Clinically-aligned explainable AI for atrial fibrillation detection: A U-Net inspired multi-lead ECG analysis...

Area of Science:

• Artificial Intelligence in Medicine
• Cardiology
• Medical Imaging Analysis

Background:

• Deep learning models for Electrocardiogram (ECG) analysis achieve high accuracy but often lack interpretability, limiting clinical trust and adoption.
• Interpretable AI is crucial for understanding model decision-making in critical applications like Atrial Fibrillation (AFib) detection.

Purpose of the Study:

• To develop a U-Net-inspired deep learning framework for AFib detection that provides clinically interpretable explanations.
• To elucidate model decision-making through extraction of morphological and rhythmic ECG features.

Main Methods:

• Utilized a U-Net-inspired encoder-decoder architecture for 12-lead ECG analysis, capturing inter-lead relationships from the PTB-XL database.
• Employed layer-wise Gradient-weighted Class Activation Mapping (Grad-CAM) for visualizing hierarchical feature focus and model interpretability.
• Incorporated RR interval variability assessment with heart rate stratification and optimized thresholds for physiological variations.

Main Results:

• Achieved high diagnostic performance with an Area Under the Curve (AUC) of 99.15% and sensitivity of 97.81% on PTB-XL-Improved labels.
• Grad-CAM confirmed accurate localization of QRS complexes (>97.5%) and identification of P-wave/rhythm irregularities (>93.1%).
• Independent RR-interval variability analysis demonstrated strong performance across different heart rate groups (AUCs 91.3%-95.5%).

Conclusions:

• The developed framework successfully integrates high diagnostic accuracy with transparent, clinically relevant visual explanations for AFib detection.
• Demonstrated that layer-wise explanations can align with the clinical ECG reading sequence, enhancing trust and understanding.
• External validation is recommended for broader clinical generalization of the interpretable deep learning model.