Multi-window HRV and temporal context for sleep staging from ECG and smartwatch PPG-derived intervals

Sleep staging is essential for evaluating sleep quality, yet conventional polysomnography is costly, intrusive, and difficult to deploy in everyday settings. This study presents an HRV-centered sleep-staging framework for ECG-derived RR intervals and smartwatch PPG-derived IBI sequences. Rather than introducing a new Transformer architecture, the study systematically evaluates multi-window HRV representation and sleep-progress context for four-class sleep staging, while examining demographic information and feature-interaction strategies as auxiliary components. The framework extracts time-domain, frequency-domain, and nonlinear HRV features from multiple temporal windows, combines them primarily with sleep-progress context while optionally incorporating demographic context, and uses a lightweight Transformer encoder as a temporal integration module for classifying Wake, Light sleep, Deep sleep, and REM sleep. In the final baseline comparison, the proposed configuration achieved accuracies and Cohen’s Kappa coefficients of 0.6988/0.5326 on the SHHS 1000 subset, 0.6958/0.4830 on Huami ECG, and 0.6915/0.4615 in the Huami smartwatch PPG setting. Ablation analyses showed that multi-window HRV representation and temporal sleep-progress information provided the most consistent contributions, whereas demographic variables and feature-interaction strategies had auxiliary and dataset-dependent effects. For practical interpretation, configurations without the future-context 300-s feature are considered the relevant setting because they exclude explicitly post-epoch information, whereas configurations containing this feature are treated only as offline retrospective analyses. Further validation across independent datasets, wearable devices, prospective deployment settings, and external PPG interval-estimation pipelines is needed before broader real-world use.