ECG-DPSHAP: An Approach towards Privacy-preserving SHAP-based Explainable AI for 12-lead ECG Classification Model
Sirapop Nuannimnoi, Milzam Wafi Azhar, Kai-Hsiang Chang, Angger Baskoro, Jasmine Ariesta Dwi Pratiwi, ChingYao Huang · 2025
Deep learning (DL) methods have demonstrated immense potential in healthcare applications, particularly for medical diagnostics, where they enable automated and accurate analysis of complex physiological signals. However, integrating AI into clinical settings presents two fundamental challenges— interpretability and data privacy—both of which are essential for the ethical and responsible deployment of AI-driven healthcare solutions. Ensuring that AI models provide transparent, explainable decisions is crucial for clinician trust and patient safety, while simultaneously preserving sensitive patient data is imperative for compliance with privacy regulations. In this paper, we explore the inherent tension between explainability and privacy preservation in AI-powered medical diagnostics. We then introduce a case study on 12-lead electrocardiogram (ECG) classification and propose a novel approach, ECG-DPSHAP, which integrates differential privacy (DP) mechanisms with SHAP (SHapley Additive Explanations) to safeguard patient data while maintaining the interpretability of model predictions. Through this case study, we systematically analyze the trade-offs between privacy and explainability, identify best practices for balancing these objectives, and outline potential research directions that could further enhance the development of responsible AI solutions in healthcare. Our findings contribute to the ongoing discourse on AI ethics and provide insights into designing privacy-preserving yet interpretable models for real-world clinical applications.