WaSP-ECG: A Wave Segmentation Pretraining Toolkit for Electrocardiogram Analysis.

Area of Science:

• Artificial Intelligence
• Biomedical Signal Processing
• Machine Learning

Background:

• Representation learning in AI reduces reliance on labeled data for complex tasks.
• Electrocardiogram (ECG) analysis benefits from effective feature learning.
• Explainable AI (XAI) is crucial for clinical adoption of AI diagnostic tools.

Purpose of the Study:

• To develop and evaluate a Wave Segmentation Pretraining (WaSP) application for ECG representation learning.
• To assess the impact of WaSP on downstream ECG analysis tasks.
• To explore the potential of segmentation masks for XAI in ECG interpretation.

Main Methods:

• Developed a non-AI simulator to generate synthetic ECGs and wave segmentation masks.
• Trained U-Net models for wave segmentation on synthetic data.
• Fine-tuned pre-trained U-Net encoders on the PTB-XL dataset for ECG classification tasks (SR vs. AF, MI vs. Normal).
• Compared performance of pre-trained models against non-pretrained models.

Main Results:

• WaSP pretraining consistently improved model performance for both ECG signals and images.
• Significant performance gains were observed in ECG image analysis using WaSP.
• A hybrid AI and rule-based algorithm for AF detection showed high explainability but lower performance than end-to-end AI models.

Conclusions:

• WaSP enables AI models to learn valuable ECG features from synthetic data for real-world applications.
• Segmentation masks generated by WaSP can enhance confidence calibration in AI-driven ECG analysis.
• Combining AI segmentation masks with rule-based classifiers offers a pathway to explainable ECG diagnostics.