It is All in the Wrist: Wearable Sleep Staging in a Clinical Population versus Reference Polysomnography

Abstract

Purpose There is great interest in unobtrusive long-term sleep measurements using wearable devices based on reflective photoplethysmography (PPG). Unfortunately, consumer devices are not validated in patient populations and therefore not suitable for clinical use. Several sleep staging algorithms have been developed and validated based on ECG-signals. However, translation from these techniques to data derived by wearable PPG is not trivial, and requires the differences between sensing modalities to be integrated in the algorithm, or having the model trained directly with data obtained with the target sensor. Either way, validation of PPG-based sleep staging algorithms requires a large dataset containing both gold standard measurements and PPG-sensor in the applicable clinical population. Here, we take these important steps towards unobtrusive, long-term sleep monitoring. Methods We developed and trained an algorithm based on wrist-worn PPG and accelerometry. The method was validated against reference polysomnography in an independent clinical population comprising 244 adults and 48 children (age: 3 to 82 years) with a wide variety of sleep disorders. Results The classifier achieved substantial agreement on four-class sleep staging with an average Cohen’s kappa of 0.62 and accuracy of 76.4%. For children/adolescents, it achieved even higher agreement with an average kappa of 0.66 and accuracy of 77.9%. Performance was significantly higher in non-REM parasomnias (kappa = 0.69, accuracy = 80.1%) and significantly lower in REM parasomnias (kappa = 0.55, accuracy = 72.3%). A weak correlation was found between age and kappa ( ρ = −0.30, p<0.001) and age and accuracy ( ρ = −0.22, p<0.001). Conclusion This study shows the feasibility of automatic wearable sleep staging in patients with a broad variety of sleep disorders and a wide age range. Results demonstrate the potential for ambulatory long-term monitoring of clinical populations, which may improve diagnosis, estimation of severity and follow up in both sleep medicine and research.

SleepPPG-Net: a deep learning algorithm for robust sleep staging from continuous photoplethysmography

“…On the held out test set of 204 MESA patients, SleepPPG-Net scored a κ of 0.75 against 0.66 for BM-FE and 0.69 for BM-DTS approaches. SleepPPG-Net performance is also significantly (p < 0.001, two-sample t-test) higher than the current published SOTA results for sleep staging from PPG which stand at a κ of 0.66 [22,23], and significantly (p = 0.02, two-sample t-test) higher than the current SOTA results for sleepsleep staging from ECG which are reported at κ of 0.69 [20]. Figure 9 presents an example of the hypnograms generated by BM-FE, DB-DTS and SleepPPG-Net for a single patient.…”

“…Most works that use PPG usually do so in the context of transfer learning (TL), where models are trained on a large database of heart rate variability (HRV) measures and then fine-tuned to a smaller database of pulse rate variability (PRV) measures derived from the IBIs detected on the PPG. These works report κ performance approaching 0.66 [22,23]. Sleep staging from the raw PPG is a relatively novel approach.…”

A Comparison of Signal Combinations for Deep Learning-Based Simultaneous Sleep Staging and Respiratory Event Detection

“…Since the EEG measurement setup is quite complex and requires expert assistance for both in-laboratory and home measurements, alternative signal sources for sleep staging, such as the electrocardiography (ECG) and body movements [13], [14], have been studied. In addition, automatic deep learningbased sleep staging from photoplethysmography (PPG) has been performed with promising results [15]- [17]. Especially the detection of REM sleep from PPG has been performed with an accuracy of 87% using deep learning [16].…”

A Flexible Deep Learning Architecture for Temporal Sleep Stage Classification Using Accelerometry and Photoplethysmography

“…All approaches that used both ACC and PPG were featurebased [14]- [20]. The best performing of these, i.e., Wulterkens et al, reported similar performance to that of the proposed approached: 𝜅 = 0.62 ± 0.12 [19], while all other approaches had substantially lower performance. One feature-based approach reported a performance increase from 𝜅 = 0.55 to 𝜅 = 0.65 by pretraining their classifier on features extracted from ECG [21].…”

“…Sleep is a dynamic process with a cyclic pattern that cycles through NREM and REM sleep with a period of approximately 90-110 minutes. The most recent and best performing sleep stage classification algorithms are temporal models that are either based on recurrent frameworks, e.g., long-short term memory (LSTM) [9], [19], [25] and gated recurrent units (GRU) [22], [26] or convolutional neural network (CNN) architectures, e.g., dilated convolutions [27] and the residual U-Net architecture [28]. While there is consensus that including contextual information from neighboring epochs increase performance, the segment size that these temporal models are trained on vary considerably between studies; from minutes [26], [28], to hours [22], and to the entire recording length [27].…”