Patterns so subtle that human eyes would miss them entirely
For generations, medicine has treated the electrocardiogram as a window into the present — a snapshot of the heart's rhythm in the moment of testing. Researchers at Mass General Brigham now propose it may also be a window into the future: an AI model called ECG2Stroke, trained on more than 200,000 patient records, can estimate the likelihood of ischemic stroke up to a decade before it occurs, using only the routine 12-lead ECG and basic demographic data. Validated across three major hospitals and nearly 200,000 patients, the tool performs with an accuracy comparable to far more complex clinical methods, particularly for strokes caused by clots originating in the heart. If prospective studies confirm its promise, medicine may gain something rare — the ability to intervene before catastrophe, rather than after.
- Stroke strikes without warning for millions of people who never knew they were at high risk, making early detection one of medicine's most urgent unmet needs.
- Existing stroke risk tools require complex calculations that don't fit naturally into clinical workflows, so they are rarely used — leaving a critical prevention gap in everyday care.
- ECG2Stroke cuts through that barrier by working with equipment hospitals already use constantly, requiring nothing more than a standard electrocardiogram and a patient's age and sex.
- The model proved especially powerful at predicting cardioembolic strokes — those triggered by clots forming in the heart — by detecting subtle electrical signals in the atrium invisible to the human eye.
- Validated across over 199,000 patients at three major medical centers, the algorithm held its predictive accuracy across different hospitals and diverse populations.
- Clinical adoption still awaits prospective real-world trials, but the research signals a broader shift: AI transforming ordinary tests into instruments of anticipation rather than mere diagnosis.
An electrocardiogram takes only minutes, and hospitals run them constantly — a routine check of the heart's electrical activity. Researchers at Mass General Brigham now believe that same familiar test could do something far more consequential: warn of a stroke up to ten years before it happens.
The team developed an AI model called ECG2Stroke, trained on more than 200,000 clinical records, that estimates ischemic stroke probability using only a standard 12-lead ECG and basic information like age and sex. Published in JACC, the study found the model's predictive accuracy rivals far more complex traditional methods. The AI detects electrical patterns so subtle that human eyes would miss them entirely — patterns that correlate with elevated stroke risk years into the future.
The appeal lies in its simplicity. Neurologist and coauthor Rahul Mahajan noted that existing stroke risk scales demand complicated math and rarely make it into daily clinical practice. A tool that works with what hospitals already do could change that entirely. The algorithm was validated across Massachusetts General Hospital, Brigham and Women's Hospital, and Beth Israel Deaconess Medical Center — collectively more than 199,000 patients — and held steady across different institutions and populations.
One particularly striking finding: ECG2Stroke excelled at predicting cardioembolic strokes, the kind caused by clots that form in the heart and travel to the brain. Many such clots are linked to atrial fibrillation, and the model detected electrical signals tied to changes in the heart's upper chamber even among patients without a known arrhythmia diagnosis.
The authors are clear that widespread clinical use requires further prospective validation in real-world conditions. But the research reflects a growing momentum — AI finding signals invisible to human perception in ordinary clinical tests, shifting medicine's posture from reaction to anticipation.
An electrocardiogram takes minutes. Hospitals and clinics run them constantly—a quick way to check heart rhythm, spot irregularities, monitor cardiac trouble. Researchers at Mass General Brigham now believe that same routine test could do something far larger: predict whether someone will suffer a stroke a decade before it happens.
The team developed an artificial intelligence model called ECG2Stroke that estimates the probability of ischemic stroke up to ten years into the future using only a standard electrocardiogram plus basic demographic information—age, sex. The work, published in JACC, trained on more than 200,000 clinical records and demonstrated predictive power that rivals far more complex traditional medical methods. What appears to most people as lines and peaks on a screen or printout actually contains vast amounts of information about the heart's electrical activity. The AI learned to detect patterns so subtle that human eyes would miss them entirely—patterns that correlate with elevated stroke risk years down the road.
Unlike many established medical tools, ECG2Stroke requires no invasive procedures, no complicated calculations. A routine electrocardiogram and demographic data are enough. The model analyzes the standard 12-lead format used in everyday clinical practice. This simplicity matters. Rahul Mahajan, a neurologist and study coauthor, noted that existing stroke risk scales demand complex math and don't scale easily into daily practice, which is why they're barely used at all. A tool that works with what hospitals already do could change that.
The researchers validated the algorithm across three major medical centers. Massachusetts General Hospital contributed data on more than 100,000 people. Brigham and Women's Hospital added nearly 69,000. Beth Israel Deaconess Medical Center brought in close to 30,000. Over a ten-year follow-up period, thousands of strokes occurred, allowing the team to compare ECG2Stroke's accuracy against established clinical methods. The model held steady across different hospitals and different populations.
One striking finding: the tool proved especially skilled at predicting cardioembolic stroke—the kind that happens when clots form inside the heart, travel to the brain, and block blood vessels. Many of these clots link to rhythm disturbances like atrial fibrillation. ECG2Stroke detected electrical signals tied to changes in the atrium, the heart's upper chamber. Cardiologist Shaan Khurshid, another coauthor, suggested the tool might also help clarify cardiac abnormalities that remain poorly understood. The model maintained stable correlation even among people with and without atrial fibrillation.
Stroke itself is straightforward and brutal. Blood flow to part of the brain stops or drops sharply. Neurons lose oxygen and nutrients. Permanent damage begins within minutes. Ischemic strokes—the majority of cases—usually result from clots blocking cerebral arteries. Hypertension, diabetes, smoking, high cholesterol, certain heart problems: all raise the risk. The cruel part is that many people don't know they're at high risk until the acute event arrives. A predictive tool that identifies vulnerable patients years early could shift prevention from reaction to anticipation, allowing intensive monitoring or targeted treatment before crisis strikes.
The authors are careful to note that ECG2Stroke still needs more validation before it enters widespread clinical use. So far it has been tested on existing hospital records. The next phase involves prospective studies—real-world conditions, real patients, real workflows. But the study reflects a growing momentum in medicine: using AI to spot signals invisible to human eyes in ordinary clinical tests. If future work confirms its clinical value, models like this could identify high-risk people sooner and enable more precise preventive interventions.
Citações Notáveis
Existing stroke risk scales demand complex calculations and don't scale easily into daily practice, which is why they're barely used at all.— Rahul Mahajan, neurologist and study coauthor
A Conversa do Hearth Outra perspectiva sobre a história
Why does a heart test predict a stroke, which is a brain event? Aren't those separate systems?
They're connected more than most people realize. Many strokes come from clots that form in the heart and travel to the brain. The electrical patterns in a heart scan can hint at conditions that make clotting more likely—especially rhythm problems like atrial fibrillation.
So the AI isn't really predicting stroke. It's detecting heart problems that lead to stroke?
Partly. But it's also finding patterns so subtle that cardiologists wouldn't catch them on their own. The model trained on 200,000 records and learned to see connections between electrical signals and future stroke risk that human analysis would miss.
If it's that good, why isn't it already in hospitals?
Because it's only been tested on old records. Doctors need to see it work in real time, with real patients, in actual clinical settings. That's the next step. And there's always caution with AI in medicine—you have to prove it works before you change how people are treated.
What happens to someone if the tool flags them as high-risk?
That's still being figured out. Theoretically, they'd get more intensive monitoring, maybe preventive medications, lifestyle interventions. But the tool itself doesn't treat anyone. It just identifies who needs closer attention.
Does it work for all kinds of stroke?
It's especially good at cardioembolic stroke—clots from the heart. It's less clear how well it predicts other types. That's another reason more testing is needed.
Ten years is a long time to predict something. How confident should we be in that?
Confident enough to pay attention, but not confident enough to act on it alone. The model showed it can match traditional risk scales, which is impressive. But matching isn't the same as proving it saves lives. That's what prospective studies will show.