Machine Learning-Based Risk Stratification for Gestational Diabetes Management

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Obstetrics and Gynecology

Background:

• Gestational diabetes mellitus (GDM) diagnosis in late pregnancy limits intervention time.
• Effective GDM management is crucial for reducing maternal and fetal complications.
• Current risk stratification for GDM requires timely and accurate prediction of glycemic excursions.

Purpose of the Study:

• To develop and validate a machine learning system for stratifying GDM patients based on predicted high blood glucose levels.
• To identify patients at risk of hyperglycemia using daily glucose monitoring and electronic health record (EHR) data.
• To assess the generalizability of the predictive model across different healthcare settings.

Main Methods:

• Trained and validated regression models (linear, non-linear, tree-based) using daily blood glucose and EHR data.
• Utilized XGBoost for predicting the proportion of high glucose readings in GDM patients.
• Performed internal validation on 1148 pregnancies (Oxford University Hospitals) and external validation on 709 patients (Royal Berkshire Hospital).

Main Results:

• XGBoost demonstrated superior performance in predicting high glucose readings during internal validation (MSE: 0.021, R2: 0.482, MAE: 0.112).
• The model achieved comparable performance in the external validation cohort, indicating robustness.
• The findings suggest the model's potential for generalizable application in GDM management.

Conclusions:

• A machine learning-based system can effectively predict high blood glucose levels in GDM patients.
• The developed model shows promise for early identification of at-risk individuals, enabling timely interventions.
• The generalizability across different hospital cohorts supports its potential clinical utility in managing GDM.