CPSS: Fusing consistency regularization and pseudo-labeling techniques for semi-supervised deep cardiovascular...

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Supervised deep learning for electrocardiogram (ECG) analysis requires extensive labeled data.
• Manual ECG labeling is time-consuming and requires specialized medical expertise.
• Semi-supervised learning (SSL) offers a solution by leveraging unlabeled data.

Purpose of the Study:

• To enhance cardiovascular disease (CVD) detection performance by maximizing the utility of unlabeled ECG data.
• To develop an efficient method that reduces the dependency on large, manually labeled ECG datasets.

Main Methods:

• A novel SSL algorithm, CPSS, combining consistency regularization and pseudo-labeling techniques was developed.
• The method incorporates supervised learning on labeled ECGs and unsupervised learning on augmented unlabeled ECGs.
• VGGNet and ResNet classifiers were jointly optimized using both labeled and unlabeled data.

Main Results:

• CPSS demonstrated significant accuracy improvements over supervised learning with only 10% labeled ECGs across multiple databases.
• Achieved comparable performance to fully supervised methods while reducing labeling workload by 90%.
• A real-time CVD monitoring system was developed using the trained classifiers on an SoC.

Conclusions:

• The proposed CPSS algorithm effectively enhances CVD detection using unlabeled ECG data.
• CPSS substantially alleviates the burden of ECG labeling in deep learning applications.
• The developed monitoring system highlights the practical, real-world applicability of CPSS.