Why are fetal heart defects so hard to see?
Congenital heart disease, a heart defect present from birth, is the most common birth defect. The screening ultrasound, done between 18 and 24 weeks of pregnancy, provides five views of the heart that together can in theory detect 90% of complex forms.
In practice, sensitivity (the share of cases actually caught) can fall as low as 30%. Spotting these defects takes real expertise, which is not available everywhere.
How was the AI trained?
The researchers used 107,823 images from 1,326 scans, taken both for screening and for more detailed heart examinations, in fetuses of 18 to 24 weeks. They combined several neural networks, which are programs that learn to recognise patterns in images.
Some networks first identify the recommended views of the heart, others separate normal hearts from hearts with a complex defect. Other models also measure standard dimensions of the fetal heart and chest.
What performance did the study measure?
The AI was tested on 4,108 screening scans, of which 0.9% showed a defect. It reached an area under the curve of 0.99, a performance score where 1 would be perfect. Its sensitivity was 95%, its specificity (the ability to recognise a normal heart) 96%, and when it called a heart normal it was right 100% of the time in this test.
The authors report that the model's sensitivity was comparable to clinicians and stayed robust on images from outside hospitals and of lower quality. Its decisions relied on clinically meaningful features, and its heart measurements matched those reported by doctors.
The authors stress that it matters whether results hold up outside the hospital that built the tool. Here the sensitivity stayed robust on outside-hospital images and on lower-quality images, which matters because image quality varies a lot between machines and operators.
Key figures
- 95%Sensitivity of the AI in telling an abnormal heart from a normal one (confidence interval: 84 to 99%).
- 96%Specificity: share of normal hearts correctly recognised.
- 4,108Screening scans in the internal test set (0.9% with a heart defect).
- 30%Sensitivity of screening ultrasound in practice, at its lowest, according to the authors.
What could this mean for families?
Better detection before birth makes it possible to prepare the delivery and the baby's care in a suitable centre. The authors suggest this kind of tool could significantly improve detection of these defects, a diagnostic challenge worldwide.
The AI would be an aid for the professionals who do the scan, not a substitute. Follow-up and decisions stay with the medical team.
What are the limits of this study?
The study is retrospective: it was run on scans already done, not live during pregnancies. The share of defects in the test set is small (0.9%), and the confidence interval around sensitivity is wide (84% to 99%).
Prospective studies, in several countries and with different machines, will be needed to know whether the AI really improves screening and the care of babies.
The model will also need to be tested on scans from many more machines and countries before anyone can say how it behaves everywhere.
A pregnancy scan: asking your questions with Mima
This research AI could one day help spot heart defects in the foetus. It is not yet in everyday use. It is a delicate subject: the scan team explains what they see, with care.
Mima keeps the dates of your scans, gathers your questions for each appointment and explains medical terms in plain words. You can ask: "Does my scan include a detailed look at my baby's heart, and when is it done?" You are not alone.
Frequently asked questions
Can AI spot a heart defect in a fetus?
According to this 2021 study, yes: on 4,108 screening scans, the AI flagged 95% of abnormal hearts. It was validated on images already recorded, not yet on pregnancies followed live.
Why does current screening miss cases?
Images of the fetal heart are hard to obtain and to read. According to the authors, screening sensitivity can fall to 30% in practice.
Does the AI replace the person doing the ultrasound?
No. It is presented as an aid to improve detection. Diagnosis and follow-up stay with professionals.
Can maternity units already use it?
This study is retrospective. Prospective trials will be needed to know whether the tool improves care in practice.
The original study
Arnaout R, Curran L, Zhao Y, et al. “An ensemble of neural networks provides expert-level prenatal detection of complex congenital heart disease.” Nature Medicine, 14 May 2021. DOI : 10.1038/s41591-021-01342-5
Read the full study in Nature MedicineThis article summarises a scientific study and is not medical advice.