Corrélats cérébraux de l’anxiété catégorielle et dimensionnelle

Anxiety disorders are highly prevalent and represent a heavy burden of disease. However, despite a large group-level neuroimaging literature, robust brain markers of vulnerability or therapeutic response struggle to emerge. In recent years, individual-level prediction approaches using machine-learning have become increasingly popular in mental health research, and some promising results have been reported in small-scale neuroimaging studies (usually with Ntotal < 60 participants). These prediction results have yet to be replicated in larger, multisite samples. The present doctoral project involved the use of supervised machine-learning to prospectively predict the development of anxiety disorders in adolescents using a longitudinal dataset from the general population, IMAGEN, as well as to predict the response to psychotherapeutic treatment in phobic patients using the SPIDER-VR dataset. With IMAGEN data, machine-learning analyses were conducted using questionnaire and neuroanatomical data of non-anxious adolescents, to predict the development of a future anxiety disorder (N = 156) vs. healthy control status (N = 424). The study supported the predictive potential of sociodemographic and questionnaire data for the future onset of pooled anxiety disorders, and of gray matter volumes for future generalized anxiety disorder onset. Functional MRI metrics extracted from an emotional face processing task did not yield any above-chance level predictive performance. With SPIDER-VR data, machine learning analyses were conducted to predict the response of patients with spider phobia (N = 190) to a virtual reality exposure therapy session, using questionnaire data, structural MRI data, and various functional MRI metrics extracted from a symptom provocation task. Contrastingly to expectations, the study did not support the predictive potential of sociodemographic and questionnaire data nor neuroimaging data, with the exception of BOLD signal variance which yielded moderate predictive performance. Overall, this doctoral work challenges optimistic results from earlier smaller-scale neuroimaging- based machine-learning prediction studies in anxiety. Nonetheless, findings substantiate that easy-to-administer questionnaires show promising predictive performance for anxiety onset prediction and that structural MRI might bring incremental predictive value. Various other anxiety biomarkers have emerged in the literature with potential to improve the accuracy of anxiety-relevant predictions, and further multimodal research using large-scale datasets alongside rigorous machine- learning methodology are needed in an effort to reach clinical utility.