Constructing small for gestational age prediction models: A retrospective machine learning study.

Area of Science:

• Perinatal medicine
• Machine learning in healthcare
• Predictive analytics in obstetrics

Background:

• Small for gestational age (SGA) infants are at increased risk for adverse outcomes.
• Accurate prediction of SGA is crucial for timely intervention.
• Existing prediction methods may not fully leverage data from various pregnancy stages.

Purpose of the Study:

• To develop and compare machine learning models for predicting SGA.
• To evaluate the predictive performance of models using data from different pregnancy stages.
• To identify key predictive variables for SGA.

Main Methods:

• Retrospective study of 4,394 singleton pregnancies.
• Data categorized into four gestational time points.
• LightGBM framework with cross-validation for variable importance.
• Seven machine learning algorithms used for model development.
• Performance evaluated using ROC analysis and sensitivity.

Main Results:

• 148 (3.4%) SGA infants identified.
• Maternal height, age, and pre-pregnancy weight were key features.
• Models using admission stage variables demonstrated strong predictive performance (AUC > 0.8).
• Best model achieved an AUC of 0.85 and 73% sensitivity at 10% FPR.

Conclusions:

• Machine learning models show good predictive performance for SGA across pregnancy stages.
• The prediction model using admission stage variables performed best.
• Prenatal physical examinations are significant for SGA prediction.