Multi-Level Explainable AI for ECG-Based Atrial Fibrillation Detection: Exploring LIME, SHAP, and Grad-CAM for...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Automated electrocardiogram (ECG) analysis using deep learning shows potential for atrial fibrillation (AFib) detection.
• The "black box" nature of deep learning models hinders their clinical integration and interpretability.
• Explainable AI (XAI) methods are crucial for bridging the gap between AI performance and clinical trust.

Purpose of the Study:

• To develop and evaluate a multi-level XAI framework for AFib detection using ECG data.
• To enhance the clinical interpretability of deep learning models for AFib diagnosis.
• To align AI-driven insights with established clinical diagnostic patterns for AFib.

Main Methods:

• A ResNet-based deep learning model was developed for AFib detection using the PhysioNet AFib Dataset (5,830 ECG samples).
• A multi-level XAI framework combining LIME, SHAP, and Grad-CAM was implemented.
• A novel RR-interval aggregation method was introduced to improve the clinical relevance of XAI outputs.

Main Results:

• The ResNet model achieved high accuracy: 91.3% precision and 88.7% recall for AFib detection.
• Normal rhythm classification achieved superior performance with 98.4% precision and 98.8% recall.
• Grad-CAM successfully identified key morphological features: P-wave absence, pronounced T-waves, and irregular RR-intervals, aligning with clinical AFib indicators.

Conclusions:

• The proposed multi-level XAI framework provides clinically meaningful interpretations for AFib detection, enhancing model transparency.
• Integrating XAI methods, particularly with RR-interval aggregation, improves the alignment of AI predictions with clinical diagnostic criteria.
• This approach facilitates the adoption of AI in clinical cardiology by offering interpretable and accurate AFib detection.