Detection of findings suspicious for severe congenital heart defects on fetal echocardiograms by an artificial intelligence-based software

Abstract

Introduction

Prenatal detection of severe congenital heart defects (CHDs) is crucial for improving neonatal outcomes. However, identifying cardiac anomalies on prenatal ultrasounds remains challenging, contributing to consistently low detection rates globally.

We aim to evaluate the efficacy of an artificial intelligence (AI) software to detect examinations suspicious for severe CHD based on eight ultrasound findings during fetal echocardiography.

Method

We retrospectively included 193 fetal echocardiography examinations from singleton pregnancies, performed between 18 and 24 weeks of gestation at one pediatric cardiology center in the United States. The dataset included 84 exams with various forms of severe CHD confirmed postnatally and 108 exams without any cardiac abnormalities.

The AI-based software analyzes 4-chamber, LVOT and RVOT standard views on all grayscale 2D ultrasound clips to determine the presence of eight findings that may be associated with severe CHD: overriding artery, septal defect at the cardiac crux, abnormal relationship of the outflow tracts, enlarged cardiothoracic ratio, right-to-left ventricular size discrepancy, tricuspid-to-mitral valve annular size discrepancy, pulmonary-to-aortic valve annular size discrepancy and cardiac axis deviation. The presence of any of these findings would warrant increased attention by a fetal heart specialist.

Results

The AI detected on average 3.9 (95% CI, 3.5; 4.2) suspicious findings per case with severe CHD, significantly more than for cases with no cardiac abnormalities (0.02, 95% CI: 0; 0.04, p < 0.001). This difference was consistent across all major types of severe CHD (Figure 1). The software detected at least one suspicious finding in 82 of 84 cases with severe CHD. Among the 108 exams with no cardiac abnormalities, the AI identified 106 as having no findings present, while 2 cases were flagged with abnormal cardiac axis deviation. This resulted in a sensitivity for detecting severe CHD of 0.98 (95% CI, 0.92; 0.99) and a specificity of 0.98 (95% CI, 0.93; 0.99).

Conclusion

The AI software shows strong potential in distinguishing between severe CHD cases and cases with no cardiac abnormalities. Ultimately, AI assistance in detecting fetal ultrasounds at risk of severe CHD could enhance prenatal detection rates.