In a subset of cases, an AI model caught a dangerous heart obstruction more accurately than two expert echocardiographers looking at the exact same images.
Mayo Clinic researchers developed and externally validated the AI model to detect left ventricular outflow tract obstruction, a complication that affects roughly two-thirds of patients with hypertrophic cardiomyopathy. The findings were published in Circulation: Cardiovascular Imaging.
What the AI Actually Found
Hypertrophic cardiomyopathy is a genetic condition that thickens the heart muscle abnormally. The outflow tract obstruction it often causes restricts blood leaving the heart, triggering chest pain and shortness of breath during exertion or while lying flat.
Which patients develop that obstruction matters directly for treatment decisions and long-term management, not just diagnosis. Diagnosing it normally requires Doppler echocardiography, a technique that depends on precise ultrasound-beam alignment and real operator skill.
"We wanted to determine whether AI could recognize subtle patterns that are imperceptible to the human eye in routinely acquired B-mode ultrasound videos," said Imon Banerjee, the study's senior author and an AI researcher at Mayo Clinic in Phoenix.
Why This Matters for HCM Patients
The model used only resting, non-Doppler ultrasound videos, the kind that get captured far more routinely than specialized Doppler imaging. Combining three standard ultrasound views improved its accuracy, and it also picked up obstructions that only appear when the heart is under stress.
The study tested the model on 275 patients from Mayo's own cohort of 1,833, then externally validated it on 46 patients at a hospital in South Korea. It held up well despite real differences between that population and the one it was trained on, a meaningful signal for whether it could generalize to other hospitals and patient groups.
The Model Beat Human Experts, Sometimes
The comparison against human echocardiographers is the most striking part. In a subset of cases, the AI identified the obstruction more accurately than two expert reviewers working from the same non-Doppler images.
Banerjee was careful about what that result should and shouldn't mean.
"This technology is intended to complement, not replace, Doppler echocardiography," Banerjee said. "By enabling earlier identification of patients with potential LVOT obstruction, it could prompt confirmatory Doppler measurements, stress testing, or referral to an HCM specialty center."
The tool could be especially useful in settings without ready access to comprehensive Doppler assessment, including care delivered through portable ultrasound.
That access gap is real. Doppler echocardiography requires both specialized equipment and an operator experienced enough to align the ultrasound beam precisely, a combination that isn't available everywhere routine ultrasound imaging already is. A model that works off standard, widely available video could meaningfully widen where this kind of screening happens at all.
The study received no external funding, and Banerjee said the next step is broader prospective validation across more clinical settings and ultrasound platforms.
What This Means for Miami
South Florida's older population carries a higher cardiovascular disease burden than most of the country, and the region's hospital systems already run substantial cardiac care programs that could plausibly evaluate a tool built for exactly this kind of early screening gap.
This is also a useful real-world example of the kind of AI reliability research MAIN has covered from McGill University this year, work aimed at getting AI systems to flag uncertainty rather than answer with false confidence. A model validated against expert clinicians on a specific, well-defined task, rather than a general-purpose chatbot answering open-ended medical questions, is the version of medical AI worth taking seriously as Miami's healthcare AI sector, including companies like eMed, continues to grow.