Opening the black box: interpretability of machine learning algorithms in electrocardiography.

Area of Science:

• Cardiology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep learning (DL) models show promise for detecting cardiac abnormalities from electrocardiograms (ECGs).
• A key limitation of current DL models is their lack of interpretability, hindering clinical trust and adoption.
• Understanding which ECG components DL models use for classification is crucial for validation.

Purpose of the Study:

• To develop and validate novel frameworks for interpreting DL-based 12-lead ECG classification.
• To identify specific ECG waves (P-wave, QRS complex, T-wave) most influential in DL model predictions.
• To assess the clinical relevance of features utilized by DL algorithms in diagnosing cardiac conditions.

Main Methods:

• Designed two new interpretability frameworks compatible with any DL model.
• Tested frameworks on a Deep Neural Network trained for classifying 24 cardiac abnormalities from 12-lead ECGs.
• Evaluated the frameworks' ability to pinpoint relevant ECG segments and waves contributing to classification decisions.

Main Results:

• The proposed frameworks successfully identified the most relevant ECG waves for DL-based cardiac abnormality classification.
• In many cases, the DL model's focus aligned with cardiologists' diagnostic criteria.
• Discrepancies were observed, indicating potential reliance on non-traditional features by the DL model.

Conclusions:

• The developed frameworks enhance the interpretability of DL models for ECG analysis.
• These tools can verify if DL models utilize clinically significant features, thereby increasing trust in AI-driven cardiac diagnostics.
• The findings contribute to advancing computational methods in cardiovascular physiology.