(4) AI Significantly Improves Detection of Prenatal Ultrasounds Suspicious for Major Congenital Heart Defects by OBGYN/MFMs

Thursday, January 30, 2025

8:45 AM – 9:00 AM

Location: Aurora Ballroom, Lobby Level

Presenting Author(s)

Jennifer Lam-Rachlin, MD

Icahn School of Medicine at Mount Sinai Hospital
New York, New York, United States

Coauthor(s)

RP

Rajesh Punn, MD

Pediatrics - Cardiology, Stanford University School of Medicine
Stanford, California, United States
SB

Sarina K. Behera, MD

Palo Alto Medical Foundation, Sutter Health
Palo Alto, California, United States
MG

Miwa Geiger, MD

Icahn School of Medicine at Mount Sinai Hospital
New York, New York, United States
ML

Matthias Lachaud, MD

University of Grenoble Alpes
Grenoble, Rhone-Alpes, France
ND

Nadine David, MD

Medical Training Center
Rouen, Haute-Normandie, France
SG

Sara Garmel, MD

Michigan Perinatal Associates and Corewell Health
Dearborn, Michigan, United States
Matthew K. Janssen, MD (he/him/his)

Clinical Assistant Professor
University of Pennsylvania
Philadelphia, PA, United States
KS

Kendra Sylvester, MD

Perinatal Specialists of the Palm Beaches
West Palm Beach, Florida, United States
JK

John Kennedy, MD

Wayne State University School of Medicine
Detroit, Michigan, United States
JS

Jessica Spiegelman, MD

Icahn School of Medicine at Mount Sinai Hospital
New York, New York, United States
Mia A. Heiligenstein, MD

Fellow
Mount Sinai West
Astoria, New York, United States
Nathan S. Fox, MD

Clinical Professor
Icahn School of Medicine at Mount Sinai Hospital
New York, New York, United States
Andrei Rebarber, MD

Clinical Professor, Ichan School of Medicine at Mount Sinai
Icahn School of Medicine at Mount Sinai Hospital
New York, New York, United States
Greggory R. DeVore, MD

Clinical Professor
The Fetal Diagnostic Center of Pasadena
Pasadena, California, United States
Carolyn M. Zelop, MD

Director of Fetal Echocardiography and Perinatal Research
Valley Health System and NYU School of Medicine
Paramus, New Jersey, United States
RB

Roger Bessis, MD

Centre d’Echographie de l’Odéon
Paris, Ile-de-France, France
ML

Marilyne Levy, MD

UE3C - Unité d’explorations cardiologiques - Cardiopathies Congénitales
Paris, Ile-de-France, France
BS

Bertrand Stos, MD

UE3C - Unité d’explorations cardiologiques - Cardiopathies Congénitales
Paris, Ile-de-France, France
MD

Malo De Boisredon, MSc

BrightHeart
Paris, Ile-de-France, France
EA

Eric Askinazi, MSc

BrightHeart
Paris, Ile-de-France, France
VT

Valentin Thorey, MSc

BrightHeart
Paris, Ile-de-France, France

Submitting Author(s)

Christophe Gardella, PhD

BrightHeart
Paris, Ile-de-France, France

Coauthor(s)

AA

Alisa Arunamata, MD

Pediatrics - Cardiology, Stanford University School of Medicine
Stanford, California, United States

Objective:

Congenital heart defects (CHDs) are a leading cause of infant morbidity and mortality partly due to low prenatal detection rates. We evaluated whether an artificial intelligence (AI) system can improve the detection of CHDs on fetal ultrasound exams among both general OBGYNs and MFM specialists.

Study Design:

The AI system analyzes all grayscale 2D ultrasound cines of an exam and detects 8 morphological findings associated with severe CHDs. The presence of any such finding should justify patient referral for further examination. The AI system identifies each finding as present, absent or inconclusive, and highlights frames where findings can be assessed.

A dataset of 200 ultrasound exams from 11 centers in 2 countries was collected (single pregnancy, obstetric or detailed anatomic ultrasounds, or fetal echocardiograms, at 18-24 weeks gestation), with 100 exams having at least one suspicious finding. The ground truth for presence or absence of each finding was determined by a panel of expert fetal cardiologists. These exams were not used for the training of the AI system.

Fourteen physicians (OBGYNs and MFMs, 1-30+ years’ experience) reviewed each exam both aided and unaided by the AI system, in randomized order, and annotated them for the presence or absence of any such finding and of each individual finding, along with confidence scores. Receiver operator characteristics (ROC) area under the curve (AUC), sensitivity and specificity were computed by comparing reviews to the ground truth.

Results:

ROC AUC for detection of any finding was significantly higher for aided than unaided reviews: 0.97 (95% CI 0.96-0.99) vs 0.83 (0.74-0.91), p=0.002 (DBM-OR method). Similar results held for sensitivity: 0.94 (0.89-0.98) aided vs 0.78 (0.69-0.88) unaided and specificity: 0.97 (0.95-0.99) aided vs 0.76 (0.63-0.89) unaided. Mean reading time was shorter for aided (226 ± 218 s) than unaided (274 ± 241 s) reviews (p < 0.001).

Conclusion:

Assistance by the AI system significantly improved detection of studies suspicious for CHD by OBGYNS and MFMs. AI may play a pivotal role in improving prenatal detection of CHD.