Document Type

Conference Proceeding

Publication Date

1-2025

Publication Title

Pregnancy

Abstract

Objective: Prenatal detection of congenital heart defects (CHD) improves patient outcomes. Despite advances in imaging techniques and equipment, prenatal diagnosis remains low. We evaluated the accuracy of an artificial intelligence (AI) system to detect 2nd trimester (2T) ultrasound studies suspicious for CHD. Study Design: The AI system analyzes all grayscale 2D ultrasound cines of a study and detects 8 morphological findings suspicious for CHD. The findings were selected such that most major forms of CHD would feature at least one such finding, and the presence of any of these would flag the study for referral for further examination. We retrospectively included 877 exams (obstetric or detailed anatomic ultrasounds, or fetal echocardiograms), in singleton pregnancies, between 18-24 weeks gestation from 11 centers, including 311 exams with a major CHD, cardiomegaly or levo/dextrocardia. Each exam included 4-chamber, LVOT and RVOT views documented in cines, and was not used for training of the AI system. For each exam, 3 expert fetal cardiologists annotated the presence or absence of each finding and majority voting was used as the ground truth. The AI system predicted the presence, absence or inconclusiveness (in case of low confidence) of each finding. An exam was predicted positive if at least one finding was present, negative if all findings were absent and inconclusive otherwise. Results: Regarding the detection of any finding, the AI system had a conclusive output for 98.8% (95%CI 97.8-99.3) of exams. On exams with a conclusive output, sensitivity was 98.7% (96.7-99.5), and specificity was 97.7% (96.1-98.6). Performance per individual finding is presented in the Figure. To better represent the general low risk population despite the artificially high prevalence of exams with at least one suspicious finding (311/877), per-finding specificity was computed on exams negative to all findings. Conclusion: The AI system accurately detected 2T ultrasound exams suspicious for CHD consistent with experts in most cases. The AI system may improve the prenatal ultrasound detection of CHD and thereby improve outcomes.

Volume

1

Issue

S1

Comments

Society for Maternal-Fetal Medicine (SMFM) 2025 Pregnancy Meeting, January 27 - February 1, 2025, Aurora, CO

DOI

10.1002/pmf2.12008

Download

Share

COinS