Ultrasound-based machine learning models for assisting the prediction of neonatal size and mode of delivery

Elective Cesarean surgeries (CSs) rates continue to rise worldwide, prompting renewed interest in predictors of prenatal biometrics, which have a central role in recommendations for elective CS.

To develop machine learning (ML) models for predicting neonatal anthropometric measures such as head circumference (HC), birth weight, and the mode of delivery, either vaginal or CS.

A retrospective single-center longitudinal cohort cross-sectional study conducted at a big public hospital.

Data were drawn from 5375 pregnant women who underwent routine prenatal ultrasound examinations within 2 weeks of delivery. Dataset curation included exclusion criteria and handling of missing data prior to model development. Formal feature selection for the most predictive variables resulted in the final dataset of 3447 subjects. Four supervised ML algorithms were implemented: stochastic gradient descent, random forest, K-nearest neighbors, and stacking ensemble (SE). The models were trained and evaluated on clinical and ultrasonographic data.

The predicted newborn weight (NBW) was of comparable accuracy to the commonly used Hadlock IV formula for the estimated fetal weight. The predicted newborn head circumference (NBHC) was of superior accuracy compared to the last prenatal ultrasound measurement. The classification of the delivery mode revealed a very close association between the predicted CSs and high values of NBHC and NBW.

We developed highly accurate ML-based models for prediction of NBHC, NBW, and the mode of delivery using only the three last prenatal ultrasound measurements: biparietal diameter, abdominal circumference, and HC. The predicted mode of delivery demonstrated a very good association between CSs and high values of NBHC and NBW. Future implementation of ML algorithms in risk-based obstetric management will benefit both maternal and fetal health and wellbeing.