14 - Utilizing Machine Learning and Electronic Medical Records to Predict Preeclampsia

Thursday, February 9, 2023

2:45 PM – 3:00 PM

Location: Moscone West, Level 3

Objective: To improve preeclampsia risk prediction beyond standard guidelines, we aimed to create machine learning models using large-scale electronic health record (EHR) data to identify high-risk patients. Additionally, we sought to determine which factors were most influential in determining risk.

Study Design: A multi-center analysis of patients who received prenatal care from January 2017 to June 2022 was carried out. The data were sourced from the EHR system and data collected during the first sixteen weeks of pregnancy were used for the prediction models. The algorithms to predict preeclampsia (with and without severe features) are trained using LightGBM, a high-performance implementation of gradient boosted tree models in both weighted and unweighted versions.

Results: A total of 121,662 patients were used to train the algorithm and 6.3% were diagnosed with preeclampsia during their pregnancy. Our models have a mean area under the curve (AUC) of 0.743-0.744 for preeclampsia (95% CI: 0.74-0.77). The post-test likelihood ratio is 10 to 15 using our models. The negative predictive value is 93.9%-94.3% with a positive predictive value of 39.2-62.9% (Table 1). The factors most predictive of preeclampsia, in order of importance include number of prior term deliveries, body mass index (BMI), ZIP code, parity, chronic hypertension, maternal age, pre-pregnancy BMI, race, and alcohol use.

Conclusion: Our model has predictive accuracy above that which has been previously reported and above that which is currently utilized (Figure 1). The 10 to 15-fold increase in likelihood of a patient developing preeclampsia after screening positive on our model highlights the potential diagnostic value of implementing this predictive model. It could be applied to personalize risk assessment following a routine first prenatal office visit, without additional workup. Utilizing this model could allow initiation of aspirin prophylaxis in a more precise patient population and targeting of high-risk patients with appropriate antepartum monitoring.

Primary & Presenting Author(s)

Carly I. Hirschberg, MD (she/her/hers)

MFM Fellow
Zucker School of Medicine at Hofstra/Northwell
New York, New York, United States

Coauthor(s)

MS

Marek Sirendi, PhD

Aegis Ventures
New York, New York, United States
CH

Christina Howe, RN

Aegis Ventures
New York, NY, United States
IC

Ipsita Chatterjee, MS

Northwell Holdings at Northwell Health
New York, New York, United States
DH

Dean Hayden, MD, MPH

Aegis Ventures
New York, New York, United States
TN

Trung Nghia Ton, MD, MPH

Aegis Ventures
New York, New York, United States
ZT

Zachary Tan, MD, MSc

Aegis Ventures
New York, New York, United States
MB

Matthew J. Blitz, MD

Zucker School of Medicine at Hofstra and Northwell
Manhasset, New York, United States
AC

Adriann Combs, MSN, RN

Northwell Health
Manhasset, New York, United States
MN

Michael Nimaroff, MBA, MD

Senior Vice President Executive Director Ob/Gyn - Northwell Health; Chairman Department of Ob/Gyn
Zucker School of Medicine at Hofstra and Northwell
Manhasset, New York, United States
BR

Burton Rochelson, MD

Chief, Maternal-Fetal Medicine
Zucker School of Medicine at Hofstra and Northwell
Manhasset, New York, United States