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

Area of Science:

• Maternal Health
• Medical Informatics
• Public Health

Background:

• Clinical tools for predicting pre-eclampsia are essential.
• Artificial intelligence (AI) presents a promising avenue for identifying at-risk individuals.
• There is a need for more data on AI models trained for diverse demographic groups.

Purpose of the Study:

• To develop machine learning (ML) models for eclampsia prediction using large-scale CDC data.
• To compare ML model performance against the American College of Obstetrics and Gynecology (ACOG) checklist.
• To evaluate the applicability of these models across various population subgroups.

Main Methods:

• ML models were trained using ACOG-recommended predictors and vital statistics.
• The best-performing model was tested on six distinct demographic subsets.
• Data from 3.6 million births in 2022 was utilized for model development and validation.

Main Results:

• The ACOG checklist achieved an 81.4% true-positive rate but a 68.9% false-positive rate.
• ML models (Logistic Regression, Random Forest, LightGBM, XGBoost) showed an Area Under the ROC Curve (AUC) of 0.64, with ~50% recall and ~25% false-positive rates.
• Lower AUC was observed in Black/African American, American Indian/Alaska Native, Native Hawaiian/Other Pacific Islander, foreign-born, and lower-income populations.

Conclusions:

• The ACOG checklist, despite a high false-positive rate, offers high detection with lower computational needs.
• ML models demonstrate demographic biases, with performance disparities across different population subgroups.
• External validation of ML models tailored to specific populations is necessary for equitable eclampsia risk prediction.