Causal Insights into Gestational Diabetes Mellitus

Sheresh Zahoor^1*Anthony C. Constantinou²Fiona O'Halloran¹Louise O’Mahony³Mairead O'Riordan³Oratile Kgosidialwa³Linda Culliney³Mohammed Said Alhajri¹Mohammed Hasanuzzaman⁴

• ¹Munster Technological University, Cork, Ireland
• ²Queen Mary University, London, United Kingdom
• ³Cork University Hospital, Cork, Ireland
• ⁴Queens University Belfast, Belfast, United Kingdom

Gestational diabetes mellitus (GDM), defined by the onset of hyperglycaemia during pregnancy, remains the most prevalent metabolic complication in pregnancy and is associated with increased risks of adverse maternal, neonatal, and long-term metabolic outcomes. This study analysed a clinically curated dataset from patients diagnosed with GDM at a major Irish maternity hospital over a defined study period (2014–2016, 2020), with the aim of identifying potential causal relationships that could support more targeted and effective interventions. A knowledge graph was first constructed, integrating clinical expertise, established literature, and insights generated using the GPT-4 large language model. To complement this knowledge-driven framework, 20 structure learning algorithms were applied to independently infer Causal Bayesian Networks (CBNs). To account for variability across individual models, a model-averaging approach was employed to generate a consensus-based causal structure. This integrative model offered a more stable representation of underlying relationships and produced quantifiable insights that may enhance clinical decision-making. Clinicians involved in the study emphasised that the ability to quantify these relationships improved confidence in patient care strategies and supported more personalised, evidence-based practice. Key findings from the model-averaged CBN highlighted critical pathways in GDM management, including the influence of birth weight on neonatal intensive care unit (NICU) admissions and the impact of dietary intervention on maternal glucose regulation. Sensitivity analysis confirmed birth weight, gestational age at delivery, and mode of delivery as major determinants of maternal and neonatal outcomes. Non-modifiable factors, such as a history of multiple pregnancies and prior GDM, also contributed to risk stratification, underscoring the value of early screening and tailored care. This study applied structure learning techniques to observational clinical data to identify clinically relevant relationships, providing a basis for generating hypotheses that could refine intervention strategies and improve patient outcomes in GDM care.