Development of an Explainable-AI Enabled Decision Support System for Improved Risk Assessment of Atrial Fibrillation in Cardiac Patients during Hospital Stay

Cardiovascular disease (CVD) is the primary cause of hospitalization and mortality worldwide, implying a critical burden on the healthcare system. Enhancing CVD risk assessment requires the integration of heterogeneous data sources to provide accurate, robust, and explainable predictions. This study focuses on developing an explainable artificial intelligence decision support system to predict the risk of in-hospital postoperative atrial fibrillation (AF). The use case was selected through extensive discussions and strong collaboration with healthcare professionals from different centers to be aligned with clinical needs and to provide practical applicability, AF being the most common complication after a cardiac surgery. The proposed pipeline includes data preprocessing, feature extraction, feature selection, model training, and explainability analysis, ensuring that methods are transferable from research to practice. A retrospective Italian dataset of 2,445 patients admitted to hospital following an acute myocardial infarction (AMI) was analyzed, incorporating clinical and ECG-derived features. Explainable AI (XAI) techniques such as SHAP and MDI were employed to provide interpretable insights, which are visualized through a user-friendly software framework tailored to support clinical decision-making. The performance of these models will be cross-validated with Finnish data as well as prospective Italian data. The system's implementation balances performance and accessibility, aiming to facilitate wide applicability across diverse populations and healthcare settings. Moreover, Ethical Legal and Societal Aspects (ELSA) interviews have been conducted to ensure patient and clinician acceptance of AI-driven CVD risk assessment.Clinical Relevance - This study presents an AI-driven decision support system, addressing a well-defined clinical use case, that integrates multi-modal data and explainability techniques to enhance personalized CVD risk assessment and bridge the gap between research and clinical practice, while also taking into account Ethical, Legal and Societal aspect.