A novel XAI framework for explainable AI-ECG using generative counterfactual XAI (GCX)

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Current AI explainability methods for ECGs often highlight segments without detailing their impact on predictions.
• Understanding AI decision-making in ECG interpretation is crucial for clinical adoption and trust.

Purpose of the Study:

• To introduce a generative counterfactual framework for enhancing the explainability of AI models interpreting ECGs.
• To enable clinicians to explore "what-if" scenarios by generating counterfactual ECGs that modify predictive values.

Main Methods:

• Developed a generative counterfactual approach to create synthetic ECGs.
• Generated counterfactual ECGs to illustrate the impact of specific morphological and rhythmic changes on AI predictions.
• Validated the generated counterfactuals against known clinical knowledge, including alterations related to potassium imbalance and atrial fibrillation.

Main Results:

• The framework successfully generated counterfactual ECGs that align with clinical understanding.
• Demonstrated how specific ECG changes (e.g., T wave amplitude, PR interval) influence AI-ECG predictions.
• Showcased the ability to visualize incremental modifications in ECGs and their effect on AI predictions.

Conclusions:

• Generative counterfactual XAI provides a powerful tool for understanding AI-ECG interpretation beyond static attribution maps.
• This approach has the potential to significantly increase clinician trust and confidence in AI-ECG diagnostic systems.
• Offers a promising avenue for improving the explainability and clinical reliability of AI in cardiovascular diagnostics.