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: Ultrasound/Imaging

Poster Session 4

(1228) Assessment of Machine Learning Models for First-Trimester Detection of Fetal Congenital Heart Defects

Thursday, February 12, 2026

3:30 PM - 5:00 PM

Submitting Author and Presenting Author(s)

Rebecca Horgan, MD

Assistant Professor
Eastern Virginia Medical School at Old Dominion University
Norfolk, Virginia, United States

Coauthor(s)

JC

Jonathan Colen, PhD

Old Dominion University, Joint Institute on Advanced Computing for Environmental Studies
Norfolk, Virginia, United States
DL

Daniel Lersch, PhD

Thomas Jefferson National Accelerator Facility
Newport News, Virginia, United States
LE

Lana El-Kassis, MD

Eastern Virginia Medical School at Old Dominion University
Norfolk, Virginia, United States
AP

Achyut Paudel, PhD

Old Dominion University, Joint Institute on Advanced Computing for Environmental Studies.
Norfolk, Virginia, United States
AA

Alfred Abuhamad, MD

Eastern Virginia Medial School at Old Dominion University
Norfolk, Virginia, United States
HR

Heather Richter, PhD

Hampon Roads Biomedical Research Consortium
Norfolk, Virginia, United States
MS

Malachi Schram, PhD

Thomas Jefferson National Accelerator Facility
Newport News, Virginia, United States

Objective:
To evaluate machine learning (ML) models for early detection of congenital heart defects (CHD) using first-trimester ultrasound.

Study Design:
We retrospectively analyzed de-identified grayscale and color Doppler ultrasound cine loops obtained between 11 to 13+6 weeks’ gestation from our institution’s database between 2018 and 2024. The dataset included 8,820 clips from 927 patients, 56 of whom had confirmed CHD. To evaluate the potential of ML for early CHD detection, we tested three model types. The first approach used visual pattern analysis, extracting key imaging features from each clip to train a classification model. The second involved deep learning models trained directly on the raw ultrasound images to identify abnormalities. The third applied a general-purpose medical AI model (MedGemma), which interpreted each clip using knowledge derived from a broad range of medical imaging data. Each model produced a final prediction for each patient by summarizing results across all available clips. All models were evaluated on patients not used during training to ensure unbiased performance.

Results:
The best-performing model was a support vector machine (SVM) that analyzed key visual patterns from each ultrasound clip. It achieved an area under the ROC curve (AUC) of 0.84. At 90% specificity, the model detected 62% of CHD cases, with sensitivity increasing to 81% at 70% specificity (Table 1, Figure 1). In comparison, deep learning models trained directly on the ultrasound images underperformed, with AUC values up to 12% lower. The general-purpose medical foundation model (MedGemma) demonstrated poor accuracy, misclassifying 74% of normal cases as CHD, underscoring the limitations of non-specialized AI in early fetal cardiac screening (Table 1, Figure 1).

Conclusion:
Machine learning models demonstrate potential as adjunctive tools for the early detection of CHD. With further optimization and validation in larger, diverse populations, ML-based approaches may enhance prenatal screening protocols.