Researchers at the Swiss Federal Institute of Technology have trained an artificial intelligence system to spot congenital heart defects in newborns with an accuracy that falls in the 80 to 90 percent range. The findings were shared in a study published in the International Journal of Computer Vision, reflecting a significant step toward machine-assisted newborn heart screening.
The development team built the AI by curating a large training dataset drawn from ultrasound videos of 192 newborn hearts. The dataset included sequences showing the beating heart from multiple angles, capturing natural motion and anatomical variations. Alongside the imaging data, each frame carried diagnostic labels assigned by seasoned pediatric cardiologists, providing the ground truth that the AI used to learn patterns associated with different heart conditions. This combination of dynamic imaging and expert interpretation aimed to teach the model the subtle cues that indicate typical or atypical cardiac structure and function in newborns.
To evaluate the model’s diagnostic capability, researchers tested it on a separate collection of ultrasound sequences that the AI had not previously examined. The results indicated that the algorithm could correctly identify the presence or absence of a defect in a substantial majority of cases, yielding a diagnostic accuracy between 80 and 90 percent. Moreover, the model demonstrated a notable ability to gauge the severity of detected defects, achieving an accuracy range of roughly 65 to 85 percent in classifying how serious the conditions were. This indicates the potential for AI to assist not only in initial detection but also in triaging cases by severity, which can help clinicians prioritize urgent evaluations and treatment planning.
The researchers emphasize that health care facilities equipped with well-trained specialists may eventually integrate this AI system as a support tool. In real-world settings, the technology could help reduce the workload on clinicians by handling routine image interpretation tasks and flagging cases that warrant human review. When paired with expert oversight, such AI tools have the potential to enhance diagnostic confidence, shorten turnaround times for critical cases, and support more consistent assessments across different operators and imaging centers. While the study shows promise, the authors also note that continued validation across diverse populations and imaging setups is essential to ensure robustness and generalizability before widespread clinical deployment.
As this field advances, ongoing collaboration between engineers, pediatric cardiologists, and health care administrators will be crucial. The goal is to translate high-precision imaging analysis into practical workflows that improve outcomes for newborns with heart conditions while maintaining safety, privacy, and patient trust. The intersection of medical imaging, machine learning, and clinical expertise is moving toward tools that empower clinicians with faster, more reliable insights, ultimately supporting better early intervention for infants who need it most.