Selective screening for gestational diabetes mellitus (GDM) remains a widespread strategy. Variation in the criteria identifying at-risk women questions its accuracy, with implications for clinical outcome and resource allocation. Our aim was to develop and externally validate a multivariable prediction model with improved performance compared to the current French pre-screening selection strategy.

Data was derived from the population-based 2021 (derivation sample) and 2016 (external validation sample) French National Perinatal Survey (ENP). Independent predictors of GDM were identified using a multivariable logistic regression model. Predictive performance was assessed through the area under the receiver operating characteristic curve. Diagnostic performance was assessed through sensitivity, specificity, and accuracy. Sensitivity analyses were conducted: (1) in maternity centers with quasi-universal screening, (2) outcome strictly defined as GDM cases associated with large-for-gestational-age births and (3) implementing doubly robust estimators of sensitivity and specificity.

The study population included 10,834 women in the derivation sample and 11,633 in the validation sample, where the prevalence of GDM was respectively 19.4% and 13.2%. Maternal age, body mass index, obstetric history, family history of diabetes, and maternal country of birth were independent predictors of GDM. The prediction model demonstrated a statistically significant improvement in specificity (0.49 [95% CI: 0.48–0.50] vs. 0.46 [95% CI: 0.45–0.47]) and overall diagnostic accuracy (0.53 [95% CI: 0.52–0.54] vs. 0.50 [95% CI: 0.49–0.51]).

The prediction model modestly improved performance while maintaining the same screening rate, though it is unlikely to justify additional complexity of implementation, therefore limiting its added-value in clinical practice. These findings suggest that available clinical predictors already capture most of the predictive information relevant for selective screening.