A Novel K-Means with SHAP Feature Selection and ROA-Optimized SVM for Sleep Monitoring from Ballistocardiogram Signals

Sleep quality is closely associated with cardiovascular, metabolic, and mental health outcomes, yet the clinical gold standard, polysomnography (PSG), is costly and intrusive for long-term home monitoring. Ballistocardiography (BCG) enables unobtrusive in-bed sensing and is therefore attractive for low-burden sleep assessment in natural environments. However, most existing BCG studies are PSG-referenced and mainly focus on sleep staging, while movement and out-of-bed episodes are often treated as artifacts rather than modeled jointly. In this study, we propose an interpretable unsupervised proxy-state modeling framework for three-state in-bed monitoring from BCG signals under an unlabeled setting. BCG recordings were segmented into 30 s windows with 50% overlap, and multi-domain features were extracted from waveform morphology, spectral power, heart rate-related dynamics, and wavelet energy distribution. K-means clustering (K = 3) was used to construct cluster-derived proxy labels, TreeSHAP-based feature ranking together with inner-CV-guided Top-N subset selection was used for training-only feature screening, and multiple classifiers were compared under a strict leave-one-subject-out protocol, with an ROA-optimized RBF-SVM achieving the best overall performance. Using data from 32 volunteers, the framework achieved an accuracy of 0.9932 ± 0.0047 (mean ± SD), together with consistently strong Macro-F1 and MCC scores. Overall, it outperformed the alternative methods compared in this study.