Redefining and subtyping of major depression based on brain functional connectivity signatures with high generalization: an ensemble hybrid framework

Abstract

Current clinical diagnosis of major depression is solely based on subjective symptoms and signs rather than the underlying biological mechanisms, which has hindered the improvement in treatment effectiveness and outcomes. Redefining and subtyping major depressive disorder (MDD) based on neural circuits has been an emerging consensus. Here, we proposed a novel ensemble hybrid deep-learning framework based on graph neural network for redefining and subtyping of MDD with the brains "fingerprinting", i.e. nuanced functional connectivity profiles. A graph neural network (GNN) based supervised model was first employed to discriminate MDD patients from healthy controls. The graph embedding features obtained from the supervised classifier was then employed for subtyping of MDD with an unsupervised clustering by fast search and find of density peaks (CFDP) approach. The generalization accuracy and reproducibility of the proposed framework were further validated with one of the largest resting-state fMRI data set for MDD which included 2428 participants recruited from 25 independent sites. Our classifier showed a high generalization of 75% for leave-one-site-out cross validation. In addition, we obtained three biologically homogenous MDD subtypes characterized by impairments in subcortical-limbic, DMN-FPN, VN-SMN circuits, respectively. Notably, we demonstrated that subtyping of MDD according to graph embedding features showed high robustness and reproducibility across different sites. The proposed framework is not only essential for redefining and subtyping of MDD, but also paves the way towards precision medicine of mental disorders.