Sleep Stages Classification Using ECG Signals and Deep Learning: A Comparative Study

Abstract

Sleep disorders hinder good sleep quality. Therefore, a demand for bedtime monitoring arose to reveal those sleep disorders. Polysomnography is a sleep study used for sleep quality monitoring, which highly relies on electroencephalogram (EEG) signals. It depends on mapping every 30 seconds of the collected bio-signals into a sleep stage. Poor sleep quality will impact the normal sleep stage rhythm. Polysomnography has been considered the gold standard. However, it has drawbacks as it is expensive, not convenient for all patients, and it can add artifacts. Thus, researchers are trying to figure out methodologies that utilize alternative signals that are more inexpensive and convenient than EEG signals. Our approach classifies sleep stages into three classes (Wake, NREM, REM) utilizing the deep learning capabilities represented in pre-trained models to be used with scalograms. Scalograms are generated from 17 electrocardiogram (ECG) signals extracted from the polysomnography recordings within the Multi-Ethnic Study of Atherosclerosis (MESA) dataset. A comparison between the efficiency of AlexNet and VGG16 models revealed the superior excel of AlexNet over VGG16 using the same data and hyperparameters. The accuracy for the AlexNet trial was 87.3%, and F1-score was 0.87.