Development and evaluation of machine learning models for diabetes risk prediction using BRFSS 2023 data

Diabetes remains a major global health burden, with its rising prevalence demanding scalable approaches to early risk detection. Traditional clinical screening is resource-intensive, whereas survey-based modelling offers a complementary tool for population-level surveillance. In this study, we develop and evaluate machine learning classifiers on the 2023 U.S. Behavioural Risk Factor Surveillance System (BRFSS) to predict self-reported diabetes and prediabetes. Following a comprehensive computational comparison, three models—elastic net logistic regression, gradient boosting, and XGBoost—were selected for hyperparameter optimisation, probability calibration, and recall-oriented threshold tuning. Given that diabetes risk prediction prioritises minimising false negatives (recall) over false positives (precision), we emphasised PR-AUC as the primary metric, with AUROC as a secondary measure. XGBoost and gradient boosting achieved the strongest overall performance (PR-AUC ≈ 0.48, AUROC ≈ 0.83), and emerged as reliable models, balancing discrimination, calibration, and sensitivity for practical screening applications while elastic net maximised sensitivity at the cost of precision.