Investigating explainable arrhythmia classification using class-specific ensemble model from segmented ECG signals.

Area of Science:

• Cardiology
• Biomedical Engineering
• Artificial Intelligence in Medicine

Background:

• Accurate arrhythmia classification from short ECGs is challenging due to limitations in existing methods.
• Prior studies often use single-lead ECGs, uniform model fusion, or lack uncertainty quantification.

Purpose of the Study:

• To develop a novel class-specific weighted ensemble for 12-lead, 10-second ECG segments.
• To provide lead-resolved SHAP explanations for improved interpretability.
• To enhance arrhythmia classification accuracy and quantify uncertainty.

Main Methods:

• Proposed a class-specific weighted ensemble fusing multiple models with per-class weights.
• Utilized a 12-lead, 10-second ECG segment analysis.
• Employed an estimation-based framework with Wilson and Newcombe 95% confidence intervals (CIs) and Cohen's h for evaluation.
• Generated lead-resolved SHAP explanations for feature importance.

Main Results:

• The ensemble demonstrated higher slow (SB) heart rate recall and fast (ST) heart rate precision compared to a Bagging baseline.
• Overall accuracy was comparable, with CIs spanning zero.
• SHAP analysis identified lead-specific contributors, such as P-wave area in lead V2, suggesting potential for minimal-lead configurations.

Conclusions:

• The proposed method achieves class-specific performance gains in arrhythmia classification.
• Lead-level interpretability via SHAP aids in understanding classification drivers.
• Findings support the development of more accurate and interpretable portable ECG devices for arrhythmia detection.