2025-12-30 中山大学(SYSU)
<関連情報>
- https://www.sysu.edu.cn/sysuen/info/1012/58201.htm
- https://www.sciencedirect.com/science/article/abs/pii/S0950705125021677
SleepECGFusion: 単一誘導心電図を用いた睡眠段階の自動分類のためのクロスモーダルディープラーニングフレームワーク SleepECGFusion: a cross-modal deep learning framework for automatic sleep stage classification using single-lead ECG
Xuanhao Qi, Li Chen, Hongze Liu, Shishi Tang, Zhipeng He, Yichen Dai, Kaizhe Zheng, Jianping Man, Yi Zhou
Knowledge-Based Systems Available online: 13 December 2025
DOI:https://doi.org/10.1016/j.knosys.2025.115132
Highlights
- A novel cross-modal deep learning framework named SleepECGFusion is proposed for accurate sleep stage classification, integrating multiscale temporal ECG sequences and time-frequency representations.
- SleepECGFusion achieves state-of-the-art performance in ECG-based sleep stage classification and demonstrates robust generalization to sleep apnea detection, yielding comparable results in healthy and apnea populations.
- This study systematically validates that longer input signal durations significantly enhance the performance of SleepECGFusion across all sleep stages, particularly for the challenging N1 stage.
- Comprehensive ablation experiments meticulously demonstrate the validity and robustness of the proposed SleepECGFusion architecture and its core components.
Abstract
Accurate and readily accessible sleep stage classification is crucial for diagnosing sleep disorders. However, conventional polysomnography (PSG) is cumbersome. Given the convenience of single-lead electrocardiogram (ECG) signals, we propose a novel cross-modal deep learning framework named SleepECGFusion. It aims to achieve high-precision automated sleep stage classification and apnea detection. SleepECGFusion ingeniously integrates multi-scale time-domain features extracted from raw ECG sequences and high-level spectral features derived from time-frequency (TF) maps for high-level spectral representation. The framework consists of a 1) temporal branch based on multiscale feature extraction (MSFE), convolutional attention module (CBAM) and bidirectional long- and short-term memory (BiLSTM) network and 2) time-frequency branch with EfficientNet-B0 as the backbone. We systematically evaluated the effect of input signal duration on model performance, thereby revealing that longer segments significantly enhance classification accuracy across all sleep stages. Furthermore, SleepECGFusion achieves state-of-the-art performance in ECG-based sleep stage classification tasks. To rigorously assess generalizability, we conducted cross-dataset performance benchmarking using SHHS1 for training and ISRUC-S3 for testing in a subject-independent manner, demonstrating robust performance across diverse populations. The framework also demonstrates strong generalization capabilities for the detection of sleep apnea via transfer learning and obtains comparable classification results between healthy individuals and sleep apnea patients. Overall, this work emphasizes the significant potential of single-lead ECG as a viable, non-invasive alternative to PSG. It lays the foundation for scalable and accessible sleep health monitoring and diagnosis.
