Challenging Black-Box Models: Interpretable Explanations for ECG Classification.

Area of Science:

• Biomedical Engineering
• Machine Learning in Healthcare
• Cardiology

Background:

• Deep learning models excel in performance but often lack transparency, limiting their clinical adoption.
• Explainable AI (Artificial Intelligence) is essential for building trust and enabling real-world applications in medicine.
• Interpretable machine learning methods are needed to understand model decisions in healthcare.

Purpose of the Study:

• To propose and evaluate a logistic regression classifier for ECG analysis that provides interpretable feature importance.
• To demonstrate that non-deep learning approaches can achieve performance comparable to deep learning methods.
• To introduce opportunities for on-the-fly counterfactual explanations in clinical decision support.

Main Methods:

• Utilized temporal aligned Electrocardiograms (ECG) as input data.
• Developed a logistic regression classifier incorporating interpretable feature importance techniques.
• Compared the performance of the proposed model against established deep learning benchmarks.
• Implemented methods for generating on-the-fly counterfactual explanations.

Main Results:

• The logistic regression classifier achieved performance comparable to deep learning models.
• Interpretable feature importance provided insights into the model's decision-making process.
• The approach facilitated the generation of actionable counterfactual explanations.
• Code and models are publicly available for reproducibility.

Conclusions:

• Non-deep learning classifiers, such as logistic regression with interpretable features, can match deep learning performance in ECG analysis.
• Explainability and counterfactuals can be integrated into clinical AI without sacrificing performance.
• This work promotes the development of more transparent and trustworthy AI tools for healthcare.