Explainable Machine Learning for Diabetes Screening Using NHANES 2021–2023 Data: A Leakage-Controlled Classification Study

Screening models for diabetes can look deceptively strong when the same variables used to define the label are also allowed into the predictors. To avoid that shortcut, we built a leakage-controlled diabetes screening pipeline using National Health and Nutrition Examination Survey (NHANES) 2021–2023 data. Diabetes status was assigned from physician diagnosis and standard laboratory thresholds, while glycated hemoglobin (HbA1c), fasting plasma glucose, and diabetes questionnaire items were excluded from model inputs. The final cohort contained 5770 adults, including 1249 diabetic participants, and 43 non-glycemic predictors covering demographic, anthropometric, cardiometabolic, lifestyle, and medical history information. Fourteen classifiers were compared under one preprocessing and validation workflow, with evaluation based on Receiver Operating Characteristic—Area Under the Curve (ROC–AUC), Area Under the Precision–Recall Curve (PR–AUC), threshold-dependent metrics, calibration, decision curve analysis, SHapley Additive exPlanations (SHAP) interpretation, and a composite General Performance Index (GPI). To make the leading model comparison fairer, Gradient Boosting, Random Forest, XGBoost, LightGBM, and CatBoost were tuned using the same 15-configuration randomized search budget with five-fold stratified cross-validation. After this analysis, XGBoost had the highest observed test ROC–AUC (0.891) and GPI (0.772), although paired comparisons showed no statistically significant ROC–AUC or PR–AUC difference between Gradient Boosting and any of the other four leading models. A non-laboratory analysis achieved a ROC–AUC of up to 0.879, and temporal validation with NHANES 2017–March 2020 yielded ROC–AUC = 0.851 for the harmonized Gradient Boosting reference model, compared with 0.888 on the corresponding 2021–2023 internal test set. The most influential predictors remained clinically plausible, including age, waist circumference, lipid markers, blood pressure, BMI, and renal/metabolic indicators.