Predicting spontaneous preterm birth using machine learning algorithms.

STUDY

DESIGN

This prospective cohort study included pregnant women from a tertiary hospital in Kunming between 2022 and 2023. Demographic characteristics, clinical history, and cervical elastography parameters were analyzed. Four machine learning (ML) algorithms-Boruta, Lasso regression, Sliding Window Sequential Forward Selection (SWSFS), and XGBoost-were applied to screen predictors and construct logistic regression models.

RESULTS

The incidence of sPTB was 9.06%. Cervical length (AUC = 0.784, cutoff ≤3.015 cm), anterior cervical angle (AUC = 0.731, cutoff >101.85°), posterior cervical angle (AUC = 0.623, cutoff >123.91°), and combined anterior + posterior cervical angles (AUC = 0.674, cutoff >204.49°) were significantly associated with sPTB. Consistently identified predictors across all algorithms included cervical length, strain rate, combined cervical angles, and history of sPTB. Among the models, the Boruta-based logistic regression achieved the best performance (AUC = 0.953).

CONCLUSION

Cervical length ≤3.015 cm, combined cervical angles >204.49°, prior sPTB, and abnormal strain rate were critical predictors of sPTB. ML-based models demonstrated high predictive accuracy and hold promise for early clinical risk stratification in this ethnic minority population.