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 outside of the session

CME Credit Offered

Abstract Award

Recording available 2/16-5/2

AIUM Credit

Foundation Awardee

Poster Icons

Abstract Award

Foundation Awardee

Poster Session 2

Category: Intrapartum Fetal Assessment

Poster Session 2

(530) Accounting for Time From Labor Onset Improves the Detection of Hypoxic-Ischemic Encephalopathy in Labor

Wednesday, February 11, 2026

3:30 PM - 5:00 PM

Submitting Author and Presenting Author(s)

Johann Vargas-Calixto, PhD, MEng (he/him/his)

Postdoctoral Fellow
Emory University
Atlanta, Georgia, United States

Coauthor(s)

YW

Yvonne Wu, MD, MPH

UCSF
San Francisco, California, United States
MK

Michael W. Kuzniewicz, MD

Kaiser Permanente
Oakland, California, United States
Marie-Coralie Cornet, MD

Neonatology Fellow
UCSF
San Francisco, California, United States
AL

Aditi Lahiri, MPH

Kaiser Permanente
Oakland, California, United States
LG

Lawrence Gerstley, PhD

Kaiser Permanente
Oakland, California, United States
JP

John Parker, MD

Perigen Inc
Cary, North Carolina, United States
PW

Philip Warrick, PhD

Perigen Inc
Cary, North Carolina, United States
RK

Robert Kearney, PhD

McGill University
Montreal, Quebec, Canada

Objective:
Hypoxic-ischemic encephalopathy (HIE) may result from events during pregnancy and delivery, often resulting in lifelong morbidity or mortality. Early detection of fetuses at risk of intrapartum HIE using cardiotocography (CTG) is still challenging. We aimed to determine if accounting for the time from labor onset (TLO) in a machine learning classifier would improve the early detection of HIE.

Study Design:
This retrospective study used CTG recordings from a cohort of 37,522 deliveries, including 342 HIE cases. We developed and compared two random forest classifiers: a time-independent (TI) model using 86 CTG features, and a time-dependent (TD) model which added the TLO as an additional feature. The CTG features were calculated from 20-minute CTG segments. Also, both models were tested for the entirety of labor, which could last up to 72 hours. The decision threshold was selected such that the false-positive rate (FPR) of both models matched the observed emergency Caesarean delivery (ECD) rate in the healthy population.

Results:
Both models yielded the same FPR equal to the ECD rate in the healthy group (27.3%). The TD model was superior in HIE detection: it detected 3% more HIE cases than the TI model one hour before delivery (55.9% vs 52.9%, p=0.0002); and it also detected more HIE cases than their ECD rate (55.9% vs 44.7%). This advantage was more pronounced for earlier detection; the TD model identified 5.3% more cases than the TI model three hours before delivery (43.5% vs 38.2%, p< 0.0001). Also, the TD model had a median detection time of 6.0 h before delivery compared to 5.3 h in the TI model (p=0.0083).

Conclusion:
Our results demonstrate that the performance of a machine learning classifier for the detection of HIE can be significantly improved by the addition of a single feature: the TLO. This indicates that providing a classifier with temporal context increases the early detection rate of HIE without increasing its FPR. Such early detection could allow timely clinical interventions that have the potential to improve neonatal outcomes.