TAC-ECG: A task-adaptive classification method for electrocardiogram based on cross-modal contrastive learning and...

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Cardiovascular diseases pose significant human health risks.
• Deep learning methods for electrocardiogram (ECG) analysis show promise but often lack flexibility for new tasks.
• Existing models require extensive retraining for different ECG classification applications, limiting clinical deployment.

Purpose of the Study:

• To develop a flexible and efficient deep learning framework for ECG classification.
• To enable rapid adaptation of ECG analysis models to diverse clinical tasks.
• To reduce the computational cost and resource requirements for multi-task ECG classification.

Main Methods:

• Proposed Task-Adaptive Classification for ECG (TAC-ECG) using cross-modal contrastive learning and low-rank convolutional adapters.
• Developed Contrastive ECG-Text Pre-training (CETP) to create a robust ECG encoder.
• Integrated a frozen pre-trained ECG encoder with a lightweight Low-Rank Convolutional Adapter (LRC-Adapter) for task-specific adaptation, requiring only adapter training.

Main Results:

• Evaluated on four datasets (CPSC2018, Cinc2017, PTB-XL, Chapman) for multi-category ECG classification.
• Achieved highly competitive results with significantly fewer trainable parameters (approx. 3%) compared to fully fine-tuned methods.
• Demonstrated the effectiveness and practicality of TAC-ECG across various network architectures and classification tasks.

Conclusions:

• TAC-ECG provides a flexible and efficient approach for ECG classification.
• The method allows for rapid adaptation to diverse tasks, enhancing clinical diagnostic practicality.
• TAC-ECG reduces resource consumption and deployment costs in multi-tasking scenarios.