Mayo AI Flags HCM Obstruction on Resting Ultrasound Videos

A Mayo Clinic AI model scored 0.84 for spotting hypertrophic cardiomyopathy obstruction on resting ultrasound, with no Doppler, in 46 patients at a South Korean hospital.

The tool reads three standard B-mode views and calls a gradient above 20 mm Hg. Drugs, myectomy, alcohol septal ablation, and myosin inhibitors still wait on a Doppler study and, for invasive care, 50 mm Hg.

Resting Gray-Scale Videos Now Carry a Stress-Echo Job

Hypertrophic cardiomyopathy, or HCM, thickens heart muscle. In about 75% of patients the thickened septum and the mitral valve crowd the left ventricular outflow tract, or LVOT, at rest or after a provoke. Blood leaving the heart is squeezed. Chest pain and shortness of breath on effort, or while lying flat, often follow.

The catch in daily practice is that a quiet resting scan can look calm while the gradient is waiting for a Valsalva strain, standing, a meal, or a treadmill. Guidelines already tell labs to provoke when the resting gradient sits under 50 mm Hg, and to add exercise echo when symptoms persist and bedside tricks still look negative. Many community rooms skip that extra step. Alignment for continuous-wave Doppler is fussy. Stress lists are short. Portable probes often never open a Doppler package at all.

Imon Banerjee, Ph.D., an AI researcher at Mayo Clinic in Phoenix and associate professor of radiology, is senior author on the new work. The group asked whether gray-scale video already holds the physiology that Doppler is supposed to measure, including obstruction that may appear only when the heart is loaded.

Measuring LVOT obstruction typically requires Doppler echocardiography, which depends on precise ultrasound-beam alignment and operator expertise. We wanted to determine whether AI could recognize subtle patterns that are imperceptible to the human eye in routinely acquired B-mode ultrasound videos and identify patients with LVOT obstruction earlier, enabling timely confirmatory Doppler evaluation and referral when appropriate.

Imon Banerjee, Ph.D., AI researcher, Mayo Clinic in Phoenix

Mayo’s newsroom said the model can identify obstruction from routine ultrasound videos and that, in a subset, it beat two expert echocardiographers who read the same non-Doppler clips. That comparison has no published score. It still makes the mechanical point: if two experts cannot reliably see obstruction on B-mode, the clinic was never going to replace Doppler with a pair of eyeballs.

The official account posted the same claim on August 19, 2026, the day the briefing went out.

Mayo Clinic researchers have developed and externally validated an artificial intelligence (AI) model that can identify a potentially significant heart obstruction from routine ultrasound videos without relying on specialized Doppler imaging.

The technology could help… pic.twitter.com/69yjQeJx61

— Mayo Clinic (@MayoClinic) August 19, 2026

On September 11, 2026, the same account posted a short video that framed the work as a way to find high-risk HCM earlier. The Society of Diagnostic Medical Sonography passed the briefing along the day before. The people who already record parasternal long-axis and apical windows on every complete echo are the ones who would feed this model. No extra view list is required. The scarce work remains the aligned Doppler cursor and the stress lab.

Why the 20 mm Hg Cut Is Not a Treatment Line

The paper, Beyond Doppler: Scalable AI Detection of LVOT Obstruction in HCM, went online August 6, 2026, in Circulation: Cardiovascular Imaging. Its job is classifying LVOT obstruction above 20 mm Hg from routine two-dimensional windows, with no Doppler input.

That number is not the line cardiologists use to name obstruction, and it is not the line used to offer septal reduction. The 2020 AHA/ACC HCM guideline treats a peak gradient of 30 mm Hg as obstruction present, and it uses a 50 mm Hg threshold for septal reduction in patients whose symptoms stay after drugs. If the resting gradient is under 50 mm Hg, provocative testing should be done. Symptomatic patients who still look negative get an exercise echocardiogram.

THE GRADIENT LINES THAT ACTUALLY CHANGE CARE

A positive model score is therefore a queue ticket. It is not a myectomy decision, and it is not a prescription for a cardiac myosin inhibitor. Banerjee said as much in plainer language: confirmatory Doppler, stress testing, or a move to an HCM specialty center. The study received no outside funding.

Dobutamine is the provoke the guideline tells labs not to use for this question, because it can manufacture a gradient that does not map to daily life. The model never had to take a side in that fight. It only saw resting B-mode.

EchoPrime’s 12 Million Clips Feed a Three-View Fusion

The architecture is a cross-attention video-to-video fusion built on EchoPrime embeddings. EchoPrime is a view-informed vision-language model trained on more than 12 million videos paired with reports, from 275,442 studies in 108,913 patients at Cedars-Sinai Medical Center. Code and weights were released. A view classifier sorts B-mode and color Doppler into 58 standard windows. The Mayo team did not retrain that foundation from scratch. They read three routine transthoracic views through it and fused the clips.

Training used a Mayo derivation cohort of 1,833 patients. Performance was scored on a held-out internal test of 275 patients and on 46 patients from a hospital in South Korea. Generic single-view baselines were weak once the scanners and the population shifted. EchoPrime, already taught on echo video, was stronger on one window. Late fusion of the three windows produced the 0.84 external score that the journal account summarized as parasternal long-axis plus apical views.

SCORES WITH DOPPLER LEFT OUT

The jump from a generic single view to EchoPrime is the large one. Fusion then adds a thin slice, about 0.01 to 0.05 above the best one-view EchoPrime band on the Korean set. Pretraining on echo video did more than the extra windows. The extra windows still mattered, because naive single-view models sat as low as 0.47 once they left Mayo.

The journal abstract also says the fused model showed “substantial cost savings” against a screen-all Doppler approach. No dollar figure, no test-count, and no time-per-scan sat in the public summary, so that claim stays qualitative. What the table does show is a triage score, not a measured gradient in mm Hg.

Forty-Six Korean Exams Against a Mayo Training Set

External validation is the sentence every methods paper needs, and 46 exams is a short sentence. The authors call the Korean set a real population shift and still report 0.84. A set that small cannot settle how the score moves by machine vendor, by probe frequency, by sonographer habit, or by the mix of resting versus provocable obstruction. It can only show that the signal did not collapse on first contact with another hospital.

WHAT THE PUBLIC FILE STILL CANNOT SETTLE

• External size: Forty-six patients is a first look, not a multi-vendor trial.
• Gradient target: The model calls above 20 mm Hg, not the 30 mm Hg or 50 mm Hg lines that change drugs and surgery.
• Human baseline: A win over two experts on non-Doppler clips is described, without a published accuracy for either side.
• Stress-only cases: The briefing says the model helped flag obstruction that may appear only under stress, without a count of those patients.
• Deployment: No prospective run, no listed ultrasound platforms, and no clearance as a medical device in the briefing.

Those limits fit the silence around the paper. Three weeks after Mayo’s note, and a month after the journal posting, there is no loud argument among practicing echocardiographers on the record. The result still reads as a methods paper that has not yet been asked to change a lab’s order set.

The Filter Fits the Handheld Probe Mayo Already Uses

Banerjee’s own next sentence is the deployment map. The same resting clips could be read on portable ultrasound, or in rooms where a full Doppler exam is not sitting on the cart. That is a different customer from the HCM center that already runs treadmill echo and knows how to align a continuous-wave beam down the LVOT.

Mayo has already tested the first half of that stack. In a separate ECG-lab project, novices acquired point-of-care cardiac ultrasound on 486 outpatients. An AI electrocardiogram alone had a 32% positive predictive value for selected structural heart disease, including thick walls that can mean HCM or amyloid. Adding the handheld clips raised that figure to 64%, and overall accuracy moved from 67% to 88%, with a number-needed-to-screen of 8 for one patient who then needed a formal echocardiogram. A later novice study used Philips Lumify with real-time AI guidance on 995 adults and kept a two-step path, AI-ECG first, focused ultrasound second.

Those projects hunt disease. They do not grade an LVOT gradient. The new model is the missing middle: once a handheld or a limited cart study has already been recorded, the same B-mode loops can be scored for obstruction physiology before anyone books a stress echo. Complete TTEs already store the three windows. Handheld exams often store at least a parasternal long-axis. The model’s appetite for three views will be tighter on a two-clip pocket study than on a full lab tape, and that gap is still unmeasured.

Prospective Tests Across Machines Come Next

Banerjee said the next work is prospective validation across broader clinics, ultrasound platforms, and patient groups. Until those runs exist, a lab that turns the score into an automatic stress-echo order will be guessing at false-positive load. A lab that ignores the score will keep missing provocable obstruction in rooms that never open Doppler.

The useful use, on the evidence in hand, is a sort. Resting B-mode that already lives on the archive gets a probability of a gradient above 20 mm Hg. High scores go to confirmatory Doppler, and if the resting number stays shy of 50 mm Hg, to a real provoke. Low scores do not certify a nonobstructive heart. They only say this particular resting tape did not look like the tapes the network was trained to call positive.

That is a narrower product than the headline. It is also the product the numbers describe.

Frequently Asked Questions

How Common Is Hypertrophic Cardiomyopathy in Young Adults?

The 2020 AHA/ACC guideline cites unexplained asymptomatic hypertrophy in young adults in the United States at 1 in 200 to 1 in 500, while symptomatic hypertrophy inferred from medical claims has been estimated at under 1 in 3,000 adults, with a larger unrecognized burden in the general population.

Why Do Guidelines Warn Against Dobutamine to Provoke an HCM Gradient?

The AHA/ACC writing committee advises against dobutamine when the question is provocable LVOT obstruction and eligibility for septal reduction, citing a lack of specificity; British echo guidance makes the same call because the infusion can induce outflow obstruction even in people who do not have HCM.

What Is Systolic Anterior Motion of the Mitral Valve?

Systolic anterior motion, or SAM, is the mitral valve being pushed toward the septum in systole, which both narrows the outflow tract and lets the mitral leaflets fail to coapt, so a posterior or lateral mitral regurgitant jet often rides along with the gradient.

How Often Should Clinically Stable HCM Patients Repeat an Echocardiogram?

Key perspectives from the 2020 AHA/ACC guideline say clinically stable patients should be considered for surveillance echocardiograms every 1 to 2 years to watch hypertrophy, obstruction, and cardiac function, with a shorter path if symptoms change.

Disclaimer: This article is news reporting and analysis of a published research study and related clinical guidelines. It is informational only and is not medical advice, a diagnosis, a treatment recommendation, or a device endorsement. Readers should not change imaging, referral, medication, or procedure plans based on this summary and should consult a cardiologist or other qualified clinician who can review their own studies. Figures, cohort sizes, scores, and regulatory status reflect the cited papers and briefings as dated in the text and may change with later validation or product review.