Interpretability of Machine Learning Models for Cardiac Arrhythmia Classification

Context: Cardiovascular diseases, particularly cardiac arrhythmias, cause global mortality. Electrocardiograms (ECG) are essential for diagnoses. However, the automated analysis of ECG signals using machine learning faces interpretability challenges, which are crucial for accepting these models in medical practice. Problem: The classification of cardiac arrhythmias using machine learning faces resistance in the medical field due to the black-box nature of these models, which hinders the understanding of decisions and reduces professional trust. There is a need for interpretable models that reveal the critical factors in diagnosis, fostering the reliable use of artificial intelligence in healthcare. Solution: This work proposes an interpretable model for classifying cardiac arrhythmias, using machine learning to identify and visually explain the ECG features contributing to the diagnosis, with explanations at the model, class, and specific signal levels. IS Theory: This work is based on the Technology Acceptance Model (TAM), which suggests that "perceived usefulness" and "perceived ease of use" influence the intention to use a system. In healthcare, these perceptions relate to the model’s ability to provide clear and helpful explanations, facilitating medical professionals’ adoption of artificial intelligence (AI). Method: Instead of using raw biosignals, extracted ECG features are employed to enhance interpretability. This approach provides model-agnostic explanations at both local and global levels. Interpretability techniques are applied to clarify the contribution of each feature to the diagnosis. Summary of Results: Features such as the variability and median of RR and PR intervals and the signal-to-noise ratio of ECG signals are crucial for accurate arrhythmia classification. Contributions and Impact on IS: The approach allows for understanding the influence of specific ECG features on diagnosis, helping to identify patterns that support classification decisions and promoting the adoption of AI with trust and responsibility in medical practice.