In the quiet work of reading a beating heart on ultrasound, a question that has long belonged to human hands has now been tested against a machine. In a randomized trial across 21 Northern California medical centers, cardiologists who finalized reports drafted by artificial intelligence proved as discerning as those working from human drafts, and did so in nearly half the time. The result is promising but provisional, a single step toward a future in which the physician's judgment is kept, and perhaps given back some of the hours it has long been spent.
AI-Drafted Echocardiogram Reports Match Human Accuracy and Cut Cardiologist Read Time by Nearly Half
Cardiologists could not tell which drafts the AI had written
Mark
The headline number is 46% less time. Where does that come from?
Mimi
Cardiologists averaged 4.6 minutes per scan finalizing an AI draft, versus 8.5 minutes producing a report without AI help.
Mark
And accuracy held up?
Mimi
By the study's measure, yes. Cardiologists substantially changed 50.6% of AI drafts and 49.7% of human-prepared ones, which met the noninferiority criteria.
Luke
Worth noticing what that measure is, though. Half of all drafts, AI or human, needed substantial changes. It tells us AI is as editable as human drafts, not that it is right.
Mimi
Fair. It's a proxy: how much the cardiologist had to rewrite. The blinding does help, since reviewers couldn't tell who wrote the draft.
Mark
Why does time matter so much here?
Mimi
Volume. Lee said a few minutes per read add up across a day, and there's a projected shortfall of about 7,000 cardiologists by 2036.
Luke
Those faster turnaround and more-patient-time benefits are Lee's expectations, though. The study didn't measure them.
Mimi
Right. It also didn't test long-term patient outcomes, and the cardiologists knew they were being studied, which could have shifted their behavior.
Luke
And it's one health system, 1,200 scans from two weeks, presented as preliminary. Real-world scale is still unknown.
Le Pouls
- A trial of 1,200 heart ultrasounds pits AI-written drafts against those made by sonographers and cardiologists, and the reviewers cannot reliably tell them apart.
- Cardiologists changed 50.6% of AI drafts substantially, nearly identical to the 49.7% rate for human drafts, clearing the noninferiority bar.
- Review time fell from 8.5 minutes to 4.6 minutes per scan, a saving that could add up to real relief across a demanding clinical day.
- With a projected shortfall of about 7,000 cardiologists by 2036 looming, the pressure to find efficiency is building fast.
- The study did not measure long-term patient outcomes, and clinicians knew they were being observed, so larger real-world trials must come before any wide rollout.
In the quiet work of reading a beating heart on ultrasound, a question that has long belonged to human hands has now been tested against a machine. In a randomized trial across 21 Northern California medical centers, cardiologists who finalized reports drafted by artificial intelligence proved as discerning as those working from human drafts, and did so in nearly half the time. The result is promising but provisional, a single step toward a future in which the physician's judgment is kept, and perhaps given back some of the hours it has long been spent.
Twenty cardiologists in Northern California spent part of a study period finalizing echocardiogram reports without knowing who had written the first draft. Some drafts came from sonographers, some from other cardiologists, and some from an artificial intelligence tool. When the results were presented at the American Heart Association's Scientific Sessions 2026 in Chicago, the reviewers could not tell the difference.
An echocardiogram, or heart ultrasound, is the most common way clinicians look at the heart, and the report that follows shapes diagnosis and treatment. Usually a sonographer captures the images and a cardiologist reads them and writes the assessment. This trial asked whether an AI tool could analyze the same images and produce a full preliminary report for the cardiologist to check.
The team drew on 1,200 scans from routine care, performed over two weeks in March 2025. Each exam received three drafts, one from the AI, one from a sonographer and one from a cardiologist, randomly assigned and reviewed blind. Accuracy was measured by how often cardiologists made substantial changes before signing. The AI drafts were changed at nearly the same rate as the human ones, and they agreed more closely with patients' prior records.
The bigger difference was time. Cardiologists needed about 4.6 minutes per scan to finalize an AI draft, compared with 8.5 minutes to write a report unaided. Keane K. Lee, M.D., M.S., one of the study's senior authors, said the team had expected strong accuracy but was surprised by the savings. Across a day of many scans, he noted, those minutes could mean faster results and more time with patients.
Lee was clear that the physician remains responsible for the final report and for the personal care plan that follows from it. That matters against a backdrop of workforce strain, with projections of a national shortfall of roughly 7,000 cardiologists by 2036.
The researchers acknowledge that the trial did not show whether this approach improves long-term health, and that participants knew they were being studied. The findings are preliminary, and wider real-world testing will be needed before anyone can say how it performs at scale.
Twenty cardiologists in Northern California spent part of a study period finalizing echocardiogram reports without knowing who had written the first draft. Some drafts came from sonographers, some from other cardiologists, and some from an artificial intelligence tool. According to results presented Monday at the American Heart Association's Scientific Sessions 2026 in Chicago, the reviewers could not tell the difference.
An echocardiogram, or heart ultrasound, is the most widely used form of heart imaging, and it helps clinicians diagnose conditions and choose treatments. In the usual workflow, a sonographer acquires the images and a cardiologist reviews them and writes a report assessing heart structures and function. This study tested whether an AI tool could analyze the images and draft a preliminary report to assist the cardiologist. The researchers say it is the first randomized trial of an AI tool producing a full echocardiogram report, interpretations and measurements included.
The design used 1,200 echocardiograms already acquired during routine care, performed over two weeks in March 2025 in hospital and outpatient settings across 21 Kaiser Permanente Northern California medical centers. The AI tool, a sonographer and a cardiologist each prepared a preliminary report for every exam. Those drafts were randomly assigned to cardiologists, who finalized them blind to authorship. Nine sonographers, averaging 7.2 years of experience, and 20 cardiologists, averaging 11.7 years, took part. The patients averaged 69 years old; 47% were women and 53% men. About 54% identified as white, 16% as Hispanic, about 16% as Asian, 8.8% as Black and 5.8% as other.
Accuracy was judged by how often cardiologists made substantial changes to a draft before signing off. They changed 50.6% of AI drafts, against 49.7% of drafts from sonographers and cardiologists. That gap met the study's criteria for noninferiority, meaning the AI approach was comparable in accuracy to the human-assisted ones. The AI-assisted reports were also more consistent with historical reports in patients' medical records than those in the comparison group.
The time difference was larger. Cardiologists averaged 4.6 minutes per echocardiogram when reviewing and finalizing an AI draft, compared with 8.5 minutes to produce a report without AI assistance, a saving of about 46%. Co-senior author Keane K. Lee, M.D., M.S., a cardiologist at Kaiser Permanente Santa Clara Homestead Medical Center, said the team expected strong accuracy based on earlier work but was surprised by how much time was saved. Because cardiologists read so many scans, he said, a few minutes each adds up over a day, which could mean faster results and more access for patients, and more time with patients for physicians.
Lee was explicit about limits on the technology's role. The cardiologist remains responsible for interpreting the study and certifying the final report, he said, and applies clinical judgment and knowledge of the patient to build a personalized care plan. The workforce context is relevant: a July 2026 study in the Journal of the American Heart Association cited Health Resources and Services Administration projections of a national shortfall of about 7,000 cardiologists by 2036.
The researchers note two limitations. The trial did not test whether AI-assisted interpretation improves patients' long-term health, and the cardiologists knew they were being studied, so they may have changed their usual habits. The findings are preliminary, and more research is needed on how the approach performs at larger scale and higher volume in real-world clinical settings.
Citations marquantes
AI cannot, however, replace the human team. The cardiologist remains ultimately responsible for interpreting the echocardiogram and certifying the final report.— Keane K. Lee, M.D., M.S., co-senior author, Kaiser Permanente Santa Clara Homestead Medical Center
Envie de l'histoire complète ?
Lire l'original sur Medical Xpress ↗