Interpretable Sleep Staging Using a Portable EEG-PPG Forehead Patch with SHAP-Guided Clustering

Sleep staging plays a critical role in both clinical diagnosis and home-based sleep health monitoring. Although traditional PSG provides reliable assessments, it remains costly and cumbersome. Wearable devices offer a scalable alternative, but most existing solutions prioritize accuracy while overlooking interpretability, which is a key requirement for clinical trust and deployment. This study proposes an explainable machine learning pipeline using a portable forehead patch device that records frontal EEG (two channels) and PPG signals. We analyzed 27 full-night recordings with synchronized PSG-based annotations using multiple classifiers under a leave-one-out cross-validation scheme, where CatBoost achieved the best performance with a macro-F1 score of 55.0 % and Cohen's kappa of 54.3%. To interpret model decisions, we applied SHAP for feature attribution, projected decision patterns into a twodimensional space, and identified class-specific clusters. Symbolic rules extracted from these groups revealed stagespecific patterns, such as low alpha/theta ratio in N3 and elevated fractal dimensionality in REM, aligned with known physiology. These findings support the use of compact EEGPPG systems for interpretable sleep monitoring at home and in resource-limited environments.