the best results. Using a threshold of 0.65, we detected 71 seizures, yielding a sensitivity of 66%, a warning time of 15%, and a false alarm rate (FAR) of 30 per 24 hours. A threshold of 0.85 reduced false alarms (8% warning time, FAR of 21 per 24 hours) but also resulted in lower sensitivity (57%). The performance varied between patients: some had high sensitivity and low FAR, while others showed the opposite, indicating variability in the algorithm’s effectiveness across patients. Conclusions: The results suggest that, while the false alarm rate remains somewhat high, it is possible to detect FIAS. It was also observed that seizures in some patients were more easily detectable than in others, which may be explained by the significant variability in seizure semiology.