ClusteredSHAP: Faster GradientExplainer based on K-means Clustering and Selections of Gradients in Explaining 12-Lead ECG Classification Model

The vast majority of healthcare systems are operating at or near their full capacity. Providing an inaccurate diagnosis is a further prevalent issue. Although it is common knowledge that physicians get significant training, it is nevertheless possible for them to misdiagnose patients, overlook warning signs, or commit any number of other human blunders. Deep Learning (DL) has emerged as a helpful tool in medical diagnostics, particularly in ECG (electrocardiogram) signal classification, enabling efficient and precise detection of cardiac abnormalities. However, the inherent black-box nature of DL poses challenges for its direct implementation in real-world medical settings. To address this, Explainable AI (XAI) techniques such as SHAP, LIME, and CAM have been introduced, aiming to render the complex decisions of neural networks interpretable. Among these, SHAP is recognized for its robustness and comprehensive explanation capabilities. Nonetheless, the high computational demands of traditional SHAP methods hinder their real-time application, especially in urgent medical scenarios. In this paper, we propose an optimized SHAP approach leveraging K-Means clustering to group gradients by importance, namely ClusteredSHAP. Our methodology focuses on a select cluster of high-magnitude gradients, enhancing the efficiency of the GradientExplainer. Our evaluation encompasses both the computational efficiency and the usability of the resulting explanations, through a custom questionnaire and explanation usability scores inspired by established user experience metrics and medical tool usability standards. The results show that our proposed ClusteredSHAP not only provides significantly faster explanations, but also achieves a similar level of average explanation usability scores with the GradientExplainer. Through this work, we strive to bridge XAI with clinical practice, ensuring timely, transparent, and effective patient care.