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 4

Category: Intrapartum Fetal Assessment

Poster Session 4

(1061) Large Language Model Inference versus Clinician Judgment for Predicting Intrapartum Fetal Asphyxia from Cardiotocography

Thursday, February 12, 2026

3:30 PM - 5:00 PM

Submitting Author(s)

Debjyoti Karmakar, MD

Obstetrician Gynaecologist Clinician Researcher
University of Melbourne/Mercy Hospital for Women
Mercy Hospital for Women/Heidelberg, Victoria, Australia

Coauthor(s)

ES

Erica Spessot

Mercy Hospital for Women
Mercy Hospital for Women, Victoria, Australia
EM

Enes Makalic

Monash University
Monash University, Victoria, Australia

Presenting Author(s)

Fiona Brownfoot, MBBS, PhD (she/her/hers)

University of Melbourne/Mercy Hospital for Women
Mercy Hospital for Women/University of Melbourne, Victoria, Australia

Objective:
To assess whether GPT-4V’s image-based classifications can predict fetal asphyxia

Study Design:
We conducted a retrospective diagnostic accuracy study to evaluate GPT-4 with Vision (GPT-4V) for zero-shot classification of intrapartum cardiotocography (CTG) screenshots. Prospective evaluation was not feasible due to regulatory constraints (e.g., FDA, HIPAA) and geosecurity concerns regarding data sovereignty. A proof-of-concept design was therefore essential to demonstrate technical feasibility, secure infrastructure integration, and baseline clinical utility. Using a balanced dataset of 1,000 labor episodes (500 with asphyxia), we compared three approaches: GPT-4V visual ratings alone (based on the Royal Australian and New Zealand College of Obstetricians and Gynaecologists [RANZCOG] and National Institute of Child Health and Human Development [NICHD] classification systems), structured clinical features alone, and a hybrid model combining both.

Results:
The RANZCOG-based hybrid model achieved the highest area under the receiver operating characteristic curve (AUROC 0.71; 95% CI: 0.62–0.78) and 72% sensitivity—representing a 12% absolute gain over clinician judgment at matched specificity (AUROC 0.61). GPT-4V–only models showed limited predictive value (AUROC 0.55–0.58). For predicting clinician intervention, hybrid and clinical-only models performed comparably (AUROC 0.83). All analyses were conducted in a geographically restricted, data-sovereign, encryption-enabled cloud environment compliant with Health Insurance Portability and Accountability Act (HIPAA, USA) standards.

Conclusion:

These findings demonstrate that GPT-4V, when integrated with structured clinical data, can offer scalable, infrastructure-light decision support for fetal surveillance and lay essential groundwork for regulatory engagement and prospective trials of trustworthy AI-based tools in maternal–fetal medicine.