Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

Presentation Icons

Additional registration fee required

Faculty have requested this content not be shared on social media

CME Credit Offered

Poster Icons

Foundation Awardee

No Social Media

Back

Q&A

Prematurity

Poster Session 1

(187) Development of Predictive Model for Preterm Birth Using Machine Learning with US CDC Data

Tuesday, February 13, 2024

10:30 AM – 12:00 PM

Location: Prince George's Exhibit Hall CDE, Lower Atrium Level

Presenting Author(s)

Jessica Gulati, N/A

University of Chicago
Chicago, IL, United States

Coauthor(s)

GK

Gloria Khafif, N/A

University of Chicago
Chicago, IL, United States
RD

Rosalie DePaola, MD

Lenox Hill Hospital
New York City, NY, United States
MB

Maria Teresa Benedetto, MD

Lenox Hill Hospital
New York City, NY, United States
TC

Teresa Cheon, MD

Lenox Hill Hospital
New York City, NY, United States
MS

Mio Sawai, MD

Attending Physician
Lenox Hill Hospital
New York City, NY, United States
CY

Christine Yang, MD

Lenox Hill Hospital
New York City, NY, United States

Primary Author(s)

YA

Yuzuru Anzai, MD

Lenox Hill Hospital
New York, NY, United States

Objective: This study aims to create a prognostic model for predicting preterm birth using automated machine learning algorithms.

Study Design: Using US CDC natality data (2016-2019), a multiple regression analysis identified significant risk factors for preterm birth. These factors included gestational diabetes, gestational hypertension, pre- and during pregnancy smoking status, maternal age, multifetal gestation, BMI, prepregnancy hypertension, maternal race, prepregnancy diabetes, previous preterm births, interpregnancy interval, insurance type, and maternal presence of STIs. Using 756,288 randomly selected individuals from 2018-2019, two sets of multiple logistic regression (LR) and decision tree (DT) models were generated to predict preterm birth: early screening (in first trimester) and late screening (in third trimester). The early screening model excluded risk factors that had a later onset in pregnancy. For the LR models, we set the threshold to 10% probability, reflecting the national preterm birth rate. An 80-20 train-test split was performed to validate the model.

Results: The DT and LR models performed similarly with accuracies of > 75%. However, the LR models had greater predictive power with F1 (harmonic mean of positive predictive value and sensitivity) scores of 0.333 for early and 0.304 for late screening compared to 0.229 and 0.237, respectively, for the DTs (Table 1). The early screening LR model (threshold = 0.1) had an accuracy of 83.5% and the late screening LR model (threshold = 0.1) had an accuracy of 75.1%. The LR models can also calculate an individual probability percentage for preterm birth based on the risk factors inputted.

Conclusion: The prognostic models were trained on a large, diverse national dataset, affirming their strength. Their high negative predictive value, high specificity and accuracy, and the LR models’ ability to output a probability for preterm birth support that the models can complement physician expertise in determining risk for preterm birth. Their limitations include low sensitivity and low positive predictive value, necessitating further model development.