Heart attack: an AI that helps emergency teams know who to reassure and who to keep

Why is a heart attack hard to recognize in A&E?

A heart attack, or myocardial infarction, happens when part of the heart is starved of blood. Many people arrive in A&E with chest pain, but only some of them are actually having a heart attack.

To sort them, doctors measure troponin, a protein released into the blood when the heart muscle is under strain. Current rules compare this result with a fixed cut-off, the same for everyone. Yet troponin varies with age, sex, other illnesses and the time since symptoms began. As a result, some patients wait for hours or are admitted to hospital when they do not need to be.

What did the researchers do?

The team, based in Scotland, trained machine learning models, meaning programs that learn from examples. They used data from 10,038 patients who came in with a suspected heart attack at ten Scottish hospitals.

The tool, called CoDE-ACS, combines troponin (on arrival or repeated) with simple information such as age, sex, time since symptoms began, blood pressure and kidney function. It gives a score from 0 to 100 that matches the chance of a heart attack.

The researchers then tested it on 10,286 other patients from seven cohorts in six countries. This step, called external validation, is key to checking that a tool works beyond the place where it was built.

What were the results?

CoDE-ACS told apart patients having a heart attack from the others very well: its area under the curve was 0.953, on a scale where 1 would be perfect. It performed well across the different patient subgroups studied.

On arrival, it rated 61% of patients as unlikely to be having a heart attack, compared with 27% using fixed thresholds, with similar safety. It also rated fewer patients as highly likely (10% versus 16%), but with greater accuracy.

People rated as unlikely had fewer heart-related deaths at 30 days (0.1%) than the other groups (0.5% and 1.8%).

What could this change for patients?

If the tool is adopted, many people who are not having a heart attack could be reassured sooner, after a single blood test, and go home instead of waiting for hours or being admitted.

Patients who really are having a heart attack could also be spotted and treated earlier. The score remains a decision aid: the doctor makes the call.

What are the limits?

The authors say so themselves: the tool needs prospective validation, meaning testing in real time in emergency departments, and a measure of its real effect on care.

It was designed around one specific type of highly sensitive troponin test. And it never replaces calling for help: if you have chest pain, call emergency services straight away (999 in the UK, 911 in the US).