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

Area of Science:

• Obstetrics and Gynecology
• Medical Informatics
• Biostatistics

Background:

• First-trimester preeclampsia (PE) screening is crucial for early intervention.
• The Fetal Medicine Foundation (FMF) algorithm is a widely adopted standard for PE risk assessment.
• Machine learning (ML) models offer potential alternatives for improving screening accuracy.

Purpose of the Study:

• To compare the performance of commonly used ML classifiers against the FMF algorithm for first-trimester PE screening.
• To evaluate ML models using both maternal factors only (a priori) and augmented with biophysical/biochemical markers (a posteriori).
• To assess screening performance in a Romanian cohort.

Main Methods:

• Analysis of 1583 singleton pregnancies screened between 11-14 weeks' gestation.
• Comparison of logistic regression, random forest, and XGBoost ML models against FMF risk estimates.
• Evaluation using AUC-ROC, DeLong tests, sensitivity at 10% FPR, and decision-curve analysis (DCA).

Main Results:

• The FMF algorithm demonstrated superior performance compared to ML models in both a priori and a posteriori screening scenarios.
• FMF showed higher sensitivity at a 10% false-positive rate (50.0% vs. 33.3% a priori; 71.4% vs. 57.1% a posteriori).
• Decision-curve analysis favored the FMF algorithm, especially for the a posteriori model.

Conclusions:

• The FMF algorithm's competing-risks framework outperformed or matched ML classifiers in this cohort.
• Current ML approaches may not surpass established algorithms like FMF for PE screening.
• Further research is needed to optimize ML applications in obstetric screening.