An artificial intelligence model developed at Imperial College London can scan a standard electrocardiogram and flag signs of heart failure and valve disease in under two seconds, catching warning signs that are invisible to the naked eye even for trained cardiologists. The research, led by Dr Ahmed El-Medany, a British Heart Foundation clinical research fellow, and senior investigator Professor Fu Siong Ng, was presented on August 31, 2026 at the European Society of Cardiology Congress in Munich. In testing, the AI correctly identified up to 81% of patients with heart failure caused by reduced pumping function of the heart’s main chamber, and up to 90% of patients with aortic stenosis, a common and serious heart valve disease, directly from a routine ECG trace.
A Decades-Old Test Gets a New Reader
The ECG has barely changed since the early 20th century: electrodes on the chest, a wavy line on paper or screen, and a clinician trained to spot abnormal rhythms. For most of that history, the test has been read the same way by generations of doctors, catching obvious arrhythmias but missing subtler structural problems that don’t announce themselves in the waveform. Heart failure and valve disease, in particular, are notoriously easy to miss early because symptoms like breathlessness and fatigue are nonspecific and often blamed on aging or deconditioning. By the time patients get an echocardiogram, a more expensive and less accessible scan that can directly visualize pumping function and valve damage, the disease has frequently progressed. The Imperial team’s bet is that the ECG itself has always contained more diagnostic information than the human eye can extract, and that machine learning can finally unlock it.
How the Model Was Trained and Tested
El-Medany’s team built the model on a training dataset of 10.6 million ECGs, with roughly 72,475 of those records paired to confirmed echocardiogram results so the AI could learn to associate waveform patterns with actual structural heart disease. The model was then validated against two independent groups of US hospital patients, one of 5,442 people and a much larger group of 61,520, to confirm results held up outside the training population. Performance metrics reported at the ESC Congress showed discrimination scores of 0.86 to 0.9 for detecting reduced heart pumping function and 0.73 to 0.85 for aortic stenosis, shorthand for how reliably the model separates sick patients from healthy ones. Detection rates varied by cohort, with heart failure identification running as high as 81% and aortic stenosis detection reaching 90%, well above chance but still short of a stand-alone diagnostic.
What the AI Sees That Cardiologists Don’t
Dr El-Medany described the findings bluntly: “These results suggest there is potentially far more information hidden within a routine ECG than we can recognise by looking at it ourselves.” That is the central claim driving AI-ECG research across cardiology right now, not just at Imperial but at institutions including Mayo Clinic and UT Southwestern, which have published their own 2026 studies on AI-enabled ECGs detecting atrial fibrillation, sleep apnea and stroke risk from the same simple test. The appeal is partly practical: an ECG costs a fraction of an echocardiogram, takes minutes rather than an appointment with a sonographer, and is already routine in emergency rooms, primary care clinics and pre-operative assessments worldwide. If an algorithm can triage which traces deserve a follow-up scan, it could compress the time between a hidden problem and a confirmed diagnosis from years to days.
From the Lab to the Hospital Floor
The research has already moved past retrospective data analysis into a live NHS trial. The AI is currently being tested on ECGs from 590 real patients across five hospitals, Chelsea and Westminster, West Middlesex University Hospital, and Hammersmith and St Mary’s in London, alongside Bristol Royal Infirmary and Southmead Hospital in Bristol. Professor Ng, a consultant cardiologist at Chelsea and Westminster Hospital NHS Foundation Trust as well as an Imperial academic, is overseeing that rollout. The team estimates the technology is roughly two years away from routine clinical use if the prospective results confirm what the retrospective data showed, a timeline that reflects both the scale of validation still required and the regulatory steps standing between a promising algorithm and a tool doctors can rely on for real patients.
The Case for Caution
Not everyone treats these numbers as a finished product, and the researchers themselves are careful to say so. Dr Sonya Babu-Narayan, the British Heart Foundation’s clinical director, has stressed that the tool cannot be used on its own to definitively diagnose or rule out heart failure or valve disease; a positive flag is meant to trigger a confirmatory echocardiogram, not replace one. That caveat matters because detection rates in the 77% to 90% range, while impressive against a baseline of nothing, still mean a meaningful share of cases get missed or flagged incorrectly. False positives could flood already-stretched echocardiography services with unnecessary referrals, while false negatives risk giving patients and clinicians misplaced reassurance. There are also open questions about how the model performs across different ethnic groups and ECG machine manufacturers outside the US and UK hospitals used in training.
What Comes Next
If the NHS trial data matches the retrospective results, the practical implication is a cheap, near-instant screening layer sitting on top of a test that is already performed billions of times a year worldwide, potentially catching heart failure and valve disease years before symptoms force a diagnosis. Imperial’s team, backed by British Heart Foundation funding, will need to show the tool performs prospectively, not just on historical records, and regulators in both the UK and US will want evidence that frontline clinicians can act on its flags without either overwhelming specialist services or developing blind trust in a black-box score. For a field that has spent a century reading the same wavy line, the next two years will determine whether AI finally teaches cardiology something the human eye was never built to see.
Photo: InspiredImages / PIXABAY via Pixabay
