Machine-Learning-Based Prediction of Preterm Birth in Women with Huge Uterine Fibroids: A Stratified Cohort Analysis
Simon Shenhav1,2, Eyal Sadeh3, Yaniv S Ovadia1,4
1Obstetrics and Gynecology Division, Barzilai University Medical Center, Ashkelon 7830604, Israel.
Background/Objectives: Preterm birth remains a major cause of neonatal morbidity and mortality, and risk stratification in pregnancies with uterine fibroids is limited. This study evaluated whether detailed phenotyping of huge uterine fibroids provides predictive information. Methods: This retrospective single-center study included 192 singleton pregnancies: 64 with a huge uterine fibroid (maximum diameter ≥ 10 cm) and 128 fibroid-free controls. We analyzed the full and fibroid-only cohorts. Machine learning (ML) models were compared across full fibroid-feature, alternative fibroid-feature, non-fibroid, and clinical benchmark configurations. Results: In validation analyses, the alternative fibroid-feature configuration was selected as the best-performing configuration in both cohorts. The best validation models were Random Forest for the full cohort and Logistic Regression for the fibroid-only cohort, achieving F1-scores of 0.67 and 0.80 and areas under the receiver operating characteristic curve (AUCs) of 0.94 and 0.90, respectively. On the held-out test set, models achieved F1-scores of 0.50 and 0.57, with AUCs of 0.82 and 1.00, respectively; uncertainty and calibration remained limited by the very small number of positive events. SHapley Additive exPlanations analysis showed that fibroid-related variables contributed to model output. Conclusions: Detailed fibroid phenotyping may add predictive information beyond clinical variables, but these exploratory findings require further validation in larger cohorts.
Background/Objectives: Preterm birth remains a major cause of neonatal morbidity and mortality, and risk stratification in pregnancies with uterine fibroids is limited. This study evaluated whether detailed phenotyping of huge uterine fibroids provides predictive information. Methods: This retrospective single-center study included 192 singleton pregnancies: 64 with a huge uterine fibroid (maximum diameter ≥ 10 cm) and 128 fibroid-free controls. We analyzed the full and fibroid-only cohorts. Machine learning (ML) models were compared across full fibroid-feature, alternative fibroid-feature, non-fibroid, and clinical benchmark configurations. Results: In validation analyses, the alternative fibroid-feature configuration was selected as the best-performing configuration in both cohorts. The best validation models were Random Forest for the full cohort and Logistic Regression for the fibroid-only cohort, achieving F1-scores of 0.67 and 0.80 and areas under the receiver operating characteristic curve (AUCs) of 0.94 and 0.90, respectively. On the held-out test set, models achieved F1-scores of 0.50 and 0.57, with AUCs of 0.82 and 1.00, respectively; uncertainty and calibration remained limited by the very small number of positive events. SHapley Additive exPlanations analysis showed that fibroid-related variables contributed to model output. Conclusions: Detailed fibroid phenotyping may add predictive information beyond clinical variables, but these exploratory findings require further validation in larger cohorts.
