Data-driven Phenotypic Clusters of Gestational Diabetes Mellitus and Associations with Risk of Perinatal Complications and Postpartum Diabetes

OBJECTIVE Management of gestational diabetes mellitus (GDM) largely follows a uniform approach, despite growing recognition of GDM heterogeneity. We aimed to identify data-driven GDM clusters employing machine learning techniques using clinical data and assess their associations with perinatal complications and postpartum diabetes risk. RESEARCH DESIGN AND METHODS In a population-based cohort study, 37,544 individuals with GDM were followed up through 12 years postpartum. In the discovery (70%) and validation sets (30%), we applied dimension reduction and clustering methods using routinely available sociodemographic, behavioral, and clinical variables. Covariate-adjusted modified Poisson and Cox regression models assessed associations of GDM clusters with risk of perinatal complications and postpartum diabetes. RESULTS Four data-driven GDM phenotypic clusters were identified: Cluster 1 (C1; 65.6%), C2 (14.5%), C3 (12.0%), and C4 (7.8%) in the discovery set, with similar distributions in the validation set (C1-C4: 66.7%, 14.0%, 12.0%, 7.4%). C2-C4 compared to C1 (“Late-diagnosed, Lower-BMI and Post-load Hyperglycemia GDM”) were associated with higher risks of perinatal complications and new-onset postpartum diabetes, especially C4 [“Early-diagnosed, Comorbidity-related and High-Glucose-Challenge-Test GDM”: adjusted relative risks (aRR) for severe maternal morbidity: 1.43 (95% CI: 1.19-1.72); neonatal-intensive-unit-admission: 1.53 (1.41, 1.66); adjusted hazard ratio for diabetes: 4.32 (3.94-4.73)]. Within the largest C1, three subclusters were identified, with differential risks of perinatal complications, but not postpartum diabetes. CONCLUSIONS Our study identified distinct data-driven GDM phenotypic clusters with differential risks of perinatal complications and postpartum diabetes. These findings may inform personalized risk assessment and management strategies tailored to GDM phenotypic clusters to possibly reduce adverse health outcomes.