Whether a multimodal machine learning pipeline integrating 9 data types (clinical/psychometric, structural and functional MRI, genetics, neurocognition, blood markers, gut microbiome) could predict 2-year relapse in 564 patients with current or lifetime MDD, using data from the Marburg-Münster Affective Disorders Cohort Study.
Balanced accuracy ranged from 47.21% to 64.43% across models; the best-performing model used only a few established clinical risk factors, and adding multimodal data did not improve predictive performance.
Single time-point baseline data limits ability to capture temporally evolving biomarkers; sample recruited from inpatient, outpatient, and general population settings in Germany, which may limit generalizability; models achieved only moderate accuracy, insufficient for individual-level clinical decisions.
Multimodal biomarker data (MRI, genetics, microbiome) did not add value over clinical risk factors alone for predicting MDD relapse at 2 years. Clinicians should continue relying on established clinical risk factors while awaiting dynamic, longitudinal modeling approaches.