Explaining deep learning for ECG analysis: Building blocks for auditing and knowledge discovery.

Area of Science:

• Artificial Intelligence
• Cardiology
• Medical Informatics

Background:

• Deep neural networks (DNNs) excel at analyzing electrocardiogram (ECG) data for cardiac condition identification.
• The 'black box' nature of DNNs limits transparency and clinical trust.
• Explainable AI (XAI) offers methods to interpret DNN decisions.

Purpose of the Study:

• To comprehensively analyze post-hoc XAI methods for ECG analysis.
• To establish quantitative evidence aligning DNN behavior with clinical decision-making.
• To explore XAI for knowledge discovery in cardiac diagnostics.

Main Methods:

• Investigated glocal (aggregated local attributions) and global (concept-based XAI) perspectives.
• Performed sanity checks to identify saliency as a robust attribution method.
• Conducted dataset-wide analyses across patient subgroups.

Main Results:

• Saliency attribution was identified as the most sensible method through rigorous sanity checks.
• Quantitative evidence demonstrates alignment between DNN model behavior and cardiologists' decision rules.
• XAI techniques were shown to facilitate knowledge discovery, including identifying myocardial infarction subtypes.

Conclusions:

• Proposed XAI methods enhance the internal validity assessment of DNNs in ECG analysis.
• XAI can serve as a foundational tool for certification processes and clinical knowledge discovery.
• This work bridges the gap between complex AI models and clinical interpretability in cardiology.