Semi-Supervised Learning for Multi-Label Cardiovascular Diseases Prediction: A Multi-Dataset Study.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Electrocardiography (ECG) is crucial for non-invasive cardiovascular disease (CVD) prediction.
• Deep learning models show promise for ECG-based CVD diagnosis but face challenges like limited labeled data and multi-disease co-occurrence.
• Current models struggle with performance on unseen datasets, hindering widespread clinical application.

Purpose of the Study:

• To develop a unified deep learning framework for multi-label CVD prediction using limited supervision.
• To address label scarcity, multi-disease co-occurrence, and poor generalization to unseen datasets in ECG-based diagnostics.

Main Methods:

• Proposed ECGMatch, a multi-label semi-supervised model incorporating an ECGAugment module for data augmentation.
• Implemented a hyperparameter-efficient framework with neighbor agreement and knowledge distillation for pseudo-label generation and refinement.
• Introduced a label correlation alignment module to capture and propagate co-occurrence information between CVDs.

Main Results:

• Demonstrated the effectiveness and stability of ECGMatch across four datasets and three protocols.
• Achieved robust performance, particularly on unseen datasets, outperforming existing methods.
• Successfully mitigated the label scarcity problem and captured multi-label CVD information.

Conclusions:

• ECGMatch offers a promising solution for multi-label CVD prediction with limited supervision.
• The model's ability to handle label scarcity and generalize to new data paves the way for more reliable diagnostic systems.
• This framework can significantly advance the application of AI in clinical ECG interpretation for complex cardiovascular conditions.