Circuit-Specific Resting-State fMRI Signatures for Stratifying First-Episode Major Depressive Disorder and Predicting Recurrence Risk

BackgroundDepression is biologically heterogeneous, and first-episode depression (FED) carries a high risk of recurrence that is poorly captured by symptom-based assessment. Early identification of patients likely to relapse, as well as reliable identification of those unlikely to relapse, is needed to support personalized intervention and efficient allocation of care. MethodsWe developed a neurocomputational framework to infer recurrence risk from resting-state fMRI functional connectivity. The framework combines a Convolutional Filtering Autoencoder (CFAE) with a k-medoids clustering algorithm (FED-kMC) to derive neurofunctional FED subtypes, followed by a Manifold Sheaves-based Ensemble Support Vector Machine (MST-LVSVM) to estimate individual-level recurrence risk and define an interpretable decision boundary. Circuit-level analyses were then used to localize connectivity pathways associated with high relapse risk. ResultsThe framework identified two neurofunctionally distinct FED subtypes with divergent recurrence trajectories and achieved an external validation accuracy of 82.61% for recurrence risk prediction. Circuit analyses highlighted dysfunction within the Medial Superior Frontal Gyrus-Hippocampus and Angular Gyrus-Precuneus pathways as neural correlates of high relapse risk, together with a decision boundary enabling early-stage risk stratification. ConclusionsIntegrating connectome-derived, circuit-level information with subtype-aware machine learning may support proactive identification of FED patients at elevated recurrence risk and facilitate targeted early interventions, bridging connectome-level analysis and clinical decision-making.