Internal and external validation of a machine learning algorithm to detect preeclampsia-related adverse outcomes in high-risk pregnancies.

OBJECTIVES: This study aimed to refine an existing machine learning (ML) algorithm for predicting preeclampsia-related adverse outcomes and to assess its generalizability and predictive performance through internal validation in a German cohort and external validation in a North American cohort.

STUDY DESIGN: A retrospective analysis was conducted using data from two cohorts: a cohort of 1,634 pregnant women in Germany and a prospective study cohort of 946 in North America, all presenting with clinical suspicion of preeclampsia.

MAIN OUTCOME MEASURES: Gradient-boosted trees and logistic regression were used to predict (1) any adverse maternal or fetal outcome, (2) delivery within 14 days before 34 + 0 weeks, and (3) delivery within 7 days after 34 + 0 weeks. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC).

RESULTS: Despite notable differences in baseline characteristics between cohorts, the refined model demonstrated strong and consistent predictive performance. For predicting any adverse outcome, AUROCs were 0.92 (95% CI: 0.87-0.96) in the German cohort and 0.87 (95% CI: 0.82-0.91) in the North American cohort. For delivery within 14 days before 34 + 0 weeks, AUROCs were 0.92 and 0.88, respectively. For delivery within 7 days after 34 + 0 weeks, AUROCs were 0.79 and 0.78.

CONCLUSIONS: The refined ML model maintained high predictive accuracy across two distinct populations, demonstrating its generalizability and potential for integration into clinical decision-making. These findings support the use of machine learning in enhancing the prediction of preeclampsia-related adverse outcomes and improving maternal and neonatal care.