Continuous momentary assessment of sleep using reliable HRV-based wearables: A possible alternative to DREEM?

Introduction:
Due to the recent exponential growth in the utilization of wearables for sleep quantification, together with rising concerns about the accuracy of such sleep measurements, it is critical to refine and validate such ambulatory sleep solutions. To address this, we here developed a sleep classification method for ambulatory sleep analysis, with robust neuronal networks and affordable yet reliable HRV-based sensors.
Method:
In a study involving 136 self-reported poor sleepers, we implemented an interbeatinterval (IBI) quality control using random forest and report sleep classification accuracy in 4 classes (wake, light, deep, REM). This data is compared to goldstandard PSG recordings, one-channel ECG, and the two HRV-based consumer wearables: the ECG breast belt Polar® H10 and the optical Photoplethysmography (PPG) heart-rate sensor Polar® Verity Sense (VS).
Results:
The model - implemented in the sleep² App - demonstrates high accuracy with gold-standard ECG data (86.3%, κ = 0.79), H10 (84.4%, κ = 0.76), and VS (84.2%, κ = 0.75) sensors; notably also being accurate on an 30sec epoch-by-epoch basis for REM (87%), deep sleep (80%), light sleep (86%) and wake classification (79%). Intraclass correlations with PSG on primary sleep parameters like TST, WASO or Sleep Efficiency are showing good to excellent agreement (≈0.9). Importantly, accurate classification was maintained across all 4 classes, even in users which were taking medication directly affecting heart activity as well as such on psychoactive medication.
Conclusion:
Until recently many scientists used EEG-based headband systems such as ©DREEM which unfortunately was widely discontinued for scientific purposes. Data suggests that arm-worn and validated HRV-based sensors such as the VS allow accurate sleep staging and may be a viable and affordable alternative for basic and clinical sleep research. The multi-convolutional neural network (MCNN) presented here is implemented in the sleep² App and is agnostic to the sensors which deliver the IBI input data needed for classification. It should therefore be easily applicable to anyone interested in continuous and ambulatory epoch-by-epoch sleep classification in 4 classes.
Conflict of Interest: Yes- MS is Co-Founder and CSO of sleep² (NUKKUAA GmbH)