An AI reads an ECG in two seconds and finds what cardiologists cannot

An ECG takes a few seconds to record and costs very little. Researchers at Imperial College London have trained an AI system that can read one in under two seconds and identify signs of heart failure and valve disease that clinicians cannot detect from the same trace.

The results were presented at the European Society of Cardiology congress in Munich and reported by the Guardian on Monday. The trial covered 67,000 patients in the United States.

The tool identified up to 81% of heart failure cases and up to 90% of valve disease cases, despite using a test that was not originally designed to detect either condition.

“Superhuman AI” is the phrase Dr Ahmed El-Medany, a British Heart Foundation clinical research fellow at Imperial, used to describe the system.

The claim is relatively specific: the signal is present in the recording, but clinicians cannot reliably extract it from the ECG on their own.

The clinical problem it is designed to address is largely a problem of capacity. “Patients can often wait several months for a heart ultrasound scan after being referred for one by their doctor,” said Professor Fu Siong Ng, professor of cardiology at Imperial.

Heart failure is also a condition where delays can have serious consequences. The disease progresses over time, and treatment is generally more effective when it starts earlier, leaving patients in a queue while their condition continues to change.

An echocardiogram requires a trained sonographer, specialised equipment and an appointment. An ECG, by contrast, can be recorded using ten electrodes and a nurse, which is why it is performed much more frequently.

The idea is therefore to use the AI for triage rather than as a replacement for clinicians. If software can analyse ECGs that are already being taken and identify patients who are more likely to have a structural heart problem, those patients could be prioritised for an ultrasound instead of moving through the waiting list in the same order as everyone else.

“Technology like the AI ECG in this research could be a solution to help fast-track patients most likely to have a heart abnormality,” said Dr Sonya Babu-Narayan, a consultant cardiologist at the British Heart Foundation.

Cost is another reason the approach could work at scale. ECGs are among the cheapest and most widely used medical tests, so adding software to analyse recordings that are already being collected would require relatively little additional infrastructure or expense.

The current trial also sits on a much larger research programme. Ng’s group trained its models on 1.6 million ECGs from Brazil that were linked to patient records, along with several million additional recordings from the United States.

The wider programme looks beyond heart failure and valve disease. The models have also been used to identify heart attacks and arrhythmias, as well as conditions outside cardiology, including diabetes and kidney disease.

The researchers have reported accuracies of 83% to 93% for heart disease and 70% to 80% for the other conditions.

The Brazilian dataset is an important part of that work because the recordings are linked to what subsequently happened to the patients.

That gives the models a way to learn which patterns in an ECG were associated with diagnoses that emerged later, rather than simply learning to recognise conditions that had already been identified when the test was taken.

The research has also moved beyond the university. The BHF-funded work is being commercialised through a spinout called Cardiovolt.ai, with Ng serving as chief medical officer and Dr Arunashis Sau as chief scientific officer.

The team’s next stated step involves hardware as well as software. Handheld ECG devices with the AI built into the workflow could eventually take the technology beyond hospitals and into settings where ECGs are easier to perform, although that would also change the role of the system from helping clinicians prioritise existing patients to potentially identifying people who have not yet been referred for further testing.

Britain has an unresolved regulatory question around this kind of technology. The country has a strong academic pipeline for cardiovascular deeptech, including research that is already producing new approaches to cardiovascular disease, but moving from an academic result presented at a medical congress to routine use in hospitals requires a separate process of clinical validation and regulatory approval.

The current system has not yet gone through that process. An AI tool that identifies patients who should be referred for further investigation is making a clinical claim, which means it would need to meet the relevant medical device requirements before being deployed at scale in the UK.

There is also a difference between what has been presented at a medical congress and what has been demonstrated in routine clinical practice.

The trial suggests that information about heart failure and valve disease can be extracted from an ECG that was not designed to detect either condition, but it does not yet show whether using that information in real clinical settings leads to earlier treatment or better outcomes for patients.