Development and temporal validation of an early-pregnancy prediction model for gestational diabetes mellitus using routine clinical and biochemical markers: a retrospective cohort study

To develop and temporally validate an early-pregnancy prediction model for gestational diabetes mellitus (GDM) using routinely available indicators. In this retrospective single-center study, women who delivered from May 2020 through December 2024 formed the development cohort, and those who delivered from January through April 2025 formed the temporal validation cohort. Continuous functional forms were examined with restricted cubic splines, followed by 10-fold cross-validated group LASSO selection. The lambda.1se model was refitted using logistic regression, internally validated using 1000 bootstrap resamples, and evaluated in the temporal validation cohort. Discrimination, calibration, and clinical utility were assessed using receiver operating characteristic curves, calibration assessment, and decision curve analysis. The development cohort included 4859 women with 951 GDM (19.6%), and the temporal validation cohort included 1179 women with 299 GDM (25.4%). The final model retained age, pre-pregnancy body mass index (BMI), family history of diabetes, fasting plasma glucose (FPG), triglycerides, total cholesterol, gamma-glutamyl transferase, uric acid, and creatinine. The development AUC was 0.728 (95% CI 0.710–0.746), and the optimism-corrected AUC was 0.721. The temporal validation AUC was 0.764 (95% CI 0.733–0.796), with a Brier score of 0.156, a calibration intercept of 0.442, a calibration slope of 1.208. Random forest, XGBoost, and LightGBM did not improve temporal discrimination over logistic regression. A nine-predictor model showed moderate discrimination and retained performance in temporal validation. The model may support exploratory early risk stratification.