Interpretable ECG Analysis for Myocardial Infarction Detection through Counterfactuals

In the evolving landscape of ECG signal analysis, the opaque nature of deep learning models poses challenges to their integration into clinical practice. This study aims to address these challenges by investigating the application of counterfactual explanations to make machine learning models more interpretable for clinicians, particularly in the context of differentiating control group subjects from myocardial infarction patients. Utilizing the PTB-XL dataset, we developed a methodology for systematic feature extraction and refinement to prepare for counterfactual analysis. This led to the creation of the Visualizing Counterfactual Clues on Electrocardiograms (VCCE) method, designed to improve the practicality of counterfactual explanations in a clinical setting. The validity of our approach was assessed using custom metrics that reflect the diagnostic relevance of counterfactuals, evaluated with the help of two cardiologists. Our findings suggest that this approach could support future efforts in using ECGs to predict patient outcomes for cardiac conditions, achieving interpretation validity scores of 23.29 $\pm$ 1.04 and 20.28 $\pm$ 0.99 out of 25 for high and moderate-quality interpretations, respectively. Clinical alignment scores of 0.83 $\pm$ 0.12 for high-quality and 0.57 $\pm$ 0.10 for moderate-quality interpretations underscore the potential clinical applicability of our method. The methodology and findings of this study contribute to the ongoing discussion on enhancing the interpretability of machine learning models in cardiology, offering a concept that bridges the gap between advanced data analysis techniques and clinical decision-making. The source code for this study is available at https://github.com/tanyelai/vcce.