Abstract Introduction REM sleep behavior disorder (RBD) poses a high risk of injury to patients and bed partners, yet symptomatic treatments are limited and clinical trials lack objective, quantifiable endpoints. We evaluated whether AI-driven computer-vision models could automatically detect injury-risk behaviors in a large, well-characterized cohort of patients with isolated RBD undergoing video-polysomnography (vPSG). Methods Motor events during REM sleep were automatically detected by optical flow on vPSG from 69 subjects (79.4% male; mean age 66.4 years) with definitive RBD, and segmented in discrete movement clips separated by ≥1 second of immobility. All clips were manually rated by two independent reviewers using the International RBD Study Group scale classifying behaviors as mild (no risk), moderate (mild risk) or severe (significant risk of injury). A third reviewer adjudicated clips with discordant ratings (14.9% of clips). Given the relative rarity of moderate and severe events, these two categories were merged into a single “injury-risk” class. We used two model types to classify injury risk: (1) heuristic models based on 4 explainable features: magnitude, velocity, complexity (angular variation) of movements, and lateral displacement; (2) a deep visual model (V-JEPA) in which video frames were fed directly as input; adopting a 80/10/10% split for training, validation, and testing with severity-class distribution preserved across each set. Results A total of 3,329 movement clips were manually rated as mild, 265 clips as moderate, and 19 as severe (total of 284 injury-risk movements). Best performance was achieved with V-JEPA and a vision transformer model, with 93% accuracy and a macro F1 score of 0.81, while feature-based random forest achieved 85% accuracy and macro F1 of 0.64. Movements classified as injury-risk were associated with greater magnitude, velocity, complexity, and lateral displacement toward the edge of the bed. Model performance with environmental variations in camera setup, lighting conditions, and room configuration remained stable in the feature-based heuristics model but declined in the vision-only model. Conclusion This work illustrates the feasibility of using computer-vision models to automatically quantify RBD injury-risk. Future studies should evaluate the extent to which such models generalize to various clinical and home environments. Support (if any)
Raval et al. (Fri,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: