Development and validation of super learner models to predict small and large for gestational age in the second...

Area of Science:

• Perinatal epidemiology
• Reproductive health
• Machine learning in obstetrics

Background:

• Accurate prediction of fetal growth deviations, small for gestational age (SGA) and large for gestational age (LGA), is crucial for optimizing perinatal outcomes.
• Current prediction models using antenatal data, especially for nulliparous women, demonstrate suboptimal performance.

Purpose of the Study:

• To develop and validate predictive models for SGA and LGA by integrating intergenerational data (grandmaternal and maternal factors) with maternal clinical information.
• To assess the incremental predictive value of incorporating "G0 predictors" (grandmaternal pregnancy history and maternal birth characteristics) into models with "G1 predictors" (maternal clinical factors at 26 weeks gestation).

Main Methods:

• Utilized a retrospective cohort of singleton births to nulliparous women in Nova Scotia, Canada (1981-2011).
• Employed Super Learner, an ensemble machine learning algorithm, to build predictive models.
• Validated models using nested cross-validation, assessing discrimination (AUC-ROC, AUC-PR) and calibration.

Main Results:

• Models incorporating G0 predictors demonstrated improved discrimination compared to G1-only models (e.g., AUC-ROC increased from 0.66 to 0.69 for SGA).
• Precision-recall curves also showed modest improvements with the inclusion of G0 predictors.
• Models combining both G0 and G1 predictors were well-calibrated.

Conclusions:

• Incorporating intergenerational information (grandmaternal and maternal birth data) offers a modest improvement in predicting SGA and LGA in nulliparous women.
• Despite enhancements, the overall predictive performance for fetal growth deviations remains suboptimal, highlighting the need for further research and improved methodologies.