Seventy-three children participated in an investigation of inhibitory control (IC) at 3 years of age. Child IC was measured under various conditions in order to determine the impact that nonverbal and/or motivational task demands had on child IC task performance. Furthermore, task performance was examined with respect to measures of language, temperament, and psychophysiology. Tasks showed different patterns of relations to each of these variables. Furthermore, performance on the Hand Game, our measure of nonverbal IC, was explained by frontal EEG activity and, surprisingly, by language abilities. In contrast, performance on two other IC tasks, Day-Night and Less is More, was not related to measures of language or frontal EEG, perhaps because children performed at chance level on these tasks, indicating that these tasks may be too difficult for 3-year-old children. Implications of these findings are discussed. / Master of Science

No / The use of electroencephalograms (EEGs) to diagnose and analyses Alzheimer's disease (AD) has received much attention in recent years. The sample entropy (SE) has been widely applied to the diagnosis of AD. In our study, nine EEGs from 21 scalp electrodes in 3 AD patients and 9 EEGs from 3 age-matched controls are recorded. The calculations show that the kurtoses of the AD patients' EEG are positive and much higher than that of the controls. This finding encourages us to introduce a kurtosis-based de-noising method. The 21-electrode EEG is first decomposed using independent component analysis (ICA), and second sort them using their kurtoses in ascending order. Finally, the subspace of EEG signal using back projection of only the last five components is reconstructed. SE will be calculated after the above de-noising preprocess. The classifications show that this method can significantly improve the accuracy of SE-based diagnosis. The kurtosis analysis of EEG may contribute to increasing the understanding of brain dysfunction in AD in a statistical way.

Neurological and developmental disorders (NDDs) impose an enormous burden of disease on children throughout the world. Two of the most common are autism spectrum disorder (ASD) and epilepsy. ASD has recently been estimated to affect 1 in 68 children, making it the most common neurodevelopmental disorder in children. Epilepsy is also a spectrum disorder that follows a developmental trajectory, with an estimated prevalence of 1%, nearly as common as autism. ASD and epilepsy co-occur in approximately 30% of individuals with a primary diagnosis of either disorder. Although considered to be different disorders, the relatively high comorbidity suggests the possibility of common neuropathological mechanisms. Early interventions for NDDs lead to better long-term outcomes. But early intervention is predicated on early detection. Behavioral measures have thus far proven ineffective in detecting autism before about 18 months of age, in part because the behavioral repertoire of infants is so limited. Similarly, no methods for detecting emerging epilepsy before seizures begin are currently known. Because atypical brain development is likely to precede overt behavioral manifestations by months or even years, a critical developmental window for early intervention may be opened by the discovery of brain based biomarkers. Analysis of brain activity with EEG may be under-utilized for clinical applications, especially for neurodevelopment. The hypothesis investigated in this dissertation is that new methods of nonlinear signal analysis, together with methods from biomedical informatics, can extract information from EEG data that enables detection of atypical neurodevelopment. This is tested using data collected at Boston Children’s Hospital. Several results are presented. First, infants with a family history of ASD were found to have EEG features that may enable autism to be detected as early as 9 months. Second, significant EEG-based differences were found between children with absence epilepsy, ASD and control groups using short 30-second EEG segments. Comparison of control groups using different EEG equipment supported the claim that EEG features could be computed that were independent of equipment and lab conditions. Finally, the potential for this technology to help meet the clinical need for neurodevelopmental screening and monitoring in low-income regions of the world is discussed.

Throughout the 1990s ADHD diagnosis and medication rates have increased rapidly, and this trend continues today. These sharp increases have been met with both public and clinical criticism, detractors stating over-diagnosis is a problem and healthy children are being unnecessarily medicated and labeled as disabled. However, others say that ADHD is being under-diagnosed in some populations. Critics often state that there are multiple factors that introduce subjectivity into the diagnosis process, meaning that a final diagnosis may be influenced by more than the desire to protect a patient's wellbeing. Some of these factors include standardized testing, legislation affecting special education funding, and the diagnostic process. In an effort to circumvent these extraneous factors, this work aims to further develop a potential method of using EEG signals to accurately discriminate between ADHD and non-ADHD children using features that capture spectral and perhaps temporal information from evoked EEG signals. KNN has been shown in prior research to be an effective tool in discriminating between ADHD and non-ADHD, therefore several different KNN models are created using features derived in a variety of fashions. One takes into account the heterogeneity of ADHD, and another one seeks to exploit differences in executive functioning of ADHD and non-ADHD subjects. The results of this classification method vary widely depending on the sample used to train and test the KNN model. With unfiltered Dataset 1 data over the entire ANT1 period, the most accurate EEG channel pair achieved an overall vector classification accuracy of 94%, and the 5th percentile of classification confidence was 80%. These metrics suggest that using KNN of EEG signals taken during the ANT task would be a useful diagnosis tool. However, the most accurate channel pair for unfiltered Dataset 2 data achieved an overall accuracy of 65% and a 5th percentile of classification confidence of 17%. The same method that worked so well for Dataset 1 did not work well for Dataset 2, and no conclusive reason for this difference was identified, although several methods to remove possible sources of noise were used. Using target time linked intervals did appear to marginally improve results in both Dataset 1 and Dataset 2. However, the changes in accuracy of intervals relative to target presentation vary between Dataset 1 and Dataset 2. Separating subjects into heterogeneity classes does appear to result in good (up to 83%) classification accuracy for some classes, but results are poor (about 50%) for other heterogeneity classes. A much larger data set is necessary to determine whether or not the very positive results found with Dataset 1 extend to a wide population. / Master of Science / Throughout the 1990s ADHD diagnosis and medication rates have increased rapidly, and this trend continues today. These sharp increases have been met with both public and clinical criticism, detractors stating over-diagnosis is a problem and healthy children are being unnecessarily medicated and labeled as disabled. However, others say that ADHD is being underdiagnosed in some populations. Critics often state that there are multiple factors that introduce subjectivity into the diagnosis process, meaning that a final diagnosis may be influenced by more than the desire to protect a patient’s wellbeing. Some of these factors include standardized testing, legislation affecting special education funding, and the diagnostic process. In an effort to circumvent these extraneous factors, this work aims to further develop a potential method of using EEG signals to accurately discriminate between ADHD and non-ADHD children using features that capture spectral and perhaps temporal information from evoked EEG signals. KNN has been shown in prior research to be an effective tool in discriminating between ADHD and non-ADHD, therefore several different machine learning models are created using features derived in a variety of fashions. One takes into account the heterogeneity of ADHD, and another one seeks to exploit differences in executive functioning of ADHD and non-ADHD subjects. The results of this classification method vary widely depending on the sample used to train and test the KNN model, classification accuracy has ranged from 65% to 94%, and the cause for this variation was not identified. A much larger data set is necessary to determine whether or not the very positive results found with Dataset 1 extend to a wide population.

The study’s objective was to assess whether resting-state regional functional connectivity and current source density (CSD) measured during smoking abstinence predict smoking progression across 18 months, depressive traits, and nicotine-enhanced reward sensitivity (NERS) in young light-nicotine (NIC) smokers using low-resolution brain electromagnetic tomography analysis (LORETA). A secondary goal was to assess whether depressive traits moderate the ability of connectivity and regional CSD to predict NERS. Brain regions of interest (ROIs) hypothesized to predict smoking progression, NERS, and depressive traits include structures with high-density nicotinic acetylcholine receptors (nAChRs) and reward-related areas. A total of N=108, 14-hour NIC-deprived young (age 18-24) light (5-35 NIC uses/week) smokers underwent electroencephalogram (EEG) recording while at rest (i.e., viewed a white crosshair on a black background) for 8 minutes then completed the PRT, an assessment of reward sensitivity, after smoking a placebo (0.05 mg NIC) and NIC (0.8 mg NIC) cigarette using a within-subjects design allowing for the assessment of NIC-induced changes in reward sensitivity. All EEG power and LORETA activity bands underwent regression analysis to discover if EEG-assessed brain activity can predict smoking progression, depressive traits, NERS, and their potential interaction. Localized brain regions include 1) reward-related structures, 2) depressive trait-related structures, and 3) large-scale neural (e.g., salience network (SN), default mode network (DMN), executive control network (ECN)) and substance use disorder networks (e.g., orbital frontal cortex (OFC), insula, dorsal lateral prefrontal cortex (dlPFC) anterior cingulate cortex (ACC)). Weaker resting-state connectivity (rsC) between the insula and ACC (i.e., SN) predicted greater smoking progression at 18 months (theta1 and theta2) and greater depressive traits (delta and theta1), while greater rsC within the SN predicted greater NERS (alpha2 and beta 2/3[23.19 – 25.14 Hz]). Greater NERS was also predicted by greater alpha2 connectivity between the 1) ACC and posterior cingulate cortex (PCC) and 2) ACC and left dlPFC. Greater depressive traits were also predicted by 1) weaker delta and theta2 connectivity between the bilateral insula, 2) weaker delta, theta1, and theta2 between the insula and dlPFC, 3) weaker delta and theta1 between the insula and subgenual cortex, 4) greater theta2 in the right vs. left default mode, and 5) greater delta (2.44 – 3.41 Hz) in the left vs. right default mode network. Both greater depressive traits and greater NERS were predicted by weaker 1) theta2/alpha1 (6.59 – 9.52 Hz) between the insula and dlPFC and 2) alpha1 (7.5 – 9.5 Hz) between the left orbital frontal cortex and right dlPFC. These findings provide the first evidence that differences in EEG-assessed brain connectivity in young light smokers are associated with nicotine-enhanced reward sensitivity, depressive traits, and smoking progression. Notably, weaker low-frequency rsC within the salience network predicted depressive traits and smoking progression, while greater high-frequency rsC predicted greater nicotine-enhanced reward sensitivity. These findings suggest that salience network rsC and drug-enhanced reward sensitivity may be useful tools and potential endophenotypes for reward sensitivity and drug-dependence research.

As of 2016, diagnosis of ADHD in the US is controversial. Diagnosis of ADHD is based on subjective observations, and treatment is usually done through stimulants, which can have negative side-effects in the long term. Evidence shows that the probability of diagnosing a child with ADHD not only depends on the observations of parents, teachers, and behavioral scientists, but also on state-level special education policies. In light of these facts, unbiased, quantitative methods are needed for the diagnosis of ADHD. This problem has been tackled since the 1990s, and has resulted in methods that have not made it past the research stage and methods for which claimed performance could not be reproduced. This work proposes a combination of machine learning algorithms and signal processing techniques applied to EEG data in order to classify subjects with and without ADHD with high accuracy and confidence. More specifically, the K-nearest Neighbor algorithm and Gaussian-Mixture-Model-based Universal Background Models (GMM-UBM), along with autoregressive (AR) model features, are investigated and evaluated for the classification problem at hand. In this effort, classical KNN and GMM-UBM were also modified in order to account for uncertainty in diagnoses. Some of the major findings reported in this work include classification performance as high, if not higher, than those of the highest performing algorithms found in the literature. One of the major findings reported here is that activities that require attention help the discrimination of ADHD and Non-ADHD subjects. Mixing in EEG data from periods of rest or during eyes closed leads to loss of classification performance, to the point of approximating guessing when only resting EEG data is used. / Master of Science / As of 2016, diagnosis of ADHD in the US is controversial. Diagnosis of ADHD is based on subjective observations, and treatment is usually done through stimulants, which can have negative side-effects in the long term. Evidence shows that the probability of diagnosing a child with ADHD not only depends on the observations of parents, teachers, and behavioral scientists, but also on state-level special education policies. In light of these facts, unbiased, quantitative methods are needed for the diagnosis of ADHD. This problem has been tackled since the 1990s, and has resulted in methods that have not made it past the research stage and methods for which claimed performance could not be reproduced. This work proposes a combination of machine learning algorithms and signal processing techniques applied to EEG data in order to classify subjects with and without ADHD with high accuracy and confidence. Signal processing techniques are used to extract autoregressive (AR) coefficients, which contain information about brain activities and are used as “features”. Then, the features, extracted from datasets containing ADHD and Non-ADHD subjects, are used to create or train models that can classify subjects as either ADHD or Non-ADHD. Lastly, the models are tested using datasets that are different from the ones used in the previous stage, and performance is analyzed based on how many of the predicted labels (ADHD or Non-ADHD) match the expected labels. Some of the major findings reported in this work include classification performance as high, if not higher, than those of the highest performing algorithms found in the literature. One of the major findings reported here is that activities that require attention help the discrimination of ADHD and Non-ADHD subjects. Mixing in EEG data from periods of rest or during eyes closed leads to loss of classification performance, to the point of approximating guessing when only resting EEG data is used.

Proposé dans le cadre d’une triple collaboration entre le Centre de Recherche en Science et Technologie de l'Information et de la Communication (CReSTIC) de l'Université de Reims Champagne-Ardenne (URCA), le Laboratoire d'Ingénierie et Sciences des Matériaux (LISM) de l'Université de Reims Champagne-Ardenne (URCA) et le Laboratoire Electronique et Télécommunication (LET) de l'Université Mohammed 1er, ce projet a pour objectif d’associer les compétences de ces laboratoires, le traitement du signal pour le CReSTIC et le LET et la caractérisation de l’endommagement des agro matériaux composites pour le LISM. Le travail de cette thèse consiste à développer une méthode de Contrôle Non Destructif CND) par Emissions Acoustiques (EA). En effet, durant le processus de dégradation des matériaux composites (sollicitations mécaniques, vieillissement), plusieurs mécanismes d’endommagement à l’échelle microscopique peuvent intervenir selon la nature du composite et de ses constituants (fibres et résine). L’EA permet d’analyser et d’identifier plus en détails ces mécanismes d’endommagement, à l’aide d’une classification. Cette méthode de classification sera basée sur l’utilisation de méthodes à noyaux (typiquement séparateur à vastes marges) dans le domaine temps-fréquence. Il s’agira de déterminer un (ou des) noyau(x) adapté(s) au problème posé, basé sur une mesure de similarité entre les signaux. Ce travail permettra ainsi d’analyser et classifier les mécanismes d’endommagement sans emploi de descripteurs. Cette nouvelle méthode de CND par EA permettra de fournir des informations pertinentes et de vérifier efficacement la fiabilité et l'état de santé en temps réel de structures en service, sans perturber l'exploitation tout en réduisant les coûts de maintenance. / The aims of this thesis consists in developing a method of Non-destructive testing (NDT) by Acoustic Emissions (AE). Indeed, during the process of degradation of composite materials several mechanisms of damage in the microscopic scale can intervene according to the nature of the composite and its constituents (fibers and epoxy). AE allows to analyze and to identify more in detail these damages mechanism, by means of a classification. This method of classification will be based on the use of kernel methods in the time - frequency domain. It will be a question of determining one adapted kernel in the proposed problem. This work will so allow to analyze and to classify mechanisms without using descriptors.

In dieser Arbeit untersuche ich der Gesichterkognition zugrundeliegende biologischen Mechanismen auf der genetischen, neuronalen und verhaltensbasierten Ebene. Die neuronale Aktivität wurde mittels ereigniskorrelierter Potenziale (EKPs) untersucht und ihre Latzenzvariabilität innerhalb der Person wurde durch eine innovative Methode, Residue Iteration Decomposition (RIDE), gemessen. Die erste Studie demonstriert die Reliabilität von RIDE für die Extraktion von Einzeltrialparametern der P3b Komponente, welche in der zweiten Studie die Basis für die Untersuchung der Innen-Subjekt-Variabilität (ISV) bei der Geschwindigkeit der Gesichterkognition bildet. Die zweite Studie untersucht individuelle Unterschiede in ISV in ihrer genetischen Variation, gemessen an der Verhaltens- und neuronalen Ebene während einer Gesichterkognitionsaufgabe. Die Ergebnisse zeigen, dass ISV nicht nur mit dem COMT Val158Met Polymorphismus zusammenhängt, sondern auch von der geforderten kognitiven Verarbeitung abhängt. Zudem ist die ISV in der Reaktionszeit teilweise durch die ISV in der Geschwindigkeit zentralkognitiver Prozesse erklärbar. Studie 3 liefert neuartige Informationen für die N1/N170 Forschung. Mit einem differentialpsychologischen Ansatz konnten wir nicht nur vorangegangene Ergebnisse zur Vorhersagekraft der N170 für individuelle Unterschiede in der Gesichterkognition replizieren, sondern auch die individuellen Unterschiede in der N170 in einen allgemeinen und einen gesichtsspezifischen Teil mit unterschiedlicher Vorhersagekraft zerlegen. Darüber hinaus konnten wir zeigen, dass top-down Modulationen der N170 unterscheidbare und qualitativ unterschiedliche Beziehungen zu Fähigkeiten der Gesichterkognition aufweisen. Insgesamt zeigen die integrierten Ergebnisse der Studien meiner Dissertation die psychologische Bedeutsamkeit der intra- und interindividuellen Variabilität in der Gesichterkognition für die Erforschung der ihr zugrundeliegenden biologischen Mechanismen. / The biological mechanisms underlying face cognition from an inter- and intra-individual variability perspective at the genetic, neural, and behavioral levels are investigated. The neural activities related to face processing are measured by event-related potentials (ERPs) and their trial-by-trial latency variability are estimated using a novel and well-established method, Residue Iteration Decomposition (RIDE). Study 1 demonstrates the reliability of RIDE in extracting single-trial parameters of the P3b component. In the Study 2, individual differences in ISV of face processing speed, measured at both behavioral and neural levels during a face processing task, are studied in their genetic variation. The results suggest that individual differences in ISV are related not only to the COMT Val158Met polymorphism, but also to the type of cognitive processing (e.g., memory domain). Moreover, we showed that ISV in reaction time can be partially explained by ISV in the speed of central cognitive processes. Furthermore, the individual differences approach in Study 3, provided valuable and novel information beyond the common group-mean approach applied in the N1/N170-related research. Based on this approach, not only we could replicate previous findings that the N170 predicts individual differences in face cognition abilities, but also we could decompose individual differences in the N170 into a domain-general and a face-specific part with different predictive powers. Moreover, we showed that top-down modulations on the N170 have separable and qualitatively different relationships to face cognition abilities. In summary, the integrated results from different studies in my dissertation demonstrate the psychological importance of the information provided by inter- and intra-individual variability in face processing in the investigation of its underlying biological mechanisms.

Un traitement possible de l'épilepsie partielle pharmaco-résistante consiste en l'exérèse de la région cérébrale responsable des crises. La difficulté est de localiser cette zone et d'en définir l'étendue. L'objectif de cette thèse est d'apporter des données permettant de préciser la localisation et le volume des régions pathologiques, en exploitant deux modalités : l'imagerie TEP (Tomographie par Emission de Positons) et l'EEGHR (EEG Haute Résolution : signaux cérébraux recueillis sur le scalp avec un nombre important d'électrodes et une fréquence d'échantillonnage élevée). En imagerie TEP, il s'agit de segmenter les zones d'hypométabolisme qui sont liées aux régions responsables des crises. Des méthodes de comparaisons statistiques à l'aide d'outils de type SPM (Statistical Parametric Mapping) entre les images TEP de sujets pathologiques et de sujets sains ont été appliquées, en effectuant des tests d'hypothèse voxel à voxel entre les différentes images. Pour pouvoir être comparées à une population de référence, les différentes images ont subi des transformations non linéaires afin que chaque voxel corresponde à la même région anatomique chez tous les sujets. Deux algorithmes ont été appliqués : une méthode SPM classique et une méthode Block-Matching. Les résultats sont comparés par analyse subjective clinique et également sur des données simulées. En ce qui concerne l'EEG-HR, la localisation spatiale et temporelle de sources d'événements intercritiques (pointes et ondes lentes) a été réalisée par résolution du problème inverse. Ceci a permis de localiser les sources électriques intracérébrales d'intérêt qui sont à l'origine des signaux acquis sur le scalp. Enfin, une représentation des données multimodales (images TEP et signaux EEG-HR) dans un même référentiel a permis d'accroître les connaissances sur les relations existant entre les activités électriques et métaboliques et ainsi de mieux définir les régions épileptogènes / A possible treatment for drug-resistant partial epilepsy involves the resection of the brain region which generates crisis. The difficulty is to locate this area and to determine its extent. The objective of this thesis is to provide data to specify the location and the volume of pathological regions, using two modalities: PET (Positron Emission Tomography) imaging and HR-EEG (High-Resolution EEG: brain signals collected on the scalp with a large number of electrodes and a high sampling rate). In PET imaging, hypometabolic areas associated with regions generating seizures have to be segmented. Statistical comparisons methods using tools such as SPM (Statistical Parametric Mapping) between images of pathological and healthy subjects have been applied. Voxelwise statistical analyses between the different images were thus used to highlight the hypometabolic areas. For comparison with a reference population, nonlinear transformations were applied to the images so that each voxel corresponds to the same anatomical region in every subject. Two algorithms were applied: a conventional SPM method and a Block-Matching method. The results were compared by subjective clinical analysis and also on simulated data. Regarding the HR-EEG, the spatial and temporal source localizations of interictal events (spikes and slow waveforms) were done by solving the inverse problem. This allowed to localize intracerebral electrical sources generating the signals acquired on the scalp. Finally, a representation in the same space of multimodal data (PET images and HR-EEG) allowed to increase the knowledge on the relationship between electrical and metabolic activities and to better define the epileptogenic regions

L’imagerie de source corticale joue un rôle important pour la compréhension fonctionnelle ou pathologique du cerveau. Elle permet d'estimer l'activation de certaines zones corticales en réponse à un stimulus cognitif donné et elle est également utile pour identifier la localisation des activités pathologiques, qui sont les premières étapes de l'étude des activations de réseaux neuronaux sous-jacents. Diverses méthodes d'investigation clinique peuvent être utilisées, des modalités d'imagerie (TEP, IRM) et magnéto-électroencéphalographie (EEG, SEEG, MEG). Nous souhaitions résoudre le problème à partir de données non invasives : les mesures de l'EEG de scalp, elle procure une résolution temporelle à la hauteur des processus étudiés Cependant, la localisation des sources activées à partir d'enregistrements EEG reste une tâche extrêmement difficile en raison de la faible résolution spatiale. Pour ces raisons, nous avons restreint les objectifs de cette thèse à la reconstruction de cartes d’activation des sources corticales de surface. Différentes approches ont été explorées. Les méthodes les plus simples d'imagerie corticales sont basées uniquement sur les caractéristiques géométriques de la tête. La charge de calcul est considérablement réduite et les modèles utilisés sont faciles à mettre en œuvre. Toutefois, ces approches ne fournissent pas d'informations précises sur les générateurs neuronaux et sur leurs propriétés spatiotemporelles. Pour surmonter ces limitations, des techniques plus sophistiquées peuvent être utilisées pour construire un modèle de propagation réaliste, et donc d'atteindre une meilleure reconstruction de sources. Cependant, le problème inverse est sévèrement mal posé, et les contraintes doivent être imposées pour réduire l'espace des solutions. En l'absence de modèle bioanatomique, les méthodes développées sont fondées sur des considérations géométriques de la tête ainsi que la propagation physiologique des sources. Les opérateurs matriciels de rang plein sont appliqués sur les données, de manière similaire à celle effectuée par les méthodes de surface laplacien, et sont basés sur l'hypothèse que les données de surface peuvent être expliquées par un mélange de fonctions de bases radiales linéaires produites par les sources sous-jacentes. Dans la deuxième partie de ces travaux, nous détendons la contrainte-de rang plein en adoptant un modèle de dipôles distribués sur la surface corticale. L'inversion est alors contrainte par une hypothèse de parcimonie, basée sur l'hypothèse physiologique que seuls quelques sources corticales sont simultanément actives ce qui est particulièrement valable dans le contexte des sources d'épilepsie ou dans le cas de tâches cognitives. Pour appliquer cette régularisation, nous considérons simultanément les deux domaines spatiaux et temporels. Nous proposons deux dictionnaires combinés d’atomes spatio-temporels, le premier basé sur une analyse en composantes principales des données, la seconde à l'aide d'une décomposition en ondelettes, plus robuste vis-à-vis du bruit et bien adaptée à la nature non-stationnaire de ces données électrophysiologiques. Toutes les méthodes proposées ont été testées sur des données simulées et comparées aux approches classiques de la littérature. Les performances obtenues sont satisfaisantes et montrent une bonne robustesse vis-à-vis du bruit. Nous avons également validé notre approche sur des données réelles telles que des pointes intercritiques de patients épileptiques expertisées par les neurologues de l'hôpital universitaire de Nancy affiliées au projet. Les localisations estimées sont validées par l'identification de la zone épileptogène obtenue par l'exploration intracérébrale à partir de mesures stéréo EEG. / Cortical Source Imaging plays an important role for understanding the functional and pathological brain mechanisms. It links the activation of certain cortical areas in response to a given cognitive stimulus, and allows one to study the co-activation of the underlying functional networks. Among the available acquisition modality, electroencephalographic measurements (EEG) have the great advantage of providing a time resolution of the order of the millisecond, at the scale of the dynamic of the studied process, while being a non-invasive technique often used in clinical routine. However the identification of the activated sources from EEG recordings remains an extremely difficult task because of the low spatial resolution this modality provides, of the strong filtering effect of the cranial bones and errors inherent to the used propagation model. In this work different approaches for the estimation of cortical activity from surface EEG have been explored. The simplest cortical imaging methods are based only on the geometrical characteristics of the head. The computational load is greatly reduced and the used models are easy to implement. However, such approaches do not provide accurate information about the neural generators and on their spatiotemporal properties. To overcome such limitations, more sophisticated techniques can be used to build a realistic propagation model, and thus to reach better source reconstruction by its inversion. However, such inversion problem is severely ill-posed, and constraints have to be imposed to reduce the solution space. We began by reconsidering the cortical source imaging problem by relying mostly on the observations provided by the EEG measurements, when no anatomical modeling is available. The developed methods are based on simple but universal considerations about the head geometry as well as the physiological propagation of the sources. Full-rank matrix operators are applied on the data, similarly as done by Surface Laplacian methods, and are based on the assumption that the surface can be explained by a mixture of linear radial basis functions produced by the underlying sources. In the second part of the thesis, we relax the full-rank constraint by adopting a distributed dipole model constellating the cortical surface. The inversion is constrained by an hypothesis of sparsity, based on the physiological assumption that only a few cortical sources are active simultaneously Such hypothesis is particularly valid in the context of epileptic sources or in the case of cognitive tasks. To apply this regularization, we consider simultaneously both spatial and temporal domains. We propose two combined dictionaries of spatio-temporal atoms, the first based on a principal components analysis of the data, the second using a wavelet decomposition, more robust to noise and well suited to the non-stationary nature of these electrophysiological data. All of the proposed methods have been tested on simulated data and compared to conventional approaches of the literature. The obtained performances are satisfactory and show good robustness to the addition of noise. We have also validated our approach on real epileptic data provided by neurologists of the University Hospital of Nancy affiliated to the project. The estimated locations are consistent with the epileptogenic zone identification obtained by intracerebral exploration based on Stereo-EEG measurements.