Lipid metabolism-based machine learning models for predicting large for gestational age in non-diabetic pregnancies

The majority of LGA occurs in normoglycemic pregnancies, yet prediction models integrating lipid metabolism for this population are lacking. This study aimed to develop and validate machine learning models based on lipid metabolic profiles to predict LGA risk in non-diabetic pregnant women.

We retrospectively analyzed 1,968 singleton pregnancies without diabetes. Lipid profiles were assessed at early pregnancy (11-14 weeks) and mid-pregnancy (20-24 weeks). The dataset was randomly divided into training (n = 1,574) and independent test sets (n = 394) using stratified sampling. Four machine learning algorithms-logistic regression (LR), random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost)-were developed using nested 5×5 cross-validation to prevent overfitting and ensure unbiased performance estimation. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), and feature importance was assessed using SHAP analysis.

Hyperparameter optimization demonstrated stable convergence across outer folds (coefficient of variation < 1.5% for all models). On the independent test set, XGBoost achieved the highest performance with an AUC of 0.776 (95% CI: 0.730-0.822), significantly outperforming LR (AUC 0.688, P < 0.001). SHAP analysis identified pre-pregnancy BMI (mean SHAP value: 1.48), gestational weight gain (0.93), and mid-pregnancy triglycerides (0.73) as the strongest predictors. Lipid metabolism-related parameters collectively accounted for approximately 40% of the model's predictive capacity.

Machine learning models incorporating lipid metabolic profiles demonstrated superior performance in predicting LGA among non-diabetic pregnancies. This approach enables early risk stratification and may guide personalized prenatal management.