Abstract Preterm birth (PTB) affects an estimated 15 million infants worldwide each year, raising an enormous challenge in the field of maternal-fetal medicine. This work explores machine learning approaches for distinguishing between term and preterm birth from three-channel electrohysterogram (EHG) measurements of 300 pregnancies, specifically featuring a novel “model error” feature based upon autoregressive modeling of uterine electrical activity. We compared five different machine learning classifiers—K-Nearest Neighbors (K-NN), Support Vector Machine (SVM), Logistic Regression, Naive Bayes, and Gradient Boosting—using a variety of performance indicators such as accuracy, precision, recall, F1-score, and AUC-ROC. Addition of the model error feature, calculating the departures from autoregressive predictions as a measure of unpredictability in uterine contractions, substantially enhanced performance in every case considered. The greatest accuracy (92.50%) and highest AUC-ROC (94.69%) were achieved by Gradient Boosting, in close succession by SVM (92.36% accuracy, 95.85% AUC-ROC). Feature importance analysis verified model error as the strongest predictor. These results, though promising, require external validation in populations with different distributions to establish their potential for clinical use. This work demonstrates the potential for physiologically-informed feature engineering to combine effectively with traditional machine learning methods in PTB prediction, though considerations toward computational efficiency and practical in-patient implementation are worthy of greater scrutiny.
Powar et al. (2026) studied this question.