Early prediction of gestational diabetes mellitus with clinical characteristics, cell-free DNA and genetic variants

BACKGROUND: Gestational diabetes mellitus (GDM) remains a prevalent and heterogeneous pregnancy complication with limited strategies for early identification. We aimed to investigate efficient approaches for early prediction of GDM with clinical and genetic risk factors. METHODS: A previously developed machine-learning model based on clinical characteristics achieved an area under the curve (AUC) of 0.77. To improve predictive accuracy, we further collected non-invasive prenatal testing (NIPT) results from 595 pregnant women (295 with GDM, 300 without). A cumulative polygenic risk score (PRS) was calculated using 1,170 selected single nucleotide variants (SNVs). Logistic regression, support vector machines, random forest, decision tree, linear model and naïve Bayes machine learning models were employed. External validation was performed with an additional 2,350 blood samples independently collected from two other centers. RESULTS: Logistic regression analysis showed that the PRS alone achieved an AUC of 0.75 for GDM discrimination. From cell-free DNA (cfDNA) sequencing performed during NIPT, we identified 357 gene transcripts with differential coverage at transcription start sites. A cfDNA-based linear model achieved an AUC of 0.83 using a subset of 166 signature genes, which reached 0.85 when combined with clinical features. Integration of clinical features, cfDNA, and SNVs yielded the highest performance using a random forest model (AUC = 0.89, specificity = 0.74, sensitivity = 0.89). For external validation, a clinically practical model incorporating clinical features and cfDNA achieved an AUC of 0.83 using linear approach. CONCLUSIONS: Our GDM prediction model has reached high accuracy fully using accessible clinical and genetic data routinely generated from current antenatal testing, enabling early screening and interventions for women at risk.