ECG Semantic Integrator (ESI): A Foundation ECG Model Pretrained with LLM-Enhanced Cardiological Text.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Deep learning significantly improves electrocardiogram (ECG) analysis accuracy and efficiency in cardiac healthcare.
• A key challenge is learning robust ECG representations without extensive labeled datasets.

Purpose of the Study:

• To develop a novel multimodal contrastive pretraining framework (ECG Semantic Integrator - ESI) for learning robust ECG representations from ECG signals and associated text.
• To address the limitation of scarce labeled datasets in deep learning for ECG analysis.

Main Methods:

• Proposed ECG Semantic Integrator (ESI), a multimodal contrastive pretraining framework using dual objective functions (contrastive and captioning loss).
• Developed a retrieval-augmented generation (RAG)-based Large Language Model (LLM) pipeline, Cardio Query Assistant (CQA), to generate detailed ECG textual descriptions.
• Compiled a large-scale multimodal dataset of over 660,000 ECG-text pairs for pretraining.

Main Results:

• ESI learned robust and generalizable representations for 12-lead ECG.
• Demonstrated substantial improvements in downstream tasks like arrhythmia detection and ECG-based subject identification compared to strong baselines.
• Outperformed existing supervised, self-supervised, and multimodal pretraining methods.

Conclusions:

• Multimodal pretraining combining ECG signals and text descriptions significantly enhances ECG analysis.
• ESI offers a promising approach for advancing cardiac healthcare diagnostics through improved ECG interpretation.
• The developed framework and dataset pave the way for more accurate and efficient AI-driven cardiovascular disease detection.