Chest X-rays: an AI that puts urgent cases at the front of the queue

Why do some chest X-rays wait for days?

In a hospital, hundreds of chest X-rays arrive every day. Radiologists read them in a certain order, which does not always reflect what they show. An X-ray revealing a serious problem can therefore wait behind normal ones.

For a patient, this wait matters: the sooner the image is read, the sooner care can begin.

This is called triage: placing each exam in the right queue according to its urgency, as in a hospital emergency department.

What did the team build?

They developed a system based on convolutional neural networks, programs that learn to recognise images. It works in real time: it sorts each adult X-ray by the urgency of what it shows.

To train it, they had 470,388 fully anonymised X-rays taken from 2007 to 2017. They first had a language-processing program read the reports written by radiologists, which put each X-ray into a category: critical, urgent, non-urgent or normal. The image system then learned to find that category by looking at the image alone.

What did the simulation show?

The researchers tested the system on an independent set of 15,887 X-rays. It detected normal X-rays with a sensitivity of 71% and a specificity of 95%.

Above all, they simulated what would happen if exams were read in the order chosen by the AI and compared it with past data. The average reporting delay fell from 11.2 to 2.7 days for critical findings, and from 7.6 to 4.1 days for urgent findings.

What could this change for patients?

If the system were used in a real department, X-rays showing a serious problem would be read sooner, without changing the number of radiologists. The authors conclude that automated real-time triage is feasible, with clinically acceptable performance.

The AI does not make the diagnosis: it changes the reading order, and the radiologist reads the exam.

In practice, this means an exam revealing a serious problem could move ahead of normal X-rays. The authors speak of saving time for the cases that need it most.

What are the limits?

The results come from a simulation using past exams, in a single institution. The delay gains reported are those of the simulation, not those measured in a department that had adopted the system.

The urgency categories were derived from reports written by humans. It will need checking in real conditions that the triage does not, by mistake, delay an important exam.

The 95% specificity for normal X-rays is good, but the 71% sensitivity means about three in ten normal X-rays are not recognised as normal: the tool is not perfect.