Evaluation of machine learning and logistic regression-based gestational diabetes prognostic models

Area of Science:

• Obstetrics and Gynecology
• Medical Informatics
• Epidemiology

Background:

• Gestational diabetes mellitus (GDM) prediction models require temporal evaluation to ensure accuracy.
• Changes in population demographics and GDM prevalence necessitate model updates and validation.

Purpose of the Study:

• To temporally evaluate existing GDM prediction models.
• To update GDM models as needed.
• To compare the performance of machine learning (ML) and regression-based GDM models over time.

Main Methods:

• Utilized a temporal validation dataset of 12,722 singleton pregnancies (2021-2022).
• Evaluated Monash GDM Logistic Regression (LR) and ML models (versions 2 and 3).
• Assessed model performance using discrimination (AUC), calibration, and decision curve analysis (DCA).

Main Results:

• All models demonstrated similar discrimination performance (AUCs ~0.73).
• Models showed overestimation, which improved with recalibration.
• All models provided better net benefits than treat-all or treat-none strategies.

Conclusions:

• GDM models remained robust after recalibration despite population characteristic changes.
• Original ML model performance significantly decreased during validation.
• Dynamic models are superior for adapting to temporal changes and calibration drift.