Interpretable heartbeat classification using local model-agnostic explanations on ECGs.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Electrocardiogram (ECG) interpretation is crucial for cardiovascular disease management but is subjective and prone to errors.
• Machine learning (ML) models offer diagnostic support but often lack interpretability, hindering clinical adoption.
• Explainable Artificial Intelligence (XAI) aims to address the transparency limitations of ML models in healthcare.

Purpose of the Study:

• To develop and evaluate an XAI solution for enhancing the explainability of heartbeat classification from ECG data.
• To introduce a conceptual framework for explainable time series analysis incorporating temporal dependencies.
• To assess the clinical utility of XAI-generated visual explanations for ECG interpretation.

Main Methods:

• Utilized several state-of-the-art, model-agnostic XAI methods for heartbeat classification.
• Proposed an original method integrating the time series' derivative to capture temporal dependencies between samples.
• Validated results on the MIT-BIH arrhythmia dataset, employing performance analysis and 1-D Jaccard's index for comparison.

Main Results:

• The proposed XAI method, using raw ECG signals and their derivatives, effectively incorporates temporal dependencies for improved classification explanation.
• Performance analysis confirmed that the explanations align with the underlying ML model's behavior.
• The 1-D Jaccard's index demonstrated the similarity between subsequences identified by interpretable models and XAI methods.

Conclusions:

• The developed XAI approach enhances the interpretability of ML-based heartbeat classification, addressing a key limitation in clinical application.
• The method's ability to capture temporal dependencies is vital for generating meaningful explanations.
• User study results suggest potential for visual explanations to serve as clinical diagnostic aids or training tools.