GATE: Graph and Text Exchange for Zero-Shot ECG Classification with LLM Prompts.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Electrocardiography (ECG) is crucial for diagnosing heart conditions, but supervised learning is hindered by limited annotated data.
• Existing self-supervised learning (SSL) methods for ECG struggle with semantic accuracy, spatial details, and medical knowledge integration.

Purpose of the Study:

• To develop a multimodal self-supervised learning (SSL) framework, GATE (Graph-And-Text Exchange), to improve ECG representation learning.
• To address limitations of current SSL techniques by integrating graph-structured ECG data with clinical text reports.

Main Methods:

• GATE utilizes a spatiotemporal graph encoder to capture complex ECG lead dependencies.
• A lexical knowledge-embedded codebook enhances clinical report semantics for better graph-text alignment.
• Integration with a large language model and knowledge base enables zero-shot classification via enriched disease descriptions.

Main Results:

• GATE significantly outperforms state-of-the-art self-supervised and multimodal methods on three real-world ECG datasets.
• The framework demonstrates strong performance in both low-resource and zero-shot classification scenarios.
• Remarkable results were achieved even when trained on just 1% of labeled data, showing high generalization capability.

Conclusions:

• GATE offers a powerful approach to enhance ECG representation learning through multimodal SSL.
• The framework shows significant clinical potential, particularly for improving diagnostic accuracy with scarce labeled data.
• GATE's ability to leverage both graph and text data paves the way for more robust and interpretable AI in cardiology.