(632) Preterm preeclampsia prediction using intelligible machine learning

Thursday, February 9, 2023

4:00 PM – 5:30 PM

Objective: Our goal was to develop a predictive model for preterm preeclampsia (preeclampsia with birth < 37 weeks’ gestation) using an externally validated intelligible machine learning approach.

Study Design: This retrospective study used clinical data from 153,432 births at 20 U.S. hospitals (01/2016 - 12/2021). Explainable Boosting Machine (EBM) and Logistic Regression (LR) models were trained to predict preterm preeclampsia using data from 13 hospitals (110,407 births), and externally validated using data from 7 different hospitals (43,025 births). The 30 features used in the models were all available early in pregnancy and included sociodemographic characteristics, pre-pregnancy health, and previous pregnancy complications.

Results: The incidence of preterm preeclampsia was 1.9% (n=2919) (Table 1). The strongest predictors of preterm preeclampsia in the EBM were BMI, nulliparity, age, self-reported history of pre-pregnancy mental health issues, and chronic hypertension. BMI exhibited a near-linear contribution to preterm preeclampsia risk between 22 kg/m2 and 45 kg/m2, with the highest BMI being associated with the highest risk. A significant increase in risk for preterm preeclampsia was observed at around 45 years of age (Figure 1). The EBM yielded an AUC of 0.77 (95% CI 0.75 – 0.78) and outperformed the LR AUC of 0.74 (95% CI 0.73 – 0.75). An AUC of 0.72 (95% CI 0.72 – 0.73) was achieved for EBMs when trained only using the top 5 most important features. The EBM trained on all features identified 65% for preterm preeclampsia cases for a screen positive rate (FPR) of 25%.

Conclusion: EBMs can be used to predict preterm preeclampsia using routinely collected data available early in pregnancy. EBMs also revealed the factors that contributed most to risk in our study population, namely maternal BMI, nulliparity, and age (with a sharp increase in risk at 45 years). Importantly, EBMs highlighted a near-linear relationship between increasing BMI (one of the most modifiable risk factors) and increasing risk for preterm preeclampsia.

Primary Author(s)

Tomas M. Bosschieter, BSc (he/him/his)
PhD Student
Stanford University
Stanford, California, United States

Coauthor(s)

Zifei Xu, BEng (she/her/hers)
Student
Stanford University
Stanford, California, United States

Hui Lan, BS
Stanford University
Stanford, California, United States

BL
Benjamin Lengerich, PhD
Massachusetts Institute of Technology
Cambridge, Massachusetts, United States

HN
Harsha Nori, BS
Microsoft Research
Redmond, Washington, United States

KS
Kristin Sitcov, BS
Executive Director
Foundation for Health Care Quality
Seattle, Washington, United States

IP
Ian Painter, PhD
Foundation for Health Care Quality
Seattle, Washington, United States

RC
Rich Caruana, PhD
Microsoft Research
Redmond, Washington, United States

Presenting Author(s)

Vivienne Souter, MD (she/her/hers)
Medical Director, Womens Heath
Natera, Inc.
Seattle, Washington, United States