Noninvasive retinal features for early prediction of gestational diabetes mellitus: a machine learning-based cohort study

Abstract

Gestational diabetes mellitus (GDM) is the most common complication during pregnancy and early prediction for high-risk gravidas is crucial to facilitate timely intervention. However, most newly discovered biomarkers require additional blood sampling and high costs, which limits their feasibility in routine clinical practice. This study aimed to assess whether noninvasive retinal parameters could enhance the predictive performance of established models for GDM. This prospective cohort study collected demographic characteristics, glycolipid metabolism indices, and retinal images at 11–13 +6 weeks of gestation. The primary outcome was GDM based on oral glucose tolerance test at 24–28 weeks. Variables were selected via least absolute shrinkage and selection operator (LASSO) regression. Six machine learning algorithms (logistic regression, random forest, extreme gradient boosting, categorical boosting [CatBoost], adaptive boosting, and support vector machine) were employed to develop the LIGHT (LIpid+Glucose+opHthalmic+maTernal factors) model incorporating baseline, glycolipid metabolism and retinal features. Model performance was assessed by the area under the receiver-operating-characteristic curve (AUC) and interpreted by the Shapley Additive Explanation method (SHAP). The incremental value of retinal features was evaluated by net reclassification improvement (NRI) and integrated discrimination improvement (IDI). We compared LIGHT model with (1) Baseline model based on demographic characteristics, (2) Glycolipid model including baseline and glycolipid metabolism features, and (3) Eye model using baseline and retinal features. Of the 2114 participants, 1774 pregnancies were included in the final analysis, of which 324 (18.3%) were GDM. Sixteen variables including five baseline variables, three glycolipid metabolism variables, and eight retinal features were selected for model development via LASSO regression. Among the six machine learning algorithms, CatBoost showed the best performance. The SHAP plot showed that the six most important variables of LIGHT model were body mass index, triglyceride glucose index, age, hemoglobin A1c, std of angle deviation and angle-based tortuosity. The LIGHT model achieved an AUC of 0.762 (95% CI 0.703–0.821), which was significantly higher than the Baseline (AUC 0.687, 95% CI 0.618–0.756) and the Eye model (AUC 0.619, 95% CI 0.540–0.698). Although AUC improvement over the Glycolipid model (AUC 0.716, 95% CI 0.651–0.781) was not statistically significant, the LIGHT model showed positive event reclassification (event-NRI 0.719), with a modest overall continuous NRI (− 0.198) and IDI (− 0.039). The LIGHT model demonstrates that noninvasive retinal features can enhance early GDM risk stratification, offering a potential complementary approach for prenatal screening.