Rethinking PPG-based Sleep Staging: Datasets, Metrics, and Benchmarks

Research Motivation

Automated sleep staging is traditionally performed using 30-second epochs, a standard derived from EEG-based polysomnography. While PPG is a promising, low-cost alternative for ambulatory monitoring, it consistently underperforms compared to EEG. The authors argue that this performance gap is not due to a lack of information in the PPG signal, but rather a structural mismatch: PPG features (such as heart rate variability and pulse morphology) show sharp, second-level changes at sleep stage boundaries, whereas the 30-second epoch framework effectively masks these transitions by averaging them over a long window.

To address this, the authors propose Sleep Stage Transition Detection (SSTD), a task that shifts the focus from epoch-level classification to identifying the precise second-level onset of sleep stage transitions. They developed a two-stage label expansion pipeline to convert coarse 30-second labels into high-resolution pseudo-labels. First, they use a Hidden Semi-Markov Model (HSMM) to enforce physiologically plausible state durations. Second, they apply local changepoint refinement to pinpoint the exact timing of transitions based on signal variance shifts. Finally, they enforce physiological constraints, such as prohibiting direct transitions between Wake and Deep sleep without an intermediate Light sleep stage.

Empirical analysis confirms that PPG exhibits a higher boundary-to-stable contrast ratio compared to EEG, validating the hypothesis that PPG is particularly informative at transitions. By training models on these expanded, second-level labels, the authors achieved a 3.7–5.7 percentage point improvement in accuracy on conventional epoch-level staging tasks across four diverse model architectures. Furthermore, the benefits of this boundary-aware supervision persisted during zero-shot transfer evaluations, suggesting the approach is robust to cohort and protocol shifts.