The electrocardiogram is one of medicine’s oldest, cheapest and most frequently performed diagnostic tests. A standard recording takes around 10 seconds and gives clinicians information about the electrical activity of the heart, allowing them to identify problems including arrhythmias and evidence associated with myocardial infarction. Artificial intelligence is now suggesting that those familiar lines contain far more biological information than physicians have traditionally been able to extract.
Researchers at Imperial College London have developed AI models capable of identifying subtle patterns associated with heart failure, valve disease and other conditions from routine ECGs, and the technology is being commercialized through spinout Cardiovolt.ai. British Heart Foundation-backed development used more than 1.6 million ECGs linked with medical records in Brazil alongside several million additional U.S. ECGs. The foundation reported testing performance of 83% to 93% for heart-disease tasks and 70% to 80% for some non-cardiovascular conditions such as diabetes and kidney disease.
Fresh results presented around European Society of Cardiology Congress 2026 have pushed the concept further into the public spotlight. Recent reporting described an AI system trained using more than 10 million ECGs that can process a trace in less than two seconds and, in a large U.S. evaluation, identified heart failure with accuracy of up to 81% and valve disease with accuracy approaching 90%. The technology is being evaluated in an NHS study involving patients in London and Bristol, but it is not yet a replacement for confirmatory tests such as echocardiography or specialist assessment.
How can AI see heart disease in an ECG that a cardiologist cannot?
A human cardiologist reads an ECG by recognizing patterns that generations of clinical research have associated with specific electrical abnormalities.
Artificial intelligence does not need to limit itself to those known patterns.
A deep-learning system can analyze millions of numerical relationships across waveform shape, timing, amplitude and interactions between the 12 ECG leads. Some of those relationships may correlate with structural or physiological disease even though they do not produce a visually recognizable textbook pattern.
The British Heart Foundation describes the underlying signals as markers of disease processes embedded within electrical readings and effectively invisible to the human eye.
That is why researchers sometimes use the provocative term “superhuman.” They do not mean the AI understands cardiology better than every physician in every situation. They mean the algorithm is being designed to perform tasks that human visual ECG interpretation was never capable of performing consistently.
How could an ECG reveal heart failure if it does not directly image the heart?
Heart failure changes the cardiovascular system in ways that can influence electrical activity.
Changes in chamber size, pressure, muscle structure, conduction and other characteristics can subtly alter an ECG even when no single feature is specific enough for a clinician to diagnose heart failure from the trace alone.
AI can combine numerous weak signals.
The same principle can apply to valve disease. Aortic stenosis, for example, increases pressure load on the left ventricle and can cause structural remodeling that leaves electrical signatures in the ECG.
Today, suspected structural disease is generally confirmed using echocardiography, which directly images cardiac chambers, valves and function.
AI-ECG does not eliminate that need. Its most immediate value may be deciding which apparently routine patient should move to the front of the echo queue.
Why could AI-ECG matter so much for healthcare systems?
Because ECGs are already everywhere.
Hospitals do not need to install a multimillion-dollar new scanner before the algorithm can potentially produce value. Millions of ECGs are already generated by emergency departments, outpatient clinics, general practitioners and cardiology services.
If validated and regulated, AI could analyze those existing recordings automatically.
That creates the possibility of opportunistic screening. A patient receives an ECG for palpitations, preoperative assessment or another reason, while the software simultaneously checks for hidden signals associated with ventricular dysfunction or valve disease.
The cost of collecting the raw physiological data has already been paid.
This makes AI-ECG different from screening technologies that require an entirely new test for every person.
Could this reduce long waits for echocardiograms?
That is one of the most attractive near-term use cases.
Echocardiography is non-invasive and highly useful, but trained staff and equipment capacity limit throughput. If every mildly suspicious patient is sent directly for imaging, waiting lists expand.
An AI-enhanced ECG could function as a triage layer.
Patients with strong hidden signals for heart failure or valve disease could receive priority echocardiography, while those with reassuring results might follow a different clinical pathway depending on their symptoms and overall risk.
British media reports around the latest research highlighted an NHS diagnostic backlog involving hundreds of thousands of people awaiting heart tests, which is why even modest improvement in prioritization could have system-level value.
The challenge is avoiding false reassurance. A negative AI result cannot automatically overrule symptoms or clinical judgment.
Could the same ECG detect diabetes or kidney disease?
Potentially, but this is much earlier and requires careful interpretation.
British Heart Foundation-backed research reported AI performance between 70% and 80% for selected non-cardiovascular conditions including diabetes and kidney disease.
At first glance, this seems surprising. Why should a heart’s electrical trace contain information about the kidneys or glucose metabolism?
The explanation is that systemic diseases alter physiology throughout the body. Diabetes can influence autonomic function, cardiovascular structure and vascular health. Kidney disease can affect electrolytes, fluid balance and cardiac remodeling.
AI may identify combinations of those downstream effects even when humans cannot associate them with one obvious waveform feature.
That does not mean an ECG will replace blood glucose, HbA1c, creatinine or other established laboratory testing. Instead, it could eventually act as a flag indicating that another disease deserves investigation.
Could every ECG eventually produce dozens of AI predictions?
Technically, yes. Clinically, that could become a problem.
Once researchers prove that one ECG contains latent information about multiple diseases, developers can train models for left-ventricular dysfunction, valve disease, atrial fibrillation risk, pulmonary hypertension, mortality risk and potentially numerous metabolic conditions.
The raw data do not need to be recollected for every algorithm.
But a physician receiving 27 risk scores from one routine ECG may be less informed rather than more informed.
Future platforms will therefore need to decide which predictions matter for which patient and which outputs should actually be shown.
The best medical AI may not be the system that finds the greatest number of statistical associations. It may be the one that identifies the small number of hidden signals that reliably lead to an action capable of improving the patient’s outcome.
What are the risks of letting AI reinterpret routine medical tests?
False positives could trigger unnecessary echocardiograms, specialist referrals and anxiety.
False negatives could create misplaced confidence that structural heart disease is absent.
Dataset bias is another concern. An algorithm trained primarily in one country, healthcare system or patient population can perform differently when deployed elsewhere.
That is why the scale and geographical diversity of Cardiovolt.ai’s development datasets are important but not sufficient. Prospective validation inside the healthcare systems where the software will actually operate remains essential.
The company itself states that Cardiovolt.ai is progressing through its regulatory journey and is not currently ready for routine clinical use.
Could AI eventually read ECGs from handheld or wearable devices?
This could dramatically expand the opportunity.
Twelve-lead ECG machines are common in healthcare facilities but are not continuously attached to most people. Portable single-lead and multi-lead devices are increasingly accessible through handheld hardware, patches and consumer wearables.
Researchers will need separate evidence proving that models trained on clinical 12-lead recordings can perform reliably on those noisier and less comprehensive signals.
If they can, however, latent-disease screening could migrate beyond hospitals.
A future handheld ECG might not merely tell a patient whether the heart rhythm looks irregular. It could flag signs suggesting declining ventricular function or other cardiovascular disease and recommend formal evaluation.
That would dramatically expand screening reach while simultaneously increasing the challenge of preventing inappropriate self-diagnosis.
Why is the humble ECG becoming an AI gold mine?
Because it combines three characteristics medical AI developers love: enormous historical datasets, standardized digital signals and very low acquisition cost.
Hospitals have accumulated ECG archives covering millions of patients over decades. Those records can be linked with later diagnoses, imaging, hospitalization and mortality data.
The algorithm can therefore ask retrospective questions that would have been extraordinarily expensive to study prospectively: did patients who developed heart failure three years later already have a subtle electrical signature in their earlier ECG?
When enough linked data exist, patterns can emerge.
The same phenomenon is transforming radiology, pathology and retinal imaging. Artificial intelligence is converting old diagnostic tests into data sources for questions those tests were never originally intended to answer.
Could the 10-second ECG become a general health screening platform?
That is possible, but the path from research performance to routine medicine will determine whether the idea becomes transformative or merely impressive.
Successful deployment requires more than high accuracy in a retrospective dataset. Developers must prove that algorithms work prospectively, remain reliable across demographic groups and equipment manufacturers, integrate into clinical workflow and lead to better decisions.
Regulators must define appropriate indications, and healthcare systems need clear protocols for what happens after the AI generates a positive signal.
The future could therefore be less dramatic than “AI diagnoses everything from an ECG” and more useful than that slogan suggests.
A patient has a routine 10-second heart tracing. In the background, software identifies a pattern associated with ventricular dysfunction that no clinician could reasonably see. That patient receives an echocardiogram several months earlier than they otherwise would have and begins treatment before overt heart failure develops.
If that chain can be demonstrated reliably, one of medicine’s oldest tests will have acquired a completely new purpose without changing the machine that records it.
The biggest innovation would not be making an ECG faster. It already takes seconds. It would be discovering that medicine has been collecting far more information in those ten seconds than anyone realized.
