The cost of explainability in artificial intelligence-enhanced electrocardiogram models.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Artificial intelligence-enhanced electrocardiogram (AI-ECG) models offer high diagnostic and prognostic performance but lack clinical trust due to their black-box nature.
• Explainable AI (XAI) is crucial for transparent and trustworthy medical decision-making, moving beyond unreliable post-hoc methods.

Purpose of the Study:

• To develop a novel framework (VAE-SCAN) for explicit representation learning using variational autoencoders (VAEs) to capture inherently interpretable ECG features.
• To investigate the performance-explainability trade-off in AI-ECG models, focusing on interpretable feature representations.
• To model bi-directional associations between ECG features and clinical factors.

Main Methods:

• Utilized variational autoencoders (VAEs) for explicit representation learning of ECG features.
• Developed the VAE-SCAN framework to model interpretable, bi-directional associations between ECG features and clinical factors.
• Evaluated the impact of different ECG representations on decoding performance across models with varying explainability.

Main Results:

• Demonstrated that intrinsic ECG interpretability, achieved through VAE-based representation learning, introduces a performance cost.
• Quantified the trade-off between model performance and the level of ECG feature interpretability.
• Identified implications of this trade-off for the clinical adoption of AI-ECG technologies.

Conclusions:

• Explicit representation learning with VAEs can create more interpretable AI-ECG models.
• The study quantifies the performance cost associated with intrinsic ECG interpretability.
• Findings provide guidance for developing trustworthy and clinically applicable AI-ECG systems.