Multimodal Machine Learning–Based Prediction of Relapse in Major Depressive Disorder

Abstract

Importance Early identification of patients at high risk of relapse is key to personalized preventive strategies in major depressive disorder (MDD). However, current predictive models do not adequately account for the multimodal factors contributing to MDD relapse and fail to provide clinically meaningful predictions. Objective To predict relapse within 2 years in patients with MDD by using a comprehensive machine learning pipeline that integrates an unprecedented set of multimodal data. Design, Setting, and Participants This naturalistic longitudinal prognostic study used data from the Marburg-Münster Affective Disorders Cohort Study, with baseline assessments collected from October 2014 to May 2019 and follow-up approximately 2 years later, between October 2016 and June 2021. Data analysis was conducted between July 2024 and June 2026. The study was performed at 2 sites in Germany (Marburg and Münster), with participants recruited from psychiatric inpatient and outpatient departments and the general population. Patients meeting diagnostic criteria for current or lifetime MDD without history of neurological or severe somatic conditions were included. Exposure Natural course of illness in patients with MDD. Main Outcomes and Measures Balanced accuracy (BACC) for binary prediction of relapse vs no relapse during 2-year follow-up served as the primary outcome. Baseline features comprised 9 modalities including clinical and psychometric data, 4 structural and functional magnetic resonance imaging modalities, genetics, neurocognition, blood markers, and gut microbiome. Results A total of 564 patients with a current or lifetime MDD diagnosis were included (367 female [65.1%]; mean [SD] age, 37.34 [13.38] years). A total of 270 patients relapsed during the 2-year follow-up, while 294 did not. Mean (SD) BACC for the prediction of relapse ranged from 47.21% (5.85%) to 64.43% (6.63%). The best-performing model was solely based on few clinical relapse risk factors, while multimodal data integration did not improve performance. Conclusions and Relevance In this prognostic study, the models achieved moderate accuracy, yet the results suggest limited scope for guiding individual-level clinical decision-making. Predictive performance was largely driven by established clinical risk factors, indicating limited incremental value of multimodal measures when assessed at a single time point. Future advances may instead benefit from dynamic modeling approaches that more closely align temporally evolving multimodal markers with longitudinal clinical outcomes.