Identification et caractérisation de profils de connectivité fonctionnelle à partir de l'EEG dans les troubles neurocognitifs modérés et majeurs

As the global population continues to age, the prevalence of neurocognitive disorders (NCD) is increasing and is expected to continue. Therefore, the need for effective and efficient diagnostic methods is growing. However, many major NCDs, such as Alzheimer's Disease and Vascular Dementia share common clinical symptoms. Furthermore, in the large majority of the cases, patients display more than one disease that cause their neurocognitive symptoms. These facts make the differential diagnosis challenging. Neuroimaging tools and biomarker sampling methods are already used along with neuropsychological tests to diagnose the patients. Nevertheless, these methods are costly, invasive and do not reflect the dynamical aspect of the brain activity. Electroencephalography (EEG), on the contrary, is cheap, making it a credible alternative for low- and medium-income countries, portable, enabling its use in medical deserts, and minimally invasive. Its high temporal resolution also allows to follow brain dynamics, which enables the estimation of functional connectivity. Functional connectivity has been widely used in the literature for the distinction between cognitive conditions. However, previous studies generally focus on the distinction between controls and diseased patients, and more rarely between specific disorders. Also, they consider the labels given by a clinical diagnosis, which are sometimes inaccurate or incomplete. This thesis aims at contributing to bridge this gap, by proposing an unsupervised analysis framework based on the analysis of functional connectivity estimated with Dynamic Time Warping distance. By removing dependence on diagnosis labels during model training, we aim at reducing bias and gain better understanding of the underlying brain dynamics. In this thesis, we analyzed a wide range of NCDs to gain insights into the differences between them. We also explored functional connectivity in patients with Mixed Dementia to understand the interaction between co-occurring diseases. We also compared different mathematical techniques across the proposed analysis pipeline to optimize the pertinence of the resulting connectivity profiles and enhance their interpretability, ultimately contributing to more robust label-independent diagnostic support.