Mayo AI Flags HCM Obstruction From Everyday Ultrasound Videos

Mayo Clinic researchers have built and externally validated a multi-view AI model that flags potentially significant left ventricular outflow tract obstruction from routine resting B-mode ultrasound videos alone, without any Doppler data. Published in Circulation: Cardiovascular Imaging, the work used 1,833 Mayo patients for development, held out 275 for internal testing, and checked 46 patients from a South Korean hospital. The late-fusion version reached an external AUROC of 0.84 despite clear population shift.

That performance opens a second-order path: earlier triage that can cut the volume of full comprehensive rest-and-provocative echocardiograms and extend screening into rooms that lack expert Doppler operators.

The Numbers the Model Posted

Internal prevalence ran high. Of the 1,833 derivation patients, 1,387 (75.6 percent) had a resting or provoked LVOT gradient above 20 mmHg. The external Korean cohort was leaner at 26.1 percent positive. Single-view baselines struggled outside the training distribution, with external AUROCs as low as 0.47-0.70 for ordinary 2D CNNs and MedCLIP probes.

Domain-specific pretraining changed the picture. Linear probes on EchoPrime video embeddings alone delivered internal single-view AUROCs of 0.75 (apical four-chamber), 0.76 (apical three-chamber) and 0.80 (parasternal long-axis). External numbers held or improved to 0.79-0.83. Multi-view fusion pushed further.

The late-fusion meta-learner balanced the complementary error patterns of the two view-pair models, cutting the false-positive/false-negative trade-off that each pair showed alone. Calibration curves and Brier scores supported the probability outputs on both cohorts.

Three Standard Views and the Foundation Model

The pipeline never saw spectral Doppler. It took three routinely acquired transthoracic windows that any basic echo study already records:

• Parasternal long-axis (PLAX) for mitral-valve motion, septal thickness and direct LVOT geometry
• Apical three-chamber (AP3) for additional mitral-septal interaction and flow-direction cues
• Apical four-chamber (AP4) for global left-ventricular shape and atrioventricular context

A prior view-classifier kept only clips with ≥0.80 confidence. Each 16-frame clip (down-sampled from 32 contiguous frames) was passed through the frozen video encoder of EchoPrime, the largest echocardiography foundation model trained on more than 12 million video-report pairs from Cedars-Sinai. Bidirectional cross-attention then let PLAX embeddings exchange information with each apical view, after which a logistic meta-learner combined the two pair-wise probability outputs.

Occlusion saliency maps later showed the network concentrating on end-systolic LVOT, interventricular septum and mitral apparatus-exactly the anatomy a human reader would watch for systolic anterior motion.

What the Cost Tree Shows

The second-order payoff appears in the decision analysis. For a hypothetical 1,000 patients already receiving a standard transthoracic echo, a two-stage AI path (first a previously validated HCM detector, then this gradient model) funnels only the positives into comprehensive rest-and-provocative testing. Using Mayo Clinic internal cost estimates, that sequence lands at roughly $1.104 million versus $3 million for the screen-all approach-an approximate threefold reduction-while preserving clinical detection rates on the external operating point.

~$1.1 M AI-guided path per 1,000 patients
~$3.0 M universal comprehensive rest-and-provocative TTE
~3× reduction in estimated spend
Medicare/Medicaid unit costs produced a smaller but still clear gap ($559,500 vs $750,000).

The tree deliberately ignored real-world wait times and travel to specialty labs, so the true access gain is likely larger.

Harder Than Experts Expect on B-Mode Alone

In a subset comparison the AI outperformed two board-certified expert echocardiographers who read the identical non-Doppler video sets. The release notes that the result “highlight[s] how difficult it can be to recognize LVOT obstruction from routine two-dimensional images without Doppler measurements.”

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 at Mayo Clinic in Phoenix, senior author

Banerjee, an AI researcher at Mayo Clinic in Phoenix, stresses the tool is meant to complement, not replace, Doppler. The model simply raises the flag so the specialized measurement can be ordered sooner or the patient referred to an HCM center.

HCM Obstruction and Why the Cutoff Matters

Hypertrophic cardiomyopathy thickens the heart muscle, most often the septum. Roughly two-thirds of patients develop dynamic LVOT obstruction that can restrict blood leaving the left ventricle, producing exertional chest pain, dyspnea or symptoms when lying flat. Presence and severity drive decisions on myosin inhibitors, septal reduction and lifestyle advice. Prevalence estimates place the condition between 1 in 200 and 1 in 500 people; many remain undiagnosed.

Guidelines have long used a 30 mmHg threshold to define obstruction. The Mayo team deliberately trained at >20 mmHg because that lower cutoff matches the stability criterion used in recent myosin-inhibitor trials and therefore has direct relevance for dose adjustment and monitoring. Gradients were taken as the highest value at rest or after Valsalva, squat-to-stand or amyl nitrite. Images came from expert-acquired studies across Mayo Rochester, Arizona, Florida and the health system between 2010 and early 2023.

Readers seeking the clinical picture can start with Mayo’s own overview of hypertrophic cardiomyopathy symptoms and causes.

Portable Rooms and Specialty Referrals

Because the input is ordinary B-mode video, the same pipeline can in principle run on handheld or cart-based systems that already capture PLAX, AP3 and AP4. In settings where continuous-wave Doppler expertise or the time for full provocative maneuvers is scarce, the AI can still surface patients who warrant the extra study or a trip to a specialty HCM clinic. That is the access expansion Banerjee highlights: “It also could support evaluation using portable ultrasound or in settings where comprehensive Doppler assessment may not be readily available.”

Mayo’s own X account posted the release with modest early engagement, framing the work around flagging patients who need more testing. Early specialist conversation has stayed quiet and practical, centered on resource-limited deployment rather than any claim that Doppler is obsolete. The paper itself lists the next requirements: prospective validation across more ultrasound vendors, community hospitals, novice operators and continuous (not binary) gradient prediction, plus checks at the classic 30 mmHg and 50 mmHg therapeutic cut-points.

The technical dependence that has always limited LVOT assessment is the need for precise Doppler ultrasound beam alignment. By reading motion patterns already present in the B-mode clips every lab records, the multi-view model removes that requirement from the first screening step. The late fusion AUROC of 0.84 on an external Asian GE-only cohort shows the signal travels across populations and platforms. What remains is proving the same numbers hold when the probe is held by a less experienced hand or when the machine is a pocket device.

Until those prospective data arrive, the published study already supplies a concrete, costed pathway that turns every standard echo into a potential obstruction screen.

Frequently Asked Questions

What is left ventricular outflow tract obstruction in hypertrophic cardiomyopathy?

LVOT obstruction occurs when thickened septal muscle and systolic anterior motion of the mitral valve narrow the path blood takes out of the left ventricle, raising the pressure gradient. It is dynamic, can appear only with provocation, and is present in roughly two-thirds of HCM patients either at rest or with maneuvers; it drives symptoms and treatment choice.

Why did the Mayo model use a 20 mmHg cutoff instead of the traditional 30 mmHg?

The 20 mmHg threshold aligns with the stability definition used in clinical trials of cardiac myosin inhibitors, giving the binary classifier direct relevance for medication titration and monitoring. Earlier literature commonly used 30 mmHg simply to define presence of obstruction.

Which ultrasound views does the AI require?

Only three standard non-Doppler B-mode clips: parasternal long-axis, apical three-chamber and apical four-chamber. A separate view classifier discards low-confidence acquisitions before the gradient model runs.

How does the multi-view fusion improve on single-view performance?

Bidirectional cross-attention lets the PLAX embedding exchange features with each apical view, then a late meta-learner combines the two pair-wise probabilities. The combination balanced false-positive and false-negative rates that each pair exhibited alone and lifted external AUROC from the mid-0.80s single-view range to 0.84-0.86.

Is the AI intended to replace Doppler echocardiography?

No. The authors state it is designed to flag patients who should receive confirmatory Doppler, stress testing or specialty referral earlier, especially where skilled Doppler operators or full provocative protocols are not immediately available.

Disclaimer: This article is news reporting and analysis of a published research study for general informational purposes only. It does not constitute medical advice, diagnosis, treatment recommendations, or clinical decision support for any individual. Readers should consult a qualified cardiologist or other licensed physician before making any health-related decisions or interpreting ultrasound or AI results. All figures, performance metrics and status details reflect the cited sources as of the article date and may be updated by subsequent validation studies or regulatory review.