Why is measuring heart function tricky?
A heart ultrasound (an echocardiogram) films the beating heart using sound waves. One number in it matters a great deal: the ejection fraction, the share of blood that the left ventricle, the heart's main pump, pushes out with each beat. It is used to diagnose heart disease, to watch for harmful effects of some treatments and to guide decisions in very sick patients.
But this measure depends on people. Doctors assess it on a small number of the beats that were recorded, and two experts, even after years of training, do not always give the same number.
A more reliable and more consistent measurement would therefore help doctors decide with more confidence, for example when a treatment needs changing or a patient needs closer monitoring.
What does EchoNet-Dynamic do?
The team trained an AI that analyses whole videos, not just a few still images. It performs three tasks: outlining the left ventricle, estimating the ejection fraction, and indicating whether there is cardiomyopathy, a disease of the heart muscle.
Because it watches several beats, it can combine the information from each of them. To help research, the authors also made public a large set of 10,030 annotated ultrasound videos.
In practice, the program looks beat after beat, as a very attentive expert would if they had time to examine every contraction of the heart rather than just a few.
What are the results?
The outline of the left ventricle drawn by the AI overlaps very well with the experts' (Dice similarity coefficient of 0.92). For ejection fraction, the mean error is 4.1 points, and the AI picks out heart failure with reduced ejection fraction with an area under the curve of 0.97.
On a data set from another health system, the results remain good: a mean error of 6.0 points and an area under the curve of 0.96. A prospective evaluation with repeated human measurements confirmed that the AI's variability is comparable to or lower than that of human experts.
The area under the curve shows how well a test separates patients who have a condition from those who do not: 1 is a perfect test. Values of 0.97 and 0.96 are therefore very high.
Key figures
- 4.1 pointsmean error on ejection fraction, with an area under the curve of 0.97 for heart failure with reduced ejection fraction
- 6.0 pointsmean error on an external data set from another health system
- 10,030annotated ultrasound videos made available for research
What could this change for patients?
A tool like this could make the measurement more consistent from one exam to another and from one centre to another. It could also help spot small changes in heart function earlier, for example in patients on a treatment that can strain the heart.
The authors point to real-time diagnosis of heart disease, during the exam itself, as a longer-term goal.
For patients, the practical hope is a measurement that depends less on who happens to perform the scan.
What are the limits of this study?
The AI was trained and tested on videos that had already been recorded. On the external data set the error is larger (6.0 points versus 4.1), which shows that a tool can lose accuracy when the hospital or the machine changes.
The study does not show that using it improves patients' follow-up or survival. Before everyday use, trials in real conditions are needed, and the doctor stays responsible for the interpretation.
Finally, the results would need to be checked on other types of images and other machines, especially when image quality varies more.
A more regular heart check: how to prepare
This AI, still in research, measures the heart's pumping capacity from ultrasound videos. Such a tool could make the measurement more consistent from one scan to the next. Your cardiologist still interprets.
If you are on a treatment that can tire the heart, Mima keeps the dates of your ultrasounds and your symptoms, such as breathlessness. You can ask: "Are my measurements compared from one scan to the next, and by whom?" She prepares these questions with you.
Frequently asked questions
What is ejection fraction?
It is the share of blood that the left ventricle pushes out with each beat. It is used to judge the heart's pumping capacity.
Is the AI more accurate than a cardiologist?
According to the authors, it surpasses human experts on key tasks and its measurements are more reproducible. It does not replace the doctor, who keeps the interpretation.
Why is the error larger on the external data set?
Because the images come from another health system, with different machines and habits. The results are still good but slightly less precise (6.0 points of error versus 4.1).
Is this tool used in hospitals?
The study is a demonstration on recorded videos. It does not yet prove a benefit for patients in routine care.
The original study
Ouyang D, He B, Ghorbani A, et al. “Video-based AI for beat-to-beat assessment of cardiac function” Nature, 25 March 2020. DOI : 10.1038/s41586-020-2145-8
Read the full study in NatureThis article summarises a scientific study and is not medical advice.