Why is measuring heart function so tricky?
An echocardiogram is a painless ultrasound video of the heart. It is often used to calculate the ejection fraction: the share of blood that the left ventricle, the heart's main pumping chamber, pushes out with each beat. This number drives big decisions, such as whether to continue a chemotherapy that can strain the heart, or whether to place an implantable defibrillator, a small device that corrects dangerous heart rhythms.
The catch is that this measurement depends on a human hand. A trained professional traces the outline of the heart chamber on screen, and no two people trace it exactly the same way. AI tools that measure ejection fraction already existed, but none had been tested in a blinded randomized trial.
How did the researchers set up the trial?
The team at Cedars-Sinai Medical Center in Los Angeles took 3,769 echocardiograms performed in 2019. They excluded 274 because the images were too poor to measure. The remaining 3,495 exams, from 3,035 patients, were randomized, just like in a drug trial: half got a first reading by the AI, and half got a first reading by one of 25 sonographers, the specialists who perform and measure heart ultrasounds, with 14 years of experience on average.
Next, 10 cardiologists reviewed every exam in their usual software and adjusted the outline if needed, without knowing who had made the first reading. That is what blinded means. The main question was simple: in how many exams did the final number differ from the first one by more than 5 points? The AI had been trained on videos from another hospital, Stanford, and had never seen images from Cedars-Sinai.
Did cardiologists correct the AI less often?
Yes. Cardiologists substantially changed the first reading in 16.8% of exams prepared by the AI, compared with 27.2% of exams prepared by sonographers. So the AI was not just as good: it did better on this measure. And the blinding held: cardiologists could not reliably tell who had made the first reading.
The AI also saved time. Sonographers no longer had to trace the outline, which took them about 2 minutes per exam. Cardiologists needed a median of 54 seconds to review an AI-prepared exam, versus 64 seconds for a sonographer-prepared one. Another useful finding: the reading crossed the 35% threshold, which doctors use when discussing a defibrillator, between the first and final review in 1.3% of AI exams, versus 3.1% in the other group.
Key figures
- 3,495echocardiograms randomized between the AI and sonographers
- 16.8%of AI-prepared exams substantially changed by cardiologists, versus 27.2% with a sonographer
- 54 smedian cardiologist review time for an AI-prepared exam, versus 64 s with a sonographer
Keeping track of your ejection fraction with Mima
This trial does not change your exam yet: your echocardiogram is still performed and approved by your care team. But it shows that ejection fraction can shift a little from one reading to the next, which matters if you get a treatment that can strain the heart.
Mima helps you keep track of your echocardiograms and their results over time, and prepare your questions for appointments, such as: "Has my ejection fraction changed since my last exam, and is that a real change?"
What could this mean for patients?
A more consistent ejection fraction matters for people who are monitored closely: during some cancer treatments that can strain the heart, or with heart failure, when the heart does not pump well. Less variation from one reading to the next means less risk of making a decision based on a slightly wrong number.
The authors point out that measurement tools already cleared by the FDA had not been tested this way. In the trial, a cardiologist always signed off on the final result.
What are the limits of this trial?
The trial took place at a single hospital, with its own patients and habits. It used exams that had already been recorded, so the sonographers who traced the outlines were not the ones who had scanned the patients. The 274 poor-quality exams were left out, even though they are often the hardest to measure.
Most importantly, the study was not designed to measure long-term effects on patients' health. It shows a steadier measurement, not yet better care. The authors' next step is to test the AI across several hospitals.
Frequently asked questions
What is ejection fraction on an echocardiogram?
It is the share of blood the left ventricle pushes out with each heartbeat. Doctors use it to judge how well the heart pumps and to guide several treatments.
Can AI read an echocardiogram instead of a cardiologist?
Not in this trial. The AI only made the first measurement, and a cardiologist always reviewed and approved the final result.
What is a blinded randomized trial?
Exams or patients are split into groups by chance, and the people judging the results do not know which group each one belongs to. This reduces bias.
Is AI already used for echocardiograms in the US?
The authors note that some FDA-cleared tools for measuring ejection fraction exist. You can ask your cardiologist whether software helps measure your echocardiogram.
The original study
He B, Kwan AC, Cho JH, et al. “Blinded, randomized trial of sonographer versus AI cardiac function assessment” Nature, April 5, 2023. DOI : 10.1038/s41586-023-05947-3
Read the full study in NatureThis article summarizes a scientific study and is not medical advice.