Explainable AI for Diabetes Risk Assessment: Enhancing Clinical Decision Support with Interpretable Machine Learning Models

This paper presents a comprehensive study on the application of Explainable Artificial Intelligence (XAI) for diabetes risk assessment, focusing on the interpretability of machine learning models in clinical decision support systems. While machine learning has demonstrated high accuracy in predicting diabetes, the lack of transparency in decision-making processes limits its adoption in healthcare. We employ interpretable models and model-agnostic explanation techniques such as SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-Agnostic Explanations) to enhance the understanding of predictive outcomes. Using real-world datasets including the PIMA Indians Diabetes Dataset and NHANES, we evaluate both the performance and explainability of various models. Our experiments show that the XGBoost model achieved the highest prediction accuracy of 86.5%, with an AUC-ROC of 0.91, followed by the Neural Network (MLP) with 85.2% accuracy and AUC-ROC of 0.90. Interpretable models like Logistic Regression and Decision Tree reported lower accuracies of 78.2% and 81.0% respectively, but offered high interpretability. The combination of these predictive models with SHAP and LIME enhances clinical applicability by aligning model reasoning with physician insight. Case evaluations demonstrated that LIME's local explanations matched physician reasoning in 87% of tested scenarios.➢ Integration of SHAP and LIME in both training and clinical phases➢ Case-wise physician matching study (87%)➢ Performance-explainability tradeoff analysis across 6 ML models