ECG-Derived Respiration for Sleep-Wake Stage Classification

Abstract

Sleep disorders affect millions of people worldwide. Polysomnography (PSG) is a sleep study that is commonly used to diagnose sleep disorders, such as using sleep staging. However, PSG can be labor intensive, time consuming, expensive, and may not be easily available. Sleep and wake cycles can cause variation in heart rate and respiration which can be estimated using electrocardiogram (ECG), available as wearable sensors. As such, this work studies the use of single-lead ECG for detecting sleep and wake stages, in particular, using the heart rate variability (HRV) and ECG-derived respiration (EDR) signals. Various temporal and spectral descriptors are extracted from the HRV and EDR signals for this purpose. Sequential backward feature selection is employed to select the discriminative features for classification using logistic regression. The proposed method is evaluated on a dataset of more than 85 hours of ECG recordings from 16 subjects in leave-one-subject-out cross-validation. An accuracy of 75% ($\text{AUC} =0.83$) is achieved using the EDR features in classifying sleep and wake stages. This increased to an accuracy of 80% ($\text{AUC} =0.88$) when combined with HRV features. The proposed method demonstrates potential to be used for screening sleep disorders using ECG.