Inter-patient multi-label ECG classification via low-rank adaptation fine-tuned large language models with dynamic...

Area of Science:

• Cardiology
• Artificial Intelligence
• Machine Learning

Background:

• Traditional electrocardiogram (ECG) analysis faces challenges with patient variability and multi-label disease correlations.
• Existing Large Language Model (LLM) approaches often limit adaptability by freezing internal parameters, hindering personalized signal pattern analysis.

Purpose of the Study:

• To develop an advanced framework for ECG analysis that overcomes inter-patient variability and models complex multi-label disease correlations.
• To enhance the diagnostic accuracy of ECG interpretation using adaptable LLMs and dynamic graph convolutional networks.

Main Methods:

• Integration of LLMs fine-tuned with Low Rank Adaptation (LoRA) for adaptive signal feature encoding.
• Utilizing a multi-label dynamic graph convolutional network to capture intricate correlations between various cardiac diseases.
• Employing refined semantic embeddings for accurate classification, reducing the need for extensive patient-specific labeled data.

Main Results:

• Achieved F1-scores of 0.781 (5-class) and 0.370 (20-class) on the PTB-XL dataset, surpassing state-of-the-art methods by 12.4% and 12.9%.
• Demonstrated robust cross-dataset generalization on CPSC2018 with an F1-score of 0.606, an 18.8% relative improvement.
• Successfully addressed inter-patient variability and disease correlations in ECG analysis.

Conclusions:

• The proposed framework offers a scalable and accurate solution for cardiac diagnosis, particularly in clinical settings with limited annotated data.
• This novel approach enhances diagnostic capabilities by effectively modeling complex ECG signal patterns and disease relationships.