A Machine Learning Approach to Predicting Diabetes Using Healthcare Data

Using the Pima Indians Diabetes dataset, this study evaluates multiple supervised learning models for diabetes risk prediction using a nested, stratified cross-validation protocol to mitigate optimistic bias. Logistic Regression, Random Forest, Decision Tree, and XGBoost classifiers are compared using discrimination, calibration, and statistical significance testing. The results indicate that Logistic Regression and Random Forest achieve comparable ROC–AUC performance, while Logistic Regression provides the best probabilistic calibration. Model interpretability is assessed using SHAP, revealing glucose concentration, body mass index, and age as the dominant risk factors. These findings highlight the importance of rigorous evaluation and explainability when applying machine learning models in healthcare screening tasks.