Comparison of Machine Learning Models and the FMF Competing-Risks Algorithm for First-Trimester Preeclampsia...

Background/Objectives: First-trimester preeclampsia (PE) screening is most widely implemented using the Fetal Medicine Foundation (FMF) algorithm, which combines maternal factors with biophysical and biochemical markers via a competing-risks/Bayes framework to produce individualized risks and guide prophylaxis decisions. We aimed to compare commonly used machine-learning (ML) classifiers (logistic regression, random forest, XGBoost) against FMF a priori and a posteriori risk estimates in a Romanian screening cohort. Methods: We analyzed 1583 singleton pregnancies screened at 11-14 weeks' gestation. Primary analyses excluded aspirin-treated women to reduce treatment-induced outcome modification. We evaluated two feature sets mirroring FMF structure: (1) a maternal-factor "a priori" set and (2) a "a posteriori" set additionally incorporating mean arterial pressure (MAP), uterine artery pulsatility index (UtA-PI), and Pregnancy-Associated Plasma Protein A (PAPP-A). Models were trained using stratified repeated cross-validation (5-fold × 10 repeats) and evaluated using AUC-ROC, DeLong tests, and sensitivity at 10% false-positive rate. Calibration of the model, sensitivity analyses and decision-curve analysis (DCA) were also assessed. Results: In the a priori comparison, the best ML model was logistic regression (AUC 0.796) versus FMF prior risk AUC 0.841 (DeLong p = 0.349). The sensitivity at 10% false positive rate (FPR) was 33.3% for the model versus 50.0% for FMF model. In the a posteriori comparison, the best ML model was random forest (AUC 0.844) versus FMF posterior risk AUC 0.929 (DeLong p = 0.087), with sensitivity at 10% FPR of 57.1% for ML and 71.4% for FMF. Random undersampling did not improve ML performance. Including aspirin-treated pregnancies did not significantly change our results. Conclusions: In this study, the FMF competing-risks outputs outperformed or matched ML classifiers in both maternal-only and biomarker-augmented screening, and DCA favored FMF particularly for the a posteriori model.