Non-contact REM/NREM sleep staging from piezoelectric signals using respiratory and body-movement features with auxiliary TWED-based respiratory stability measures

This is the authors' abstract. We don't add key points for this paper.

Introduction Non-contact sleep monitoring based on under-mattress piezoelectric sensing is attractive for low-burden home use, but REM/NREM discrimination remains challenging. This study aimed to investigate whether respiratory pattern stability, quantified by Time Warp Edit Distance (TWED)-based respiratory interval sequence (RIS) similarity features, could improve discrimination between the two major sleep states within sleep itself. Methods Overnight piezoelectric and polysomnography (PSG) data were collected simultaneously from 85 clinical subjects. PSG-labeled wake epochs were excluded, and the task was formulated as binary REM/NREM classification. From piezoelectric signals, we extracted conventional body-movement features, respiratory variability features, and multi-scale TWED-based RIS similarity features. Feature normalization was performed within each subject using the full-night unlabeled feature distribution, consistent with the intended post-hoc whole-night analysis scenario. An XGBoost classifier was evaluated using subject-wise nested leave-one-out cross-validation (LOOCV), with the REM-continuity fusion threshold selected within each training fold by inner subject-grouped 5-fold cross-validation. Respiratory signal extraction was additionally validated against PSG airflow in resting and full-night tests. Results Feature normalization improved performance across all feature sets. The best result was achieved by combining conventional body-movement and respiratory variability features with TWED-based RIS similarity features, yielding an accuracy of 84.39 ± 12.76%, Cohen's Kappa of 0.524 ± 0.241, REM precision of 0.600 ± 0.210, REM recall of 0.735 ± 0.226, and REM F1-score of 0.603 ± 0.211 under nested LOOCV. Compared with the normalized conventional feature set without RIS similarity, adding the TWED-based features improved both Kappa and REM F1-score. In respiratory tests, PVDF-derived respiration showed low detection error and good agreement with the airflow reference under both posture-change and overnight conditions. Discussion These findings indicate that TWED-based RIS similarity features provide useful complementary information beyond conventional descriptors and support the feasibility of using respiratory pattern stability derived from non-contact piezoelectric signals for within-sleep REM/NREM classification. At the current level of performance, the proposed method is better viewed as a low-burden adjunctive tool for offline whole-night longitudinal monitoring and trend assessment in home-like settings rather than as a replacement for PSG-based clinical diagnosis, real-time sleep staging, or full sleep-stage scoring.