Machine Learning-Based Prediction of Diabetes Using Patient Health Records

Abstract

Diabetes mellitus is a serious and increasing burden on global health, with its related complications being largely preventable by early detection and treatment. This study proposes and rigorously evaluates machine-learning models to predict diabetes risk from standard patient health record data. We benchmark eight classifiers-Logistic Regression, Random Forest, Support Vector Machine, K-Nearest Neighbours, Multilayer Perceptron Classifier XGBoost, Gradient Boosting and Gaussian Naive Bayes-on the Pima Indians Diabetes Dataset (768 instances, 8 clinical features and moderate 65:35 class imbalance) within a strictly leakage-free pipeline in which median imputation and standardization on are fitted only to the training position (never on the test group), SMOTE oversampling is applied only with respect to cross-validation folds.Approach: Models are tuned using 5-fold stratified GridSearchCV (scoring: ROC-AUC) and assessed on a hold-out set of 154 not seen patients. The best point performance (ROC-AUC = 0.9476, bootstrap 95% CI [0.9096, 0.9763]) was obtained with XGBoost. By contrast, when looking at its error rate using McNemar's paired significance tests, they are not significantly different from Gradient Boosting or Random Forest, as a result defining a statistically indistinguishable "top tier" of tree-ensemble methods rather than single-model domination. SHAP analysis suggests Insulin, Glucose, BMI and Age to be the dominating drivers as you would expect from clinical knowledge of diabetes risk factors.