Interpretable AI for Stroke Prediction: A Structured Approach Using Explainable AI Techniques

The lack of interpretability in AI-driven healthcare diagnostics poses a significant challenge to clinical adoption. This study explores the methodological integration of explainable artificial intelligence (XAI) tools using open clinical prediction dataset, the SCI-XAI pipeline, for stroke risk prediction. We apply multiple machine learning models ranging from white-box approaches (Logistic Regression, Decision Tree, Explainable Boosting Machine) to black-box models (Random Forest, XGBoost, LightGBM, and Multi-Layer Perceptron) and evaluate their trade-offs between predictive accuracy and explainability using techniques such as SHAP, LIME, and ELI5. The study uses a systematic approach involving pre-modeling, modeling, and post-modeling phases, aiming to improve model interpretability for potential use in clinical decision-support contexts. The experimental results show that ensemble models achieve superior accuracy, while traditional models provide inherent transparency. However, the SCI-XAI framework demonstrated that post-hoc explainability tools can extend such transparency to complex models. SHAP-based feature im-portance analysis identifies age, glucose levels, and BMI as the most influential predictors of stroke. The integration of structured explainability into AI based diagnostics helps bridge the gap between algorithmic prediction and clinical interpretability, offering a methodological foundation for more transparent decision-support systems.