ECG-LM: Understanding Electrocardiogram with a Large Language Model

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Electrocardiograms (ECGs) are crucial for monitoring heart conditions but require expert interpretation.
• Current deep learning models struggle to integrate ECGs with patient data for nuanced clinical insights.
• Interpreting ECGs alongside patient data is complex and resource-intensive.

Purpose of the Study:

• To develop the first multi-modal large language model (LLM) capable of processing both natural language and ECG signals.
• To enhance cardiovascular disease detection and provide advanced ECG-related question answering capabilities.
• To address the limitations of existing deep learning methods in ECG analysis.

Main Methods:

• Developed ECG-Language Model (ECG-LM), a multi-modal LLM with a specialized ECG encoder.
• Aligned ECG signal features with textual features from an LLM.
• Created a pre-training dataset using medical guidelines for text-ECG pairs.
• Fine-tuned the model with clinical conversation and real hospital data.

Main Results:

• ECG-LM surpassed existing few-shot and zero-shot models in cardiovascular disease detection.
• Demonstrated superior performance across diagnostic, rhythm, and form tasks.
• Showcased strong potential in ECG-related question answering.

Conclusions:

• ECG-LM effectively captures intricate ECG features.
• The model offers versatility in disease prediction and advanced question answering.
• This represents a significant advancement in AI-driven ECG analysis.