ECG-aBcDe: Overcoming model dependence, encoding ECG into a universal language for any large language model.

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Signal Processing

Background:

• Large Language Models (LLMs) show potential for electrocardiogram (ECG) analysis.
• Existing methods face challenges in transferability, time-scale learning, and interpretability.
• Model-specific ECG encoders and Transformer limitations hinder LLM application in ECG analysis.

Purpose of the Study:

• To introduce ECG-aBcDe, a novel ECG encoding method for universal LLM interpretability.
• To enable direct fine-tuning of pre-trained LLMs for ECG analysis without architectural changes.
• To enhance the temporal modeling and interpretability of LLMs in ECG analysis.

Main Methods:

• Developed ECG-aBcDe to transform ECG signals into a universal ECG language.
• Constructed a hybrid dataset of ECG language and natural language for LLM fine-tuning.
• Enabled bidirectional conversion between ECG and ECG language for attention heatmap extraction.

Main Results:

• Achieved competitive Rouge-L and Meteor scores.
• Significantly outperformed existing methods on Bleu-4, reaching 42.58 and 30.76.
• Demonstrated enhanced temporal modeling and interpretability for LLMs in ECG analysis.

Conclusions:

• ECG-aBcDe provides a "construct once, use anywhere" capability for LLM-based ECG analysis.
• The method addresses limitations in transferability, time-scale information learning, and interpretability.
• Presents a new paradigm for integrating LLMs into clinical decision support systems for ECG analysis.