Learning from multimodal data for classification and prediction of Alzheimer's disease.

Summary Alzheimer's disease (AD) is the leading cause of dementia in the world, affecting more than 20 million people. Its early diagnosis is essential to ensure adequate patient care and to develop and test new treatments. AD is a complex disease that requires different measures to be characterized: cognitive and clinical tests, neuroimaging, including magnetic resonance imaging (MRI) and positron emission tomography (PET), genotyping, etc. There is an interest in exploring the discriminatory and predictive capacities at an early stage of these different markers, which reflect different aspects of the disease and can provide complementary information. The objective of this PhD thesis was to evaluate the potential and integrate different modalities using statistical learning methods, in order to automatically classify AD patients and predict the evolution of the disease in its early stages. Specifically, we aimed to make progress towards a future application of these approaches to clinical practice. The thesis includes three main studies. The first one deals with the differential diagnosis between different forms of dementia based on MRI data. This study was performed using clinical routine data, which provided a more realistic assessment scenario. The second proposes a new framework for reproducible evaluation of AD classification algorithms from MRI and PET data. Indeed, although many approaches have been proposed in the literature for AD classification, they are difficult to compare and reproduce. The third part is devoted to the prediction of AD progression in patients with mild cognitive impairment by integrating multimodal data, including MRI, PET, clinical and cognitive assessments, and genotyping. In particular, we systematically assessed the added value of neuroimaging compared to clinical/cognitive data alone. Because neuroimaging is more expensive and less common, it is important to justify its use in classification algorithms.