Sleep staging using PPG-derived heart rate and accelerometer data from a smart ring with lightweight neural networks

Sleep staging plays an important role in assessing and diagnosing sleep disorders, however, polysomnography, the gold standard for sleep monitoring, is expensive, time-consuming, and often inaccessible. Wearable devices offer a promising alternative, providing a more convenient and scalable solution for long-term sleep monitoring. Compared to wrist-worn devices, smart rings offer a more stable signal acquisition environment, reducing motion artifacts and improving data reliability. This work investigates sleep staging using photoplethysmography (PPG)-derived instantaneous heart rate (IHR) signal and zero-crossing mode (ZCM) feature obtained from accelerometer data, both captured using a wearable smart ring. We propose a lightweight neural network model designed for wearable-based sleep staging, incorporating feature extraction, temporal modeling, and class balancing strategies. The IHR-based model consists of 503 k learnable parameters, while the ZCM-based model has 133 k learnable parameters, making them well-suited for efficient deployment on wearable devices. The proposed method is evaluated on a dataset of expert-annotated overnight sleep studies, using both IHR signal and ZCM features individually and in combination, in subject-independent cross-validation. The experimental results demonstrate that IHR features alone yield strong classification performance, achieving macro-average recall (unweighted average recall) values of 0.849, 0.805, 0.750, and 0.663 in two-class (wake vs. sleep), three-class (wake vs. non-rapid eye movement (NREM) sleep (N1, N2, N3) vs. REM sleep), four-class (wake vs. light sleep vs. deep sleep vs. REM sleep), and five-class (wake vs. N1 vs. N2 vs. N3 vs. REM) classification tasks, respectively. When combining IHR and ZCM features, classification performance improves further, reaching macro-average recall values of 0.866, 0.832, 0.772, and 0.671 in the respective tasks. These results highlight the effectiveness of IHR-based sleep staging and the additional benefit provided by movement-based ZCM features, particularly in two-, three-, and four-class sleep staging where we could achieve macro-average recall values of 0.750 or higher. The proposed smart ring-based system demonstrates strong potential for real-world sleep assessment by integrating multimodal physiological signals through lightweight neural networks, advancing non-invasive measurement and intelligent instrumentation.