Multimodal AI for predicting comorbid REM sleep behavior disorder in major depressive disorder.

Abstract

REM sleep behavior disorder (RBD) is a robust prodromal marker of α-synucleinopathies: idiopathic RBD carries a 10-15-year phenoconversion risk of 80-90% to Parkinson's disease (PD) and related disorders. In major depressive disorder (MDD), comorbid RBD marks a subgroup at elevated prodromal PD risk, yet is frequently missed in psychiatric practice. Here, we developed a multimodal AI framework to detect comorbid RBD in MDD. From a clinical cohort of 329 patients, we obtained 261 video clips in 31 patients during reading and spontaneous-speech tasks, including 19 patients with MDD-RBD and 12 demographically and medication-matched MDD-only controls that, to our knowledge, formed the largest cohort of its kind worldwide. We used a dual-stream multimodal model that learned facial dynamics from video and vocal features from speech, and then combined both signals to predict comorbid RBD. In 5-fold cross-validation, our best model achieved 80.5% accuracy and 0.848 F1-score. Explainability analysis highlighted lower-face tension and variability, together with brow lowering, as candidate biomarkers requiring further validation. Predicted risk correlated with RBDQ score (r = 0.53, p = 0.005) and weakly with UPDRS motor score (r = 0.34, p = 0.067). Out-of-distribution evaluation showed broadly similar patterns, supporting the promise of multimodal AI for predicting RBD in MDD and identifying interpretable potential digital markers of prodromal synucleinopathy.