A model that reads ordinary ECGs identified up to 81% of patients with heart failure and up to 90% of patients with valve disease in a 67,000-patient trial in the United States, returning a result in under two seconds. The work was funded by the British Heart Foundation, with the analysis led at Imperial College London, and was presented to thousands of delegates at the annual congress of the European Society of Cardiology in Munich, the largest cardiology conference in the world. The model does not diagnose anything. What it changes is the order of the queue.
The distinction matters more than the accuracy figures. An ECG records the heart's electrical activity — rate and rhythm — and has been used for about a century to diagnose heart attacks and rhythm disorders. It does not, on its own, reveal heart disease. That requires echocardiography, an ultrasound scan patients often wait months to receive. The new system pulls more out of the same standard trace than a human reader can see, and flags the signatures of heart failure and valve disease, two of the most common forms of heart disease, almost immediately.
It cannot confirm or exclude either condition. It only indicates that the probability is high. The clinical use is triage: patients the model marks as high-risk are moved to echocardiography ahead of the standard queue, so the diagnosis lands sooner and treatment starts sooner.
Sonya Babu-Narayan, consultant cardiologist and clinical director of the British Heart Foundation, which funded the trial, said the system produces a result from an ECG almost instantly. AI-ECG will not find every person with heart disease, she said, but it can help refer more quickly those whose probability of disease is especially high, and earlier diagnosis and treatment of heart conditions save lives and improve them.
Fu Siong Ng, professor of cardiology at Imperial College London, said patients frequently wait several months for a cardiac ultrasound after they are referred for one, and that the technology can identify the people at greatest risk of heart failure and valve disease so they can be scanned sooner.
The scale argument rests on a single number: roughly a billion ECGs are performed worldwide every year. The immediate plan is to speed up diagnosis in patients already suspected of these conditions, but Ng pointed to a second use — running the model across every ECG a hospital performs and flagging the highest-risk patients, including those whose doctor was not looking for heart disease at all. Early detection of heart failure and valve disease means finding the people who need life-saving drugs before their condition deteriorates dangerously.
Ahmed El-Medany, a British Heart Foundation research fellow who led the Imperial College London analysis, called the system "superhuman AI". The next step, he said, is portable ECG readers with AI built in that healthcare professionals can carry.
The same congress heard a second detection result in the same shape: researchers from the University of Tokyo and Institute of Science Tokyo reported that AI analysis of a five-minute video of a person's face can quickly and accurately pick up previously undiagnosed high blood pressure and type 2 diabetes, conditions millions of people carry without knowing.
Here is where I part company with the framing. "Superhuman" is a strong word for a flag that cannot confirm a diagnosis, and the numbers being quoted are sensitivity — the share of sick patients the model catches. Sensitivity alone is the easy half of the problem. Notably absent from the announcement is the other half: how many patients the model flags who turn out to have nothing wrong. A triage tool that is generous with its flags relieves no queue; it lengthens the one it was built to shorten. The two figures also arrive as "up to 81%" and "up to 90%", which is the phrasing of a best case across subgroups rather than a single headline result.
There is a second gap worth naming. The trial ran on 67,000 patients in the United States. The bottleneck the researchers describe — months-long waits for an echocardiogram after referral — is the one their own health system has, and the funder, the analysis and both cardiologists quoted are British. A model trained and tested on one population and pitched at another country's waiting list is a reasonable plan, not a completed one.
None of which makes the result small. Reading structural disease off an electrical trace is a genuine gain, and the billion ECGs a year are already being taken, already paid for, already sitting in hospital systems doing nothing beyond what they were ordered for. But the constraint that hurts patients is not the shortage of ECGs. It is the number of echocardiography slots, and that number does not change because software got better at picking who should be in them. Every patient the model correctly moves to the front of the queue moves someone else back.