Macro-Sleep Staging With ECG-Derived Instantaneous Heart Rate and Respiration Signals and Multi-Input 1-D CNN–BiGRU

Abstract

Millions of people worldwide are affected by sleep disorders. Polysomnography (PSG) is a type of sleep study that is used to diagnose sleep disorders, such as using sleep staging. However, PSG can be burdensome, time-consuming, expensive, and may not be readily accessible. Sleep cycles can cause fluctuations in heart rate and respiration which can be estimated from electrocardiogram (ECG) signal. In addition, an ECG signal is easier to acquire compared with the multiparametric PSG. As such, in this work, we study the use of ECG data for sleep staging. In particular, the proposed method utilizes ECG-derived instantaneous heart rate (IHR) and ECG-derived respiration (EDR) signals. The IHR and EDR signals are input to a neural network that consists of a 1-D convolutional neural network (CNN) to learn the temporal and spectral characteristics in short-time windows and a bidirectional gated recurrent unit (BiGRU) to learn the long-term dependencies between the windows in both the forward and backward directions. The multi-input network analyzes the IHR and EDR signals independently. Features extracted from the analysis of the IHR and EDR signals are combined for the classification of sleep stages. The proposed approach is evaluated on a dataset of 993 overnight sleep studies, where the sleep stages are scored by experts. An average class accuracy (ACC) of 83.8% is achieved in two-class staging (wake versus sleep), 79.3% in three-class staging (wake versus NREM sleep versus REM sleep), 73.7% in four-class staging (wake versus light sleep versus deep sleep versus REM sleep), and 63.5% in five-class staging (wake versus <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${N}1$ </tex-math></inline-formula> versus <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${N}2$ </tex-math></inline-formula> versus <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">${N}3$ </tex-math></inline-formula> versus REM). Our method is seen to be effective in macro-sleep staging, particularly in two-, three-, and four-class sleep staging, achieving class ACC values greater than 76% for all classes, except light sleep. In addition, our method shows good generalizability in external validation.

Topics & Concepts