From Pulses to Sleep Stages: Towards Optimized Sleep Classification Using Heart-Rate Variability

Area of Science:

• Biomedical Engineering
• Sleep Science
• Digital Health

Background:

• Consumer wearables for sleep tracking are increasingly popular, but their accuracy is often criticized, potentially leading to
• orthosomnia
• and negative health outcomes.

Purpose of the Study:

• To optimize a previously developed sleep classification procedure for ambulatory sleep sensing using consumer wearables.
• To minimize erroneous sleep classification in self-reported poor sleepers by introducing advanced signal quality control and a novel loss function model.

Main Methods:

• Utilized a new interbeat-interval (IBI) quality control method employing a random forest algorithm.
• Implemented a loss function model for sleep classification, contrasting it with an epoch-by-epoch accuracy model.
• Compared classification performance using signals from polysomnography (PSG), single-lead ECG, Polar® H10 (ECG), and Polar® Verity Sense (PPG) wearables.

Main Results:

• The optimized loss function model achieved high overall accuracy (86.3% for ECG, 84.4% for H10, 84.2% for VS), with notable improvements in deep sleep and wake classification.
• The model demonstrated moderate to high correlations and agreement with PSG for primary sleep parameters.
• Excellent reliability was observed for most sleep parameters, and accurate 4-class sleep staging was maintained in individuals taking heart-affecting or psychoactive medications.

Conclusions:

• The optimized sleep classification model enhances accuracy and reliability of wearable sleep sensing, addressing prior criticisms.
• This approach shows promise for clinical applications, particularly for older individuals or those with common disorders, by providing dependable sleep data.
• Further validation of these algorithms could increase confidence in wearable sleep technology and facilitate its integration into clinical practice.