A large-scale prediction model to predict large for gestational age infants conceived by IVF/ICSI

To develop and internally validate a machine learning model for predicting the risk of large-for-gestational-age (LGA) birth following IVF/ICSI and to identify important parental and treatment-related predictors.

A total of 17, 741 singleton live births resulting from IVF/ICSI, categorized as appropriate for gestational age (AGA) or LGA. Data on birth outcomes, parental characteristics, and treatment-related variables were collected. The dataset was randomly divided into training and testing sets (7:3). An XGBoost model was developed and optimized using the Optuna framework and Tree-structured Parzen Estimator algorithm. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), calibration analysis, Brier score, and classification metrics. Logistic regression was used as a conventional baseline model. Model interpretability was assessed using SHAP analysis.

The XGBoost model achieved an AUC of 0.7003 in the internal hold-out test set, compared with 0.6445 for logistic regression. Calibration analysis showed a lower Brier score for XGBoost than for logistic regression (0.2040 vs. 0.2295). SHAP analysis identified embryo transfer strategy, maternal anthropometric and metabolic characteristics and several paternal characteristics as important predictors in the model.

The machine learning model demonstrated moderate discriminative ability for predicting LGA risk among IVF/ICSI-conceived singleton births in internal validation. The findings highlight the dominant association of embryo cryopreservation strategies, maternal and paternal factors with fetal overgrowth. While the model shows acceptable discriminatory ability, its clinical utility requires future prospective external validation.