From Pulses to Sleep Stages: using Heart-Rate Variability from Low-Cost Wearable Devices for Accurate four-class sleep stage classification

Introduction: Although polysomnography (PSG) conducted by human experts is considered the "gold standard" for measuring sleep, PSG and manual sleep staging require significant personnel and time resources, making thus monitoring an individual's sleep over extended periods impractical. For this reason, we aspired to develop an innovative, cost-effective and automated sleep staging four-class (Wake, Light [N1 + N2], Deep, REM) approach, utilizing deep learning, solely using inter-beat-interval (IBI) data.
Materials and Methods: This approach involves training a multi-resolution convolutional neural network (MCNN) on IBI data from 8898 manually sleep-staged recordings of full nights. Subsequently, we used transfer learning on the IBIs from two affordable consumer wearables (an optical heart rate sensor - VS, and a breast belt - H10) and evaluated the MCNN's sleep classification, epoch by epoch (30 sec.), against PSG. In addition, using the H10 we collected daily ECG data from 49 participants experiencing sleep difficulties throughout a digital cognitive-behavioural therapy for insomnia (CBT-I) program offered via the NUKKUAA™ App, to examine the objective effects of sleep training on daily objective sleep measures.
Results: Regarding the evaluation of the MCNN sleep stage classification, we observed an overall classification accuracy comparable to expert inter-rater reliability for both devices (VS: 81%, Cohen's 𝜅 = 0.69; H10: 80.3%, Cohen's 𝜅 = 0.69). Concerning the evaluation of the digital CBT-i program, participants reported significant enhancements in subjective sleep quality and sleep onset latency at the end of the training. At the same time, there was a trend for objective sleep onset latency improvement. Weekly sleep onset latency, wake time during sleep, and total sleep time exhibited noteworthy correlations with the subjective reports.
Conclusions: Combining deep learning with suitable wearables allows continuous and accurate sleep monitoring in naturalistic settings with profound implications for answering basic and clinical research questions.