Global Deep Learning Initiative to Understand Outcomes in Major Depression

ABSTRACT Major depressive disorder (MDD) is the leading cause of disability worldwide, and around half of MDD patients have treatment-resistant depression. The use and clinical benefit of rTMS have escalated greatly in recent years. As only 40-50% of patients respond to current standard rTMS, there is great interest in predicting which patients are likely to respond, what brain features best predict response, and how these features relate to the core biosignatures of MDD. To address this, and responding to NIH’s call for Precision Medicine approaches, our Global Deep Learning Initiative to Understand Outcomes in Major Depression unites international leaders in MDD and rTMS research, neuroimaging, and AI to identify generalizable predictors of rTMS response, and assess how they relate to brain biomarkers of MDD. Two major innovations are proposed. First, we use novel deep learning methods, based on convolutional neural networks, to extract predictive features from multimodal brain images (sMRI, DTI, and rsfMRI); tactics applied in whole-brain and surface-based mapping of brain function and structure, DVAEs for feature extraction, and transfer learning (to learn from auxiliary datasets and tasks) will distill predictive features while protecting individual privacy. CNNs trained on multimodal brain maps for our predictive tasks will distill additional layers of information that have not yet been fully exploited in MDD research, to better predict clinical status and treatment response. Second, our worldwide ENIGMA-MDD network will provide diverse test data from globally representative populations, to ensure that our predictive models do not break down when tested on diverse data. ENIGMA’s harmonized extraction of brain measures across worldwide cohorts will enhance rigor and ensure that analyses are well-powered and consistently performed. We include an important partnership with REST-meta-MDD, a Chinese consortium collecting multimodal imaging data from patients with MDD, to test the generalizability of our predictive models. The likely outcome of our work is a set of pre-screening tools to predict who will respond best to rTMS, and a deeper understanding of the brain signatures of MDD that predict treatment outcomes following rTMS. All tools will be made public via NITRC and ENIGMA websites, and will be tested across our ENIGMA network, guaranteeing impact of the work for large- scale outcome prediction within and outside of MDD research.