Fully-automated sleep staging for Parkinson's disease and isolated REM sleep behavior disorder.

Abstract

Isolated REM sleep behavior disorder (iRBD) is a key prodromal marker of Parkinson's disease (PD). Video-polysomnography (vPSG) remains the diagnostic gold standard, but manual sleep staging is particularly time-consuming and challenging in neurodegenerative disease. We adapted U-Sleep, a deep neural network, for automated sleep staging in PD and iRBD. A pretrained model (PUB, 19,236 PSGs), was finetuned on multicenter datasets (PACE, CBC: 112 PD, 138 iRBD, 89 controls) and evaluated on a clinical hold-out (DCSM: 81 PD, 36 iRBD, 87 controls). Predictors of staging agreement were analyzed, and low-agreement recordings were blindly rescored. Confidence-based thresholds were applied to enhance REM detection. The pretrained model achieved κ = 0.66 in PACE/CBC, improving to κ = 0.74 after finetuning (p < 0.001). In the hold-out, mean κ increased from 0.60 to 0.64 (p < 0.001). Site-specific finetuning provided minimal benefit. Confidence was a significant predictor of Cohen's κ (p < 0.001). Recordings with low model agreement also showed low human interrater agreement. Applying a confidence threshold increased REM precision from 85 to 95.6%, preserving sufficient REM sleep in 96% of subjects. This publicly available model achieves human-level agreement enabling scalable, standardized PSG analysis with model-derived confidence as a tool for further refinements.