Long-term ECGs: an AI that sorts heart rhythm problems

Why is reading a 24-hour ECG hard?

An electrocardiogram (ECG) records the heart's electrical activity. When it is worn for long periods with a small device, called ambulatory, it produces a huge amount of data to review, and automatic reading software sometimes gets it wrong.

There are many rhythm problems (arrhythmias) to tell apart. Digital ECGs are now widely available, which opens the way for programs able to learn how to analyse them.

The authors recall that computerised ECG interpretation already plays an important role in clinicians' work, but that its accuracy and capacity to handle large volumes can still improve.

What did the researchers do?

They built a deep neural network, a program that learns straight from the signals, with no hand-written rules. It had to sort 12 rhythm classes using 91,232 single-lead ECGs from 53,549 patients who wore an ambulatory recording device.

To test it, they used an independent dataset whose diagnoses had been set by a committee of board-certified cardiologists who reached agreement among themselves.

In plain terms, each recording is a long line of heartbeats. The program looks at the whole line and decides, for each stretch, which of the 12 rhythm types it is and it was never given a hand-made rule for any of them.

What do the results show?

On the test, the neural network's area under the ROC curve (a score where 1 is perfect) averaged 0.97. The F1 score, which combines the ability to catch cases with the reliability of alerts, was 0.837 for the program, against 0.780 for the average cardiologist.

When the program's specificity is fixed at the average cardiologist level, its sensitivity beats the cardiologists for every rhythm class. In other words, at the same false-alarm rate, it catches more.

This comparison is rigorous: the reference was a committee of practising cardiologists who agree among themselves, and the test was independent of the training data.

What could this change for patients?

If these results are confirmed in clinics, the authors think the approach could cut errors in automated ECG readings and make expert review more efficient, by sorting or prioritising the most urgent problems.

For the patient, the hope is faster review of long recordings, with the at-risk cases looked at first.

Another benefit: the program can sort a broad range of different rhythms, not just one or two well-known problems, which addresses a limit of earlier work.

What are the limits?

The authors themselves state that this performance still needs confirming in real clinical settings. The recordings came from one type of device, with a single lead.

The program is a reading aid: the cardiologist makes the diagnosis and decides on care.