MA-MIL: Sampling point-level abnormal ECG location method via weakly supervised learning

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Signal Processing
• Cardiovascular Diagnostics

Background:

• Current automated electrocardiogram (ECG) systems lack explainability, limiting clinical adoption.
• Supervised learning methods require extensive manual labeling for accurate abnormal ECG segmentation.
• Explainable AI is crucial for integrating automated diagnostics into clinical workflows.

Purpose of the Study:

• To develop an explainable AI framework for ECG analysis.
• To improve the accuracy of abnormal ECG segment detection and classification.
• To enhance the clinical applicability of automated ECG diagnostic systems.

Main Methods:

• A multi-instance learning (MIL) framework, termed MA-MIL, was developed.
• MA-MIL features a multi-layer, multi-instance structure aggregated at different scales.
• The framework was validated on the public MIT-BIH and a private ECG dataset.

Main Results:

• MA-MIL achieved high performance in ECG classification (Accuracy: 0.987, F1: 0.986).
• The model demonstrated strong abnormal segment detection at heartbeat and sub-heartbeat levels (Accuracy: 0.968, F1: 0.949).
• Intersection over Union (IoU) values improved by 17-31% compared to visualization methods.

Conclusions:

• MA-MIL accurately pinpoints abnormal segments within ECGs.
• The framework provides trustworthy results for clinical decision-making.
• This advancement supports the integration of AI in diagnostic cardiology.