Machine Learning Prediction Models for Gestational Diabetes Mellitus: Meta-analysis.

Area of Science:

• Endocrinology and Metabolism
• Reproductive Health
• Biostatistics and Machine Learning

Background:

• Gestational diabetes mellitus (GDM) is a prevalent endocrine metabolic disorder during pregnancy.
• Early screening has improved outcomes, but advanced prediction methods are needed.
• Machine learning (ML) models are emerging tools for GDM risk identification and early prediction.

Purpose of the Study:

• To conduct a meta-analysis of published prognostic models for GDM risk prediction.
• To compare the performance of different ML models in predicting GDM.
• To identify key predictors applicable to GDM risk models.

Main Methods:

• A systematic search of four electronic databases for ML prediction models for GDM in the general population.
• Assessment of model bias using the Prediction Model Risk of Bias Assessment Tool (PROBAST).
• Meta-analysis using Meta-DiSc software, including sensitivity, meta-regression, and subgroup analyses to address heterogeneity.

Main Results:

• Analysis of 25 studies involving women over 18 without prior vital disease.
• Pooled area under the receiver operating characteristic curve (AUROC) for ML models was 0.8492.
• Non-logistic regression models achieved a higher pooled AUROC (0.8891) compared to logistic regression (0.8151).
• Key predictors identified include maternal age, family history of diabetes, BMI, and fasting blood glucose.

Conclusions:

• ML methods offer an attractive alternative to current GDM screening strategies.
• Further emphasis on quality assessment and standardized diagnostic criteria is crucial for expanding ML model utility.
• The identified predictors can inform the development of more accurate GDM prediction models.