Objective: To uncover sleep physiological signatures that distinguish Parkinson’s disease (PD) and REM sleep behavior disorder (RBD) by reverse-engineering what a deep learning classifier learns from overnight polysomnography (PSG), then dissecting its learned representations to reveal disease-specific physiological patterns.

Background: PD and RBD share alpha-synuclein pathology, with RBD recognized as a prodromal marker for PD. Conventional PSG analyses rely on predefined features examined in isolation, and deep learning models, while promising, remain largely opaque in what they learn.

Method: We analyzed PSG from 411 individuals (healthy controls, isolated RBD, PD without RBD, PD with RBD) across five centers. A two-stage architecture first learned epoch-level embeddings via supervised contrastive learning (EEG, EOG, chin EMG), then applied attention-weighted pooling to produce subject-level continuous scores (PD-ness, RBD-ness). Post-hoc interpretability analysis characterized what these scores capture physiologically.

Results: The classifier achieved adequate discrimination (PD ROC-AUC 0.787, RBD 0.757; external validation 0.770 and 0.726), enabling post-hoc dissection of its learned representations. PD-ness and RBD-ness were functionally dissociated (R2 = 0.020), with residual correlation only in transitional groups, consistent with the dual-axis structure of alpha-synuclein propagation. PD-ness correlated with known PD sleep EEG changes (decreased delta power at central channels), validating its physiological grounding, and further revealed co-occurring decreases in DFA with increases in sample entropy, jointly characterizing multi-scale temporal disruption not evident from single-feature analyses. Notably, while conventional EEG features showed the strongest PD correlations in REM, the classifier’s attention preferentially targeted NREM stages, suggesting that NREM contains PD-relevant information beyond what conventional features capture.

Conclusion: Reverse-engineering a PSG-based deep learning classifier reveals that PD and RBD map onto dissociable physiological axes, and that the model captures PD-related neural dynamics beyond conventional EEG features, pointing to NREM microstructure as a target for future investigation.

« Back to 2026 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/uncovering-sleep-neural-signatures-of-parkinsons-disease-and-rem-sleep-behavior-disorder-from-polysomnography-through-interpretable-deep-learning/