Towards Reliable and Transparent Diabetes Diagnosis with an Explainable Ensemble Framework for Clinical Decision Support

Diabetes mellitus (DM) is a significant global health concern that requires early diagnosis and effective management. This study presents an explainable machine learning framework for diabetes prediction and clinical decision support, integrating ensemble learning with multiple Explainable Artificial Intelligence (XAI) techniques to improve predictive accuracy and model transparency. The framework was evaluated using the Diabetes Binary Health Indicators dataset, comprising 253,680 records and 22 clinical and demographic attributes. Data preprocessing involved missing-value imputation, feature scaling, and the Synthetic Minority Oversampling Technique (SMOTE) to address class imbalance. Several machine learning algorithms, including Random Forest, XGBoost, LightGBM, Support Vector Machines, Logistic Regression, and Naive Bayes, were evaluated. A soft-voting ensemble combining Random Forest, XGBoost, and LightGBM achieved a test accuracy of 92.80% and an ROC-AUC of 0.975. To enhance interpretability, the study employed multiple XAI techniques, including SHAP, LIME, Explainable Boosting Machines (EBM), Anchors, counterfactual explanations, and partial dependence plots. These methods provided global and individual-level explanations, identifying influential predictors such as body mass index (BMI), age, general health, income, and physical activity. Explanation quality was further assessed using fidelity, faithfulness, sparsity, stability, and consistency metrics. The findings demonstrate the potential of combining ensemble machine learning and explainable AI to develop accurate and transparent diabetes prediction systems. The proposed framework provides a foundation for AI-assisted clinical decision support while highlighting the need for further validation using diverse clinical datasets before real-world healthcare deployment.