Eclampsia risk prediction across diverse U.S. populations using CDC data: machine learning versus ACOG checklists

Background Various tools are employed in clinical practice to predict (pre)eclampsia. Artificial intelligence offers a promising approach for identifying people at risk. More data is needed on models trained for diverse populations. Objective This study aimed to develop machine learning prediction models for eclampsia utilizing CDC data from 3.6 million births in 2022, compare them to the American College of Obstetrics and Gynecology (ACOG) checklist, and assess their applicability to various population subgroups. Study design Machine learning models were trained on predictors outlined by the ACOG, and available pregnancy characteristics in vital statistics. The best model was then applied on 6 subsets—American Indian or Alaska, Hawaiian Native or Other Pacific Islander, Asian, Black/African American, mixed heritage, lower income, and being born abroad. Results The ACOG checklist applied on this cohort had a true-positive rate of 81.4% for predicting eclampsia and a false-positive rate of 68.9%. Logistic regression, Random Forest, LightGBM and XGBoost achieved an Area-under-the-receiver-operating-characteristic-curve of 0.64, recall rates around 50% and false-positive rates around 25%. The Area-under-the-receiver-operating-characteristic-curve was lower among individuals identifying as Black/African American, American Indian or Alaska Native and Native Hawaiian or Other Pacific Islander, and those born abroad or with lower income. Conclusion The ACOG checklist had a very high false-positive rate but achieved a high detection rate, while requiring less computational resources than the machine learning models. Demographic biases and variability in machine learning model performance highlight the need for external validation tailored to specific populations.