基于脑电的 REM 睡眠行为障碍自动睡眠分期：MSIF-Net 逐帧验证

Objective REM-stage abnormalities are central to the pathophysiology of REM sleep behavior disorder (RBD) and are clinically relevant to symptom burden. We developed an EEG-based, pathology-oriented automated sleep-staging framework for RBD and tested whether model-derived REM architecture metrics are reliable and clinically informative in real-world polysomnography (PSG). Methods The Multi-Stream Imaging Fusion Network (MSIF-Net) integrates raw EEG waveforms (1D CNN), time-frequency spectrograms (2D CNN), and 65 handcrafted descriptors via attention-based fusion. Stage 1 used patient-wise five-fold cross-validation for epoch-level Wake/NREM/REM staging. Stage 2 applied the fixed model to an independent clinical RBD cohort to estimate whole-night NREM% and REM% (of TST), evaluate agreement with routine manual summaries (Bland-Altman), and test associations between automated REM% and symptom scales (Spearman, FDR-corrected). Results MSIF-Net achieved class-wise F1 scores of 0.84 for Wake, 0.94 for NREM, and 0.80 for REM, with errors mainly reflecting REM-wake confusions. Whole-night REM% and NREM% showed close agreement with manual summaries. Automated REM% correlated inversely with PSQI ( ρ = -0.507, q ρ = -0.454, q = 0.002), but not ESS ( ρ = -0.102, q = 0.667). Conclusion MSIF-Net enables EEG-only three-class sleep staging in RBD and yields clinically consistent whole-night REM architecture estimates that capture clinically meaningful variation in sleep complaints and RBD symptom burden.