Validation of machine-learning model for first-trimester prediction of pre-eclampsia using cohort from PREVAL study

Area of Science:

• Obstetrics and Gynecology
• Artificial Intelligence in Medicine
• Biomarker Analysis

Background:

• First-trimester screening for pre-eclampsia (PE) traditionally uses competing-risks models combining maternal history and biomarker multiples of the median (MoM).
• Artificial intelligence (AI) offers a novel approach using machine-learning (ML) models that may bypass the need for MoM conversion, potentially simplifying screening.

Purpose of the Study:

• To evaluate the cross-population applicability of a previously developed ML-based first-trimester PE screening model.
• To compare the performance of the ML model against a standard competing-risks model in a diverse population.

Main Methods:

• A fully connected neural network ML model, trained on UK data (maternal factors, MAP, UtA-PI, PlGF, PAPP-A), was applied to a Spanish cohort (10,110 pregnancies).
• Screening performance was assessed using Area Under the Curve (AUC) and Detection Rate (DR) at a 10% screen-positive rate (SPR).
• Adjustments were made for differences in PlGF analyzer usage between the UK and Spanish cohorts using linear regression.

Main Results:

• The ML model achieved comparable performance to the Fetal Medicine Foundation (FMF) competing-risks model.
• Detection rates at 10% SPR for early, preterm, and all PE were 84.4%, 77.8%, and 55.7% respectively, with AUCs of 0.920, 0.913, and 0.846.
• The model demonstrated effectiveness without population-specific adaptations, though analyzer adjustments were necessary for PlGF.

Conclusions:

• A neural network-based ML model provides effective first-trimester PE screening applicable across different populations.
• Analyzer adjustments for biochemical testing are essential prior to implementing the ML model in new populations.
• The ML model shows promise for improving PE screening accuracy and accessibility globally.