TOPLINE:

Artificial Intelligence (AI)-aided readers achieved higher detection rates for congenital heart defects (CHDs) with an area under the receiver operating characteristic curve of 0.974 vs 0.825 for unaided readers. AI assistance reduced interpretation time by 48 seconds per examination while increasing clinician confidence scores from 3.90 to 4.63.

METHODOLOGY:

• Researchers evaluated 200 fetal ultrasound examinations from 11 centers, including 100 cases with suspicious findings for CHDs and 100 without, in singleton pregnancies between 18-24 weeks gestation.
• Analysis included seven obstetrician-gynecologists and seven maternal-fetal medicine specialists who reviewed examinations both with and without AI assistance, assessing eight specific findings suspicious for CHDs.
• Participants rated presence or absence of findings with confidence scores, with readings performed remotely after a 30-day washout period between aided and unaided conditions.
• Primary outcome measures focused on area under the receiver operating characteristic curve for detecting any suspicious finding, while secondary outcomes included sensitivity, specificity, and reading time.

TAKEAWAY:

• AI-aided readers demonstrated significantly improved detection (area under the receiver operating characteristic curve, 0.974; 95% CI, 0.957-0.990) compared to unaided readers (0.825; 95% CI, 0.741-0.908; P = .002).
• Detection sensitivity increased with AI assistance (sensitivity, 0.935; 95% CI, 0.892-0.978) from unaided reading (sensitivity, 0.782; 95% CI, 0.686-0.878), and specificity improved (specificity, 0.970; 95% CI, 0.949-0.991) from unaided reading (specificity, 0.759; 95% CI, 0.630-0.887).
• Average reading time decreased with AI assistance (mean difference, 226 seconds; 95% CI, 218-234) compared to without (mean difference, 274 seconds; 95% CI, 265-283; P < .001).
• Reader confidence scores improved significantly with AI assistance (mean difference, 4.63; 95% CI, 4.60-4.66) from unaided reading (mean difference, 3.90; 95% CI, 3.85-3.95; P < .001).

IN PRACTICE:

“AI assistance for ultrasound interpretation was associated with an improved detection rate of prenatal ultrasonograms suspicious for CHDs across differing patient BMIs, image quality, and ultrasound manufacturers. In the subgroup analyses related to reader characteristics, there were no significant improvements in AUROCs [area under the receiver operating characteristic curves] by years of reader experience,” the authors of the study wrote.

SOURCE:

The study was led by Jennifer Lam-Rachlin, MD, Department of Maternal Fetal Medicine, Icahn School of Medicine at Mount Sinai in New York City. It was published online in Obstetrics & Gynecology.

LIMITATIONS:

According to the authors, the study’s limitations include a higher prevalence of abnormal cases (50%) than that in the general population’s CHD incidence of about 1%. Additionally, the dataset included fetal echocardiography examinations (35.5%) which may not represent typical images interpreted by clinicians in practice. The high proportion (84.0%) of examinations with superior image quality and missing BMI data for many cases may limit generalizability.

DISCLOSURES:

The study was funded by BrightHeart. Lam-Rachlin and Arunamata served as consultants for BrightHeart. Several authors were compensated for ultrasound case reviews by BrightHeart. Malo de Boisredon, Eric Askinazi, Valentin Thorey, and Christophe Gardella declared being employees of BrightHeart; Bertrand Stos and Marilyne Levy were cofounders and shareholders in BrightHeart.

This article was created using several editorial tools, including AI, as part of the process. Human editors reviewed this content before publication.