AI ECG for heart disease detection: Imperial College's two-second triage tool

Imperial College London researchers have trained an AI system that reads a standard ECG in under two seconds and identifies signs of heart failure and valve disease that clinicians cannot reliably extract from the same trace. In a trial of 67,000 US patients, the model detected up to 81% of heart failure cases and up to 90% of valve disease cases TNW. The tool is designed as a triage mechanism to prioritize patients for echocardiography, not as a replacement for clinicians.

The AI ECG model and its detection performance

The system was trained on 1.6 million ECGs from Brazil that were linked to patient outcomes, plus several million additional recordings from the United States TNW. This outcome-linked data allows the model to learn which ECG patterns are associated with diagnoses that emerge later, rather than simply recognizing conditions already identified at the time of the test. The researchers reported accuracies of 83% to 93% for heart disease and 70% to 80% for other conditions including diabetes and kidney disease.

The clinical problem is capacity. Patients can wait months for an echocardiogram, which requires a trained sonographer and specialized equipment. An ECG, by contrast, is cheap, quick, and widely performed. Adding AI analysis to existing ECG workflows could fast-track patients most likely to have a structural heart abnormality Guardian.

Why the data strategy matters for builders

For AI builders, the key takeaway is the training data design. The Brazilian dataset links ECG recordings to longitudinal patient outcomes, enabling the model to learn predictive patterns rather than just diagnostic labels. This is a common challenge in clinical AI: most medical datasets are labeled at the time of the test, not with future outcomes. The Imperial team's approach of using outcome-linked data from a large population (1.6 million) is a practical solution that could be replicated for other screening tasks.

The model also generalizes across populations: trained on Brazilian data, validated on US recordings. That cross-population performance is encouraging but needs further testing in diverse clinical settings.

How this changes the triage workflow

If validated and approved, the tool would sit on top of existing ECG collection. A nurse records an ECG, the AI analyzes it in under two seconds, and if the suspicion score is high, the patient is moved up the waiting list for an echocardiogram. This could reduce delays for heart failure and valve disease, where early treatment matters.

The research is being commercialized through a spinout called Cardiovolt.ai, with Prof. Fu Siong Ng as chief medical officer and Dr. Arunashis Sau as chief scientific officer TNW. The team also plans to integrate the AI into handheld ECG devices, which would extend triage beyond hospital settings.

What's still missing: regulatory and clinical validation

The results were presented at a medical congress, not published in a peer-reviewed journal or demonstrated in routine clinical practice. The tool has not yet gone through UK regulatory approval for medical devices, which is required before it can be deployed at scale TNW. The Guardian quotes Dr. Sonya Babu-Narayan noting that the tool "will not detect everyone with a heart condition" Guardian.

The trial shows that information about heart failure and valve disease can be extracted from an ECG, but it does not yet prove that using this information in real clinical settings leads to earlier treatment or better outcomes. Builders evaluating similar clinical AI projects should watch for the regulatory pathway and the evidence of downstream impact, not just model accuracy.