Automated REM vs NREM sleep staging using single overnight heart rate and accelerometer data

Accurate sleep staging is crucial for diagnosing sleep disorders such as insomnia, REM behavior disorder, and sleep apnea. The standard method for sleep staging requires polysomnography, an expensive and labor-intensive procedure that most often happens in a clinical setting. In this study, we develop an algorithm that performs sleep staging for 13 participants using overnight heart rate and triaxial acceleration data collected by Apple watch. We employ a two-stage K-means/hidden Markov modeling approach that achieves a strong performance in differentiating REM from NREM when compared to polysomnography-scored sleep labels, with an average accuracy of 81.5%, macro F1-score of 0.73, and Matthews Correlation Coefficient (MCC) of 0.5. This study highlights the ability of this method to perform low-cost sleep staging using data from commonly available devices.Clinical Relevance-A Hidden Markov Model (HMM) approach using heart rate and triaxial accelerometer data can provide a scalable and cost-effective alternative to polysomnography. This approach has the potential to improve at-home sleep tracking and enhance personalized sleep health interventions.