Transfer Learning from Big data to Small Data: Leveraging Psychiatric Neuroimaging Biomarkers Discovery – Big2small

Unlike many other medical specialties, psychiatry lacks objective quantitative measures (such as blood dosage) to guide clinicians in choosing a therapeutic strategy. Brain anatomy is an imprint of the individual's genetic and environmental background. The identification of prognostic brain signatures of clinical course or response to treatment would pave the way for personalized medicine in psychiatry. Many international initiatives have aggregated large datasets (>10K subjects for the general population, ~2K for patients/healthy controls cohorts). However, their large heterogeneity and cross-sectional designs prevent us from learning predictors of individual patient outcome (response to treatment, clinical evolution). Other initiatives have recently resulted in smaller, more clinically homogeneous databases (N<500) including longitudinal follow-ups to assess response to treatment and transition to disease in patients at risk. The high cost per patient (>10K€), however, limits the feasibility to scale them up to an adequate sample size (at least a few thousand) necessary to build predictive models that are sufficiently reproducible for regular clinical application.
This project proposes 3 transfer learning strategies (deep neural networks, clustering, dimensional approach of the continuum of psychiatric pathologies) to reconcile Big and small data. These 3 strategies are divided into 3 steps (i) modeling general brain variability on large bases from the general population; (ii) transfer (fine-tuning, etc.) on (medium) case-control bases to focus models on a specific pathology; (iii) final transfer on “small” longitudinal cohorts to allow learning of prognostic models of clinical evolution or response to treatments. The success of this project would demonstrate that AI could benefit health care in mental illness, which is the world‘s leading cause of disability and direct and indirect costs.