(890) Development of a Novel Machine-Learning-Based Prediction Model to Stratify Risk for Adverse Perinatal Outcomes

Friday, January 31, 2025

10:30 AM – 12:30 PM

Submitting Author and Presenting Author(s)

Rajet Vatsa, PhD (he/him/his)

MD-PhD Student
Harvard University
Boston, MA, United States

Coauthor(s)

SL

Siguo Li, MS

Massachusetts General Hospital
Boston, Massachusetts, United States
JC

Jessica L. Cohen, PhD

Bruce A. Beal, Robert L. Beal and Alexander S. Beal Associate Professor of Global Health
Harvard T.H. Chan School of Public Health
Boston, Massachusetts, United States
Mark A. Clapp, MD, MPH (he/him/his)

Physician Investigator
Massachusetts General Hospital
Boston, MA, United States

Objective: An effective risk assessment tool for perinatal deaths is critically needed, given that over one-third of stillbirths after 28 weeks of gestation are potentially preventable. While current guidelines emphasize the role of clinical risk factors (e.g., medical comorbidities), the value of geographic, sociodemographic, and utilization-related factors in risk stratification is less known.

Study Design: We utilized electronic health records data from a large integrated healthcare system between 4/2017 and 6/2023. The analysis was restricted to births ≥ 28 weeks. We developed and internally validated two classes of machine learning models to predict a composite adverse outcome, including stillbirth, early newborn death, and 5-minute Apgar < 7, using features known before 28 weeks. The first class (“clinical”) only leveraged clinical risk factors (e.g., chronic hypertension) known to be associated with perinatal death, while the second class (“clinical+”) added additional geographic (e.g., distance to hospital), sociodemographic (e.g., household income), and utilization-related (e.g., prenatal care use) features. We tested and compared the top-performing models in each class with respect to their ability to prospectively stratify the risk of adverse perinatal outcomes.

Results: The clinical+ model (AUC: 0.63 [95%-CI: 0.59-0.69]) had modest discrimination overall, similar to the clinical model (AUC: 0.58 [95%-CI: 0.53-0.64]). However, the clinical+ model stratified pregnancies into quartiles of predicted risk with significantly different rates of the composite adverse outcome compared to the clinical model, which did not. For the clinical+ model, the adverse outcome rate was 0.6% (95%-CI: 0.3-0.9%) in the lowest and 4.6% (95%-CI: 3.9-5.3%) in the highest risk quartile.

Conclusion: A machine-learning-based model developed using clinical, geographic, sociodemographic, and utilization-related factors known at the start of the third trimester may more accurately stratify risk than using clinical factors alone. Prospective and external validation studies are needed to demonstrate its utility in practice.