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

Operative Obstetrics

Poster Session 2

(444) National and International Validation of a Prenatal Machine Learning Algorithm to Forecast Cesarean Section Risk

Tuesday, February 13, 2024

3:30 PM – 5:00 PM

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

Primary & Presenting Author(s)

Reetam Ganguli, BS (he/him/his)

Elythea
San Jose, CA, United States

Coauthor(s)

SW

Stephen M. Wagner, MD

Assistant Professor
Beth Israel Deaconess Medical Center
Brookline, MA, United States

Objective: Clinical risk assessment for cesarean delivery (CD) remains subjective with poor inter-provider reliability. Our goal was to develop an objective, individualized risk prediction model for CD using machine learning (ML) models trained on routine prenatal data. Accurately identifying risk could allow more nuanced counseling on delivery mode, inform conversations around scheduling repeat cesarean, and optimize resource planning for higher risk patients.

Study Design:
We conducted a retrospective cohort study including all patients from 2018-2020 in the Center for Disease Control’s Vital Statistics System utilizing sociodemographic and clinical data prior to delivery. Deliveries with missing data for route of delivery were excluded. The primary outcome was route of delivery through cesarean section.

An extreme gradient boosting model was trained on included patients from 2018-2020, as well as CD patients from years 2014, 2016, and 2017. Weight scaling algorithms were utilized to improve model performance. The model was retrospectively tested on patients from the year 2021 of the Vital Statistics System and prospectively tested on a cohort of women enrolled across 8 hospitals in Cameroon and Nigeria.

Results:
The model was trained on 10,541,899 patients, of which 470,2731 (44.6%) had a CD.

When retrospectively tested on 2,870,804 patients, where 1,022,481 (35.6%) had a CD, the model achieved 0.85 AUC, 78.1% accuracy, 0.72 sensitivity, and 0.78 F1 score.

When prospectively tested on 240 African patients, where 84 (35%) had a CD, the model achieved 0.79 AUC, 80.8% accuracy. 0.71 sensitivity, and 0.80 F1 score when tuned to a 0.77 threshold.

The highest weighted factors for the model were maternal body mass index/pregestational weight and maternal age.

Conclusion:
We demonstrate that prenatal ML models can predict CD likelihood with strong discrimination ability, which can aid patient education, delivery planning, and resource optimization through data-driven clinical decision support. Our findings have the potential to reduce unplanned CD and lower correlated costs for providers and health systems.