Abstract
Traditional sleep staging using contact sensors may compromise data validity. This study proposes a non-contact ballistocardiogram (BCG)-based method to improve both accuracy and comfort in sleep monitoring. BCG signals were processed using continuous wavelet transform and low-pass filtering to extract heart rate variability (HRV) and respiratory rate variability (RRV). A novel feature selection model integrating attention mechanisms with XGBoost was developed, where attention weights are used to prioritize features before iterative refinement by XGBoost. Evaluated on 10,201 sleep segments, the Fast-ABC Boost model achieved an accuracy of 89.85%, along with superior precision, recall, F1-score, and Kappa values compared to conventional methods. The attention-XGBoost fusion effectively mitigates interference from noisy and redundant features while optimizing feature relevance, demonstrating robust adaptability to the complexity of sleep signals. This innovation advances the accuracy non-contact sleep staging, enabling practical applications in home healthcare and personalized sleep management, while improving user comfort.