Related Experiment Videos
A self-supervised learning framework with hierarchical residual cross fusion network for sleep apnea detection
Yuxuan Zhao1, Haitao He2, Ke Xu1
1School of Artificial Intelligence (School of Software), Yanshan University, Qinhuangdao, China.
Artificial Intelligence in Medicine
|June 22, 2026
Summary
This study introduces a self-supervised learning framework (SSL-HRCNet) for detecting sleep apnea from ECG signals, significantly reducing the need for labeled data and improving diagnostic accuracy.
Area of Science:
- Biomedical Engineering
- Artificial Intelligence
- Cardiology
Background:
- Sleep apnea is a common, underdiagnosed respiratory disorder with severe health consequences.
- Current deep learning methods for sleep apnea detection require extensive labeled data, which is costly and time-consuming to obtain.
- There is a need for efficient and accurate sleep apnea detection methods that minimize reliance on labeled data.
Purpose of the Study:
- To propose a novel self-supervised learning framework (SSL-HRCNet) for sleep apnea detection using single-lead ECG signals.
- To overcome the limitations of data dependency in existing deep learning approaches.
- To improve the accuracy and efficiency of sleep apnea diagnosis.
Main Methods:
- A two-stage pipeline involving self-supervised pre-training and supervised fine-tuning.
- Contrastive learning using R-R intervals (RRIs) and R-peak amplitudes as positive pairs for learning cardiopulmonary representations from unlabeled ECG data.
- Hierarchical residual blocks with depthwise-pointwise convolutions for multi-scale temporal modeling and an attention-based cross fusion module for integrating RRI and R-peak features.
Main Results:
- SSL-HRCNet achieved competitive accuracy with only 10% labeled data on the Apnea-ECG dataset.
- The model outperformed existing approaches under full supervision, reaching 91.91% accuracy and 89.97% sensitivity.
- Learned representations demonstrated robustness to domain shifts and adaptability to new datasets (UCDDB), indicating strong generalization capabilities.
Conclusions:
- The proposed SSL-HRCNet framework effectively detects sleep apnea using self-supervised learning on ECG signals, significantly reducing the need for labeled data.
- SSL-HRCNet offers a promising, data-efficient, and accurate alternative for sleep apnea diagnosis.
- The model's ability to transfer learned representations highlights its potential for broad clinical application.
