At Mayo Clinic, researchers have taught an artificial intelligence to read what human eyes often cannot — the faint signatures of a dangerous heart obstruction hidden within ordinary ultrasound footage. The condition, hypertrophic cardiomyopathy, quietly narrows the heart's outflow in millions of people, many of whom go undetected until symptoms become serious. This work suggests that the boundary between specialized and routine medicine may be narrowing, and that early warning need not depend on the presence of an expert.
Mayo Clinic AI Detects Heart Obstruction from Routine Ultrasounds
Patients with undetected hypertrophic cardiomyopathy obstruction experience chest pain, shortness of breath, and risk of serious cardiac complications.
The AI outperformed expert cardiologists reviewing the same images
Mark
Why does it matter that the AI sometimes outperformed the expert cardiologists?
Mimi
It shows how genuinely hard this diagnosis is to make from the images alone. Those experts were looking at the same ultrasound videos everyone else sees. They missed things the machine caught. That's not a failure of the doctors—it's a signal that the pattern is subtle enough that human perception has limits.
Mark
But if the AI is better, why not just use it instead of Doppler?
Mimi
Because Doppler actually measures the blood flow velocity directly. It's the gold standard. The AI is pattern recognition. It's very good at saying "something looks off here," but it can't replace the physical measurement. Think of it as a smoke detector, not a fire investigation.
Mark
The South Korean validation is interesting. Why does that matter?
Mimi
It means the model didn't just memorize the Mayo Clinic patients. It learned something general about how obstruction looks on ultrasound, even in people with different genetics, body types, and healthcare backgrounds. That's the difference between a neat lab result and something that might actually work in hospitals worldwide.
Mark
What happens to a patient who has HCM but no obstruction detected?
Mimi
They still have the genetic condition. But they're lower risk for the acute symptoms—the chest pain, the breathing problems. The obstruction is what makes it dangerous and urgent. So missing it means missing the patients who need aggressive management.
Mark
Where does this get deployed first?
Mimi
Probably in places where cardiologists are scarce. Rural hospitals. Developing countries. Clinics with portable ultrasound machines. Anywhere the bottleneck is expertise, not equipment. That's where the AI adds the most value.
The Pulse
- Hypertrophic cardiomyopathy obstruction causes chest pain and breathlessness in roughly two-thirds of HCM patients, yet it often goes undetected because diagnosis requires specialized Doppler imaging most clinics don't routinely deploy.
- The AI model, trained on nearly 2,000 Mayo Clinic patients, learned to detect subtle obstruction patterns in standard ultrasound video — the kind already captured in clinics worldwide — without any specialized equipment.
- In head-to-head comparisons, the model outperformed experienced echocardiographers reviewing the same non-Doppler images, exposing just how elusive this diagnosis is without the right tools.
- Validation across a South Korean patient cohort confirmed the model's ability to generalize beyond the population it was trained on, a critical hurdle for real-world clinical adoption.
- The technology is designed not to replace specialist care but to trigger it — flagging at-risk patients earlier so that Doppler testing, stress evaluation, or specialist referral can follow before complications arise.
- Broader prospective trials across varied clinical settings and ultrasound hardware are now planned, with the goal of determining whether this early promise survives contact with the full complexity of real-world medicine.
At Mayo Clinic, researchers have taught an artificial intelligence to read what human eyes often cannot — the faint signatures of a dangerous heart obstruction hidden within ordinary ultrasound footage. The condition, hypertrophic cardiomyopathy, quietly narrows the heart's outflow in millions of people, many of whom go undetected until symptoms become serious. This work suggests that the boundary between specialized and routine medicine may be narrowing, and that early warning need not depend on the presence of an expert.
A Mayo Clinic research team has developed an AI system capable of detecting a serious heart obstruction from the kind of routine ultrasound video available in virtually any clinic — no specialized equipment required. The target condition, hypertrophic cardiomyopathy, causes the heart muscle to thicken abnormally, and in most patients this produces a blockage that restricts blood flow out of the heart. The result is chest pain, breathlessness, and elevated cardiac risk — consequences that hinge on whether the obstruction is caught at all.
Traditionally, identifying this blockage demands Doppler echocardiography, a technique requiring precise probe placement and skilled interpretation. AI researcher Imon Banerjee asked whether a machine could learn to find obstruction signals in standard, non-Doppler footage — patterns too subtle for even experienced cardiologists to reliably see. Training on over 1,800 patients and combining data from three standard viewing angles, the model learned to predict significant obstruction, including blockages that only emerge under cardiac stress.
The results held up across borders: the model performed strongly on a South Korean validation cohort despite meaningful differences from the American training population. In direct comparisons, the AI surpassed two expert echocardiographers reviewing the same non-Doppler images — a result that speaks less to the machine's brilliance than to the genuine difficulty of the task without Doppler's specialized measurements.
Banerjee frames the technology as an early alert system rather than a replacement for specialist care. A routine scan flags a concern; that concern prompts the deeper investigation. In settings where cardiologists are scarce or only portable ultrasound is available, this could meaningfully expand the reach of early HCM screening. The findings, published in Circulation: Cardiovascular Imaging, now set the stage for broader prospective trials testing whether the model's promise survives the messier conditions of everyday clinical practice.
A team at Mayo Clinic has built an artificial intelligence system that can spot a dangerous heart obstruction in ordinary ultrasound videos—the kind of scan a technician might perform in any clinic, without special equipment or advanced training. The discovery matters because it could catch a serious condition earlier, in places where expert heart specialists are scarce or unavailable.
The condition is called hypertrophic cardiomyopathy, or HCM. It's genetic. The heart muscle thickens abnormally, and in about two-thirds of patients, this thickening narrows the passage where blood leaves the heart—a blockage called left ventricular outflow tract obstruction, or LVOT. When that happens, people feel chest pain and struggle to breathe, especially when they exert themselves or lie down. Knowing which HCM patients have this obstruction is crucial because it changes how doctors treat them and what they watch for over time.
Traditionally, detecting this obstruction requires Doppler echocardiography, a specialized ultrasound technique that demands precise positioning of the probe and considerable skill from the operator. Imon Banerjee, an AI researcher at Mayo Clinic in Phoenix, wondered whether a machine could learn to see what human eyes miss in standard ultrasound videos—subtle patterns that hint at obstruction, patterns too faint or complex for even experienced cardiologists to reliably spot.
The researchers trained their model on 1,833 patients from Mayo Clinic's records. They tested it on 275 of those patients, then validated it on 46 patients from a hospital in South Korea. The AI looked only at basic, non-Doppler ultrasound footage—the kind any clinic has. By combining information from three standard viewing angles, the model learned to predict whether a patient had significant obstruction. It could even identify blockages that only appear when the heart is stressed.
The results were striking. The model worked well in the South Korean group despite major differences between those patients and the American cohort used to train it. In some cases, the AI outperformed two expert echocardiographers who reviewed the same non-Doppler images. That gap reveals how difficult the diagnosis truly is without Doppler's specialized measurements—a finding that underscores why the technology matters.
Banerjee is careful about the technology's role. It is not meant to replace Doppler echocardiography. Instead, it acts as an early alert system. A patient gets a routine ultrasound. The AI flags a potential problem. That prompts the clinic to order the specialized Doppler test, stress testing, or a referral to an HCM center. In settings where expert cardiologists are thin on the ground—or where only portable ultrasound machines are available—this could be transformative. It expands the net for catching patients who might otherwise slip through.
The work was published in Circulation: Cardiovascular Imaging. Banerjee and his team are now planning broader prospective studies across different clinical settings, ultrasound machines, and patient populations. The next phase will test whether this promise holds up in the real world, where conditions are messier and more varied than any training set can capture.
Notable Quotes
By enabling earlier identification of patients with potential obstruction, it could escalate timely detection and support evaluation in settings where comprehensive Doppler assessment may not be readily available.— Imon Banerjee, Ph.D., Mayo Clinic AI researcher
Want the full story?
Read the original at News-Medical ↗
Technology