GDM

GDM predicts the risk of gestational diabetes mellitus (GDM) in early pregnancy among Chinese women using a machine-learning model.

Key Features:

• Algorithm: Uses extreme gradient boosting (XGBoost) as the primary machine learning algorithm.
• Cohort: Trained on a population-based prospective cohort of 19,331 pregnant women registered before the 15th gestational week in Tianjin, China (October 2010–August 2012).
• Data split: Dataset was divided into training (70%) and test (30%) sets for model development and validation.
• Predictors: Incorporates risk factors including pre-pregnancy body mass index, maternal age, fasting plasma glucose at registration, and alanine aminotransferase levels.
• Comparison: Directly compared with logistic regression models for benchmarking predictive performance.
• Calibration: Demonstrated better calibration with predicted probabilities closely matching observed outcomes in the test set.
• Discrimination: Achieved higher discrimination with an area under the receiver operating characteristic curve (AUR) of 0.742 versus 0.663 for the logistic model (p < 0.001).
• Output: Produces individualized risk scores for GDM in early pregnancy.

Scientific Applications:

• Early risk prediction: Predicts individual risk of gestational diabetes mellitus in early pregnancy among Chinese women.
• Risk stratification: Identifies women at high risk for GDM to inform timely clinical interventions.
• Model benchmarking: Serves as a comparison to traditional logistic regression for evaluating machine-learning performance in clinical prediction.
• Clinical predictive modeling: Integrates clinical biochemical and demographic predictors to enhance predictive models for maternal and fetal health outcomes.

Methodology:

Modeling employed extreme gradient boosting (XGBoost) trained and validated on a 70%/30% split of a 19,331-participant prospective cohort, using predictors including pre-pregnancy body mass index, maternal age, fasting plasma glucose at registration, and alanine aminotransferase; performance was assessed by calibration and discrimination (AUR compared to logistic regression, 0.742 vs 0.663, p < 0.001).