Validation of a machine-learning-based algorithm to predict preeclampsia-related adverse outcomes on a real-world dataset.

PURPOSE: Preeclampsia is a major obstetric disorder. Machine learning (ML) models incorporating angiogenic biomarkers show promise in predicting related adverse outcomes, but refinement is needed for clinical use. This study aimed to reduce features to a clinically meaningful set and to develop and validate predictive endpoints for preeclampsia-associated outcomes.

METHODS

A model with a reduced feature set was derived from a training cohort of 1,634 patients (2, 412 visits) and then tested on a validation cohort of 402 patients (540 visits). Three machine learning models were developed to predict (1) adverse outcomes overall, (2) delivery within 14 days before 34 weeks of gestation, and (3) delivery within 7 days after 34 weeks, using 13 features versus 114 originally.

RESULTS

Reduced-feature models demonstrated comparable accuracy to original models across all endpoints. Model 1 (any adverse outcome) achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.92 (95% CI: 0.88-0.96) in training and 0.89 (95% CI: 0.84-0.93, p = 0.31) in the validation cohort, respectively. For delivery within 14 days, the AUROC was 0.92 (95% CI: 0.87-0.96) in training and 0.85 (95% CI: 0.78-0.92) in validation (p = 0.13). Delivery within 7 days showed AUROCs of 0.79 (95% CI: 0.70-0.87) and 0.80 (95% CI: 0.75-0.85) (p = 0.78).

CONCLUSION

A machine learning model with a significantly reduced number of features can accurately predict clinically relevant preeclampsia outcomes. The identified endpoints (timing of delivery and adverse events) could support clinical decision-making and help reduce maternal and neonatal morbidity and mortality.