AI-XAI-LLM: Interpretable insights into stroke risk prediction

Abstract

Early identification of individuals at stroke risk is critical for implementing timely preventive measures. Although machine learning models have demonstrated potential in predicting stroke risk, their clinical adoption is limited due to a lack of transparency. Explainable AI (XAI) techniques provide insights into predictions, but their technical metrics are often complex and difficult for clinicians to interpret. This research proposes an integrated AI-XAI-LLM pipeline that generates accurate, patient specific stroke predictions and provides post hoc explainability to highlight key factors influencing the prediction, which are then translated into clear, clinician-friendly narratives using a prompt engineered large language model. Evaluations show that this approach enhances prediction transparency and interpretability, fostering trust and encouraging the use of predictions to support decision-making.