An artificial intelligence system developed by Imperial College London researchers can extract signs of heart failure and heart valve disease from routine electrocardiograms in under two seconds. Trained on millions of ECGs and tested on 67,000 US patients, the system could eventually help doctors decide who needs an urgent heart scan; researchers are now testing the technology in NHS patients.
Artificial intelligence can identify signs of major heart conditions from an ordinary electrocardiogram in less than two seconds, according to research presented at the European Society of Cardiology Congress 2026 in Munich.
The system, developed by researchers at Imperial College London, analyses the electrical signals recorded by a standard ECG and searches for patterns associated with heart failure and heart valve disease. These patterns can be too subtle or complex for a clinician to identify from the ECG trace alone.
In testing involving 67,000 patients in the United States, the AI identified up to 81 per cent of patients with heart failure and up to 90 per cent of patients with heart valve disease. The technology has been trained using millions of ECG recordings; reporting from the Congress puts the training dataset at approximately 10.6 million ECGs.
The immediate significance is not simply speed. ECGs are already cheap, commonplace and routinely performed throughout healthcare. Around one billion are estimated to be carried out worldwide each year. If information about structural heart disease can reliably be extracted from tests that are already being performed, patients who need further investigation could potentially be identified much earlier.
What does the AI actually detect?
An electrocardiogram records the electrical activity produced as the heart beats. A conventional 12-lead ECG can reveal heart rhythm abnormalities, conduction problems and characteristic changes associated with conditions including myocardial infarction.
Heart failure and valve disease are different problems.
Heart failure occurs when the heart cannot pump blood around the body effectively enough to meet its needs. It can arise from several underlying abnormalities and does not necessarily produce an obvious pattern that a clinician can recognise on a routine ECG.
Valvular heart disease involves one or more of the heart’s valves becoming narrowed or allowing blood to leak backwards. Confirming these structural abnormalities normally requires imaging, most commonly an echocardiogram.
The Imperial system uses machine learning to analyse the ECG at a level beyond conventional visual interpretation. Rather than relying only on established features such as rhythm, rate or obvious waveform abnormalities, it examines statistical patterns distributed throughout the electrical signal.
Professor Fu Siong Ng’s Imperial research group has described these AI-derived signals as digital biomarkers. An abnormal biological process can subtly alter the electrical behaviour of the heart before those changes become useful to a human reader of the ECG. Large neural networks can be trained to associate combinations of those changes with diagnoses established through other clinical information.
The underlying idea has been developing for several years. Previous research has found that AI-enhanced ECGs can identify or predict left ventricular dysfunction, future heart failure, conduction disease and valvular abnormalities. A multinational study published in the European Heart Journal, for example, found that an AI-ECG model could stratify future heart failure risk across large patient cohorts.
Imperial researchers have also published work specifically examining AI-enhanced ECG prediction of regurgitant valve disease.
What does “under two seconds” mean?
The two-second figure refers to the time required for the AI to analyse an ECG once the recording exists.
A routine ECG itself takes longer to record. Imperial and the British Heart Foundation describe a standard recording as taking approximately ten seconds. The attraction is therefore that the computational analysis can occur almost immediately after a test that healthcare systems already perform routinely.
The system does not somehow observe a patient for two seconds and deliver a complete cardiac diagnosis. It analyses an existing electrical recording and produces an assessment of the likelihood that certain conditions are present.
That distinction matters because an AI-positive ECG would still need to lead somewhere.
For suspected heart failure or valve disease, that next step would typically include clinical assessment and an echocardiogram, an ultrasound examination capable of showing the heart’s chambers, pumping function and valves directly.
The proposed advantage is triage. A patient whose ECG produces a particularly high AI score could potentially be moved towards the front of the queue for an echocardiogram.
“Patients can often wait several months for a heart ultrasound scan after being referred for one by their doctor,” Ng told The Guardian. He said the technology could allow patients at greatest risk to be prioritised for faster investigation.
How good is 81 or 90 per cent?
The headline figures require careful interpretation.
The reported US testing found that the system identified up to 81 per cent of people who had heart failure and up to 90 per cent of those with heart valve disease. Those figures describe the proportion of affected patients successfully identified in the reported analysis.
They should not automatically be interpreted as meaning that every positive result has an 81 or 90 per cent probability of being correct.
A screening test also has to be assessed by how frequently it flags people who do not have the condition, usually expressed through measures including specificity and positive predictive value. Positive predictive value can change substantially depending on how common the disease is in the population being tested.
This becomes especially important if the technology is eventually applied to every ECG taken in a hospital. A model can be highly sensitive while still generating enough false positives to create substantial demand for follow-up scans.
The publicly reported Congress figures therefore establish that the approach can identify a large proportion of affected patients; they do not, by themselves, provide every statistic required to judge how a nationwide screening programme would perform.
That is one reason real-world clinical testing matters.
An NHS trial is already under way
The research is moving beyond retrospective datasets.
The system is being evaluated in NHS patients in London and Bristol. Reports place the current prospective study at roughly 590 to 600 patients. The purpose is to establish how the technology performs when incorporated into ordinary clinical care rather than when analysed retrospectively against an existing dataset.
This distinction is crucial for medical AI.
A model can perform strongly on carefully assembled datasets and still encounter problems when moved into hospitals. Patient populations differ; ECG machines can differ; incomplete data, unusual conditions and changing clinical pathways can affect performance. A useful system also has to provide information at the right point in the clinical workflow without overwhelming staff with alerts.
Bias is another concern. AI systems trained predominantly on one demographic or healthcare system may perform differently in another.
The British Heart Foundation notes that the Imperial programme has deliberately used datasets containing people from different ethnic backgrounds and genders. Real-world validation remains necessary before broad deployment.
Why use an ECG to find structural heart disease?
The attraction comes from scale.
Echocardiography gives doctors far more direct information about cardiac structure than an ECG. It is also more expensive, requires specialist equipment and staff, and cannot realistically be performed on every patient simply as a precaution.
ECGs are already embedded throughout healthcare.
An AI layer can therefore turn an existing test into a preliminary filter. If someone has an ECG because of chest pain, an irregular heartbeat, an operation or another medical issue, the same ten seconds of electrical data could potentially be checked simultaneously for patterns associated with otherwise unsuspected disease.
Ng has suggested this opportunistic approach as another possible application. AI could eventually analyse ECGs performed for unrelated reasons and flag patients at unusually high risk of heart failure or valvular disease for further investigation.
That could be particularly valuable for diseases that may remain unnoticed until symptoms become severe.
The research is part of a much larger AI-ECG programme
The new results are not an isolated demonstration.
Imperial’s National Heart and Lung Institute has spent years developing AI-enhanced ECG models. The research has received long-term support from the British Heart Foundation and has now led to the formation of Cardiovolt.ai, an Imperial spinout intended to commercialise the technology.
Imperial reported in June 2026 that researchers had assembled large international ECG datasets, including more than 1.6 million recordings from Brazil and several million from the United States. The resulting models have been investigated for cardiovascular conditions as well as diseases outside cardiology, including diabetes and kidney disease.
The British Heart Foundation says the wider AIRE, or AI-ECG risk estimation, platform can detect signals associated with more than a dozen conditions.
That does not mean an ECG has suddenly become a universal diagnostic test. The electrical activity of the heart is influenced by anatomy, metabolism, age, disease and the wider condition of the body. Machine learning can identify statistical fingerprints within that information that are difficult to isolate manually.
Whether each fingerprint is sufficiently reliable and clinically useful has to be established independently.
Could this be used on handheld devices?
The researchers are already considering smaller systems.
Dr Ahmed El-Medany, a British Heart Foundation clinical research fellow who led the Imperial analysis presented in Munich, has said one future challenge is the development of handheld AI-enabled ECG readers for healthcare professionals.
Wearable and portable ECG technologies are also becoming increasingly common. The British Heart Foundation has highlighted the potential combination of AI with continuous or repeated ECG measurements from future clinically reliable wearable devices.
That possibility remains further from routine use. Consumer ECGs typically collect less information than a conventional hospital 12-lead ECG, and any system used for medical diagnosis would need appropriate clinical validation and regulatory approval.
What happens next?
The most important next evidence will come from prospective clinical testing.
Researchers need to establish how accurately the technology performs in NHS patients, how many false-positive results it generates, whether doctors can use the output effectively and whether prioritising patients through AI actually leads to earlier diagnoses or better outcomes.
Those questions are more demanding than demonstrating that a neural network can recognise disease in historical data.
The technology nevertheless has a practical advantage that many proposed medical AI systems lack. It is built around a test hospitals already use on an enormous scale.
No new scanner would be required simply to generate the underlying signal. If prospective trials confirm that the models remain reliable in clinical practice, AI analysis could potentially be added to existing ECG infrastructure and used to identify patients who warrant closer examination.
For heart failure and valve disease, where delayed diagnosis can leave patients untreated until disease has progressed, that could make a familiar century-old medical test considerably more informative.
Sources
The Guardian, “‘Superhuman’ AI tool spots heart disease in less than 2 seconds,” 31 August 2026.
British Heart Foundation, “Superhuman AI-powered ECGs move a step closer to clinical use,” June 2026.
British Heart Foundation, “AI (artificial intelligence) in heart healthcare.”
Imperial College London, “Cardiovolt.ai turns heart traces into powerful diagnostic tools,” 8 June 2026.
European Heart Journal, “Artificial intelligence-enhanced electrocardiography to predict regurgitant valvular heart diseases: an international study,” published 16 July 2025.
European Heart Journal, “Heart failure risk stratification using artificial intelligence applied to electrocardiogram images: a multinational study.”
European Society of Cardiology, ESC Congress 2026 scientific programme, Munich, 28 to 31 August 2026.