Predicting preterm birth using electronic medical records from multiple prenatal visits.

Area of Science:

• Obstetrics and Gynecology
• Perinatal Medicine
• Machine Learning in Healthcare

Background:

• Preterm birth is a leading cause of neonatal morbidity and mortality.
• Accurate prediction of preterm birth is crucial for timely intervention and improved outcomes.
• Nulliparous women represent a key population for preterm birth risk assessment.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting preterm birth in nulliparous women.
• To assess the impact of incorporating data from different prenatal visit timings on predictive performance.
• To determine the added value of ultrasound measurements in enhancing preterm birth prediction.

Main Methods:

• Utilized the Nulliparous Pregnancy Outcomes Study: New Mothers-to-Be (nuMoM2b) dataset (n=8,830).
• Developed elastic net regularized logistic regression models using data from three prenatal visits (early, mid, and late gestation).
• Evaluated model performance using 5-fold cross-validation, focusing on Area Under the Curve (AUC), sensitivity, and specificity. Incorporated ultrasound data (cervical length, Pulsatility Index) in later models.

Main Results:

• Model performance improved with data from later prenatal visits, with AUC increasing from 0.6161 to 0.7087.
• Addition of ultrasound measurements significantly enhanced predictive ability.
• The final model achieved high sensitivity for predicting very preterm (0.8254) and extreme preterm (0.9295) births at the third prenatal visit.

Conclusions:

• Machine learning models utilizing readily available prenatal data can effectively predict preterm birth.
• Late-pregnancy ultrasound measurements are valuable additions for improving prediction accuracy.
• Integrating ML-based risk assessment and routine late-pregnancy ultrasounds into prenatal care can optimize outcomes for high-risk pregnancies.