A systematic review of deep learning methods for modeling electrocardiograms during sleep

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Artificial Intelligence in Healthcare

Background:

• Sleep is crucial for human health, with Polysomnography (PSG) being the standard for assessment.
• PSG is resource-intensive, limiting its widespread application.
• Electrocardiography (ECG) signals contain rich physiological information, including respiratory patterns relevant to sleep.

Purpose of the Study:

• To systematically review recent deep learning (DL) methods utilizing ECG for sleep-related tasks.
• To analyze these studies based on data, models, and specific sleep tasks.
• To identify shortcomings, summarize findings, and highlight future opportunities in ECG-based sleep analysis.

Main Methods:

• Systematic identification and analysis of recent research studies on ECG-based DL for sleep.
• Categorization of studies by data characteristics, deep learning model architectures, and sleep-related tasks.
• Evaluation of method performance, interpretability, scalability, and transferability.

Main Results:

• Deep learning methods leveraging ECG data demonstrate superior accuracy compared to traditional approaches for sleep tasks.
• Integration of multiple signal features and advanced model structures enhances predictive performance.
• ECG-based DL shows potential for non-invasive, convenient sleep monitoring.

Conclusions:

• ECG-based deep learning presents a powerful and accurate approach for sleep analysis, overcoming limitations of PSG.
• Future research should focus on developing interpretable, scalable, and transferable DL models for clinical and assisted-living applications.
• This review is the first systematic examination of ECG-based DL methods specifically for sleep-related applications.