The ECG, developed more than a century ago, is one of the most common tests in medicine, with around a billion performed worldwide each year. It measures the electrical activity of the heart, and is routinely used to diagnose conditions such as heart attacks and abnormal heart rhythms.
But an ECG often contains far more information than the human eye can typically see - which AI can now detect in less than two seconds. Researchers first trained the technology on ECGs from millions of patients, then tested it on more than 65,000 ECGs from hospital patients in the US. It was able to identify up to 81 per cent of patients who had heart failure – detecting reduced pumping function of the heart's main chamber – and also identified up to 90 per cent of patients with a common type of heart valve disease.
The technology cannot be used on its own to definitively diagnose or rule out heart failure or heart valve disease, but it gives a strong indication that someone may have these conditions from their ECG. Someone judged as high-risk could be sent rapidly for an ultrasound heart scan called an echocardiogram, if this technology became available in the future. That could mean a quicker diagnosis which might allow them to start treatment earlier.
The AI is now being tested on ECGs from 590 NHS patients across London and Bristol, and the team say the AI ECG technology could be two years away from being used routinely for patients.
"Blink of an eye"
Our clinical director, Dr Sonya Babu-Narayan, said:"It is exciting to see that AI can now deliver a read-out from an ECG in what feels like the blink of an eye.
"Technology like the AI ECG in this research, which has the potential to identify high-risk patients early, will not detect everyone with a heart condition. But it could be a solution to help fast-track the patients who are most likely to have a heart abnormality.
"When it comes to the heart, earlier diagnosis and treatment saves and improves lives."
To develop the AI, researchers gave it 10.6 million ECGs, together with the clinical reports describing them.
The model was then trained to learn the connection between the patterns in the ECG results and specific heart conditions, using 72,475 ECGs linked to scan results revealing the structure and function of the heart.
The technology has so far been tested on the ECGs of a small group of 5,442 patients in the US, and the ECGs of a larger group of 61,520 US patients, whose echocardiogram results could be compared with the AI analysis.
Researchers tested whether the AI could identify reduced left ventricular ejection fraction – a measure of how effectively the heart’s main pumping chamber squeezes blood out with each beat – as well as thickening of the heart muscle, and heart valve disease of moderate severity or higher.
The AI performed particularly well at detecting reduced heart pumping function, which is a type of heart failure - a condition affecting more than a million people in the UK. It correctly identified 77 per cent of the small group of 5,442 patients, and 81 per cent of the large group of 61,520 patients who had poor heart pumping function.
Heart valve disease is when one or more of the heart's valves do not work as they should - to control the direction of blood flow. Aortic stenosis is one of the most common types and means the valve does not open fully, which can block or restrict the flow of blood.
The AI was able to identify several forms of heart valve disease, although its performance varied depending on which valve was affected. For aortic stenosis, the AI correctly identified 90 per cent of the small group of more than 5,000 patients, and 80 per cent of the larger group of more than 61,000 patients, who had aortic stenosis.
The results presented at the European Society of Cardiology Congress use a statistical analysis called 'area under the receiver operating characteristic curve'. A score of 1.0 means the AI is always right, and 0.5 means it performs no better than someone guessing at random.
For reduced heart pumping function, the AI achieved a score of 0.86 for the set of 5,442 patients, and 0.9 for the larger group. For aortic stenosis, it achieved a score of 0.85 and 0.73, respectively.
"Hidden information"
Dr Ahmed El-Medany, a British Heart Foundation clinical research fellow, who led the analysis from Imperial College London, said:
“These results suggest there is potentially far more information hidden within a routine ECG than we can recognise by looking at it ourselves.
"It is particularly encouraging that this superhuman AI performed well across separate groups of patients. The important next questions are how it performs in NHS patients and whether the same information can be captured using much simpler portable ECG devices.”
Prof Fu Siong Ng, the senior investigator from Imperial College London, said:
“Patients can often wait several months for a heart ultrasound scan after being referred for one by their doctor. This makes it exciting that our technology could identify patients most at risk of heart failure and heart valve disease, so they could be prioritised for scans faster and more urgently.”
Find out more about our research