Multimodal model adds little to clinical factors in MDD relapse prediction

Researchers analyzed data from the Marburg-Münster Affective Disorders Cohort Study, which recruited patients with current or lifetime MDD at two sites in Germany. The naturalistic prognostic study included 564 participants (mean age 37.3 years; 65.1% female), with baseline assessments collected from 2014 to 2019 and follow-up approximately two years later. Models combined up to nine data modalities, including clinical and psychometric measures, MRI, genetics, neurocognition, blood markers and gut microbiome data. The primary outcome was balanced accuracy for predicting relapse versus no relapse.

During follow-up, 270 participants relapsed and 294 did not. Across models, mean balanced accuracy ranged from 47.21% to 64.43%. The best-performing model relied on a few clinical relapse risk factors; integrating multimodal data did not improve performance.

The findings suggest limited value for using a single baseline assessment of multiple biological and imaging measures to guide individual relapse predictions. The study was observational, and its models achieved only moderate accuracy. The authors point to dynamic modeling of changing markers as a possible direction for future research; the results do not establish a clinically actionable prediction tool.