Development of a lipid-glycaemic prediction model for gestational diabetes mellitus: a single-Centre retrospective study.

BACKGROUND

Gestational diabetes mellitus (GDM) is a common pregnancy complication. Although lipid parameters, the triglyceride-glucose (TyG) index, and TyG-body mass index (TyG-BMI) have been associated with GDM, their relative value within an integrated lipid-glycaemic prediction framework remains unclear. This study developed and internally validated an interpretable model for GDM risk estimation.

METHODS

This single-centre retrospective study included 287 pregnant women who underwent 75 g oral glucose tolerance testing and fasting lipid assessment at 24-28 gestational weeks, including 72 with GDM and 215 without GDM. Candidate predictors comprised maternal characteristics, fasting plasma glucose (FPG), routine lipid fractions, apolipoproteins, the TyG index, and TyG-BMI. Forward stepwise likelihood-ratio logistic regression was used to develop the original combined model. A structured alternative model replaced pre-pregnancy BMI, FPG, and triglycerides with TyG-BMI. Model performance was evaluated using receiver operating characteristic analysis, calibration measures, threshold-dependent diagnostic metrics, 1,000-resample bootstrap internal validation, paired DeLong testing, a nomogram, and decision curve analysis.

RESULTS

Eight variables were retained in the original combined model: maternal age, pre-pregnancy BMI, FPG, total cholesterol, triglycerides, non-high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, and apolipoprotein B. Its apparent and optimism-corrected AUCs were 0.848 (95% CI: 0.786-0.909) and 0.830, respectively; the bootstrap-corrected calibration intercept was -0.03, calibration slope was 0.86, and Brier score was 0.147. The TyG-BMI alternative model had apparent and optimism-corrected AUCs of 0.831 (95% CI: 0.765-0.897) and 0.817. The AUC difference was 0.017 (95% CI: -0.011 to 0.045; P = 0.235). Decision curve analysis indicated potential net benefit for both models across clinically relevant thresholds.

CONCLUSION

Both models showed good internally validated discrimination and acceptable calibration for GDM risk estimation at routine screening. The original model had a numerically higher AUC, but discrimination did not differ significantly between models. External validation is required before clinical implementation.