The AI that helps urgent chest X-rays reach radiologists sooner

Chest X-rays are among the most common tests in a hospital. They are also a queue: an image that can wait and another showing a possible pneumothorax or pleural effusion may enter the same workflow. The challenge is not obtaining an image. It is deciding which one needs human attention first.

That is the use case evaluated at Changi General Hospital (CGH) in Singapore in a study published in 2024. The hospital used an artificial-intelligence tool to sort chest X-rays as normal, non-urgent, or urgent before clinical review. Its purpose was not to issue an autonomous diagnosis. It was to help the radiologist’s worklist reflect clinical priority.

What the hospital put into practice

The evaluation examined Lunit INSIGHT CXR Triage, version 3.1.5.0, in a real hospital setting. The application analyses a chest X-ray and assigns a priority category. The image then remains in the normal care pathway: a clinician interprets the study and signs the report.

The dataset was substantial for a single site: 20,944 chest X-rays acquired at CGH from August through December 2023. The article appeared in the European Journal of Radiology in 2024. Forty-three radiologists, blinded to the AI output, took part in classifying the images into three groups: normal, non-urgent, and urgent. That distinction matters. It makes it possible to compare the system’s ordering with a human reference without turning a machine suggestion into the final diagnosis.

CGH was already using infrastructure for developing and testing imaging AI through the AIM SG platform. In that context, the triage system is not a standalone device. It is part of radiology operations: it receives a newly acquired image, looks for predefined signs of critical findings, and moves the case forward or back in the reading queue.

The improvement: prioritization, not replacement

The results explain why this use can help a high-volume specialty. In the urgent chest-X-ray category, the system reported 82% sensitivity and 99% specificity; its negative predictive value was 98%. In practical terms, a non-urgent result helped rule out that an image belonged to the urgent category, while an alert was highly specific and deserved earlier attention.

For X-rays classified as normal, the study reported 89% sensitivity, 93% specificity, and an area under the curve of 0.91. For non-urgent images, the corresponding figures were 93%, 91%, and 0.92. The authors also found a statistically significant reduction in turnaround times across the subgroups they analysed.

The improvement is not that AI “sees” better than the entire care team. It is about managing reading order when minutes matter. If an image with a possible emergency rises on the worklist, a radiologist can confirm, reject, or contextualize the finding earlier. Images without priority signals do not leave the queue; they simply do not compete for first review with the same urgency.

That design also matches the product’s regulatory framing. The U.S. Food and Drug Administration classifies Lunit INSIGHT CXR Triage as radiological computer-assisted prioritization software. Its 510(k) clearance does not make the program a doctor or authorize clinicians to skip professional interpretation. It describes a supporting role: prioritizing and notifying cases that may need prompt attention.

What the numbers do not establish

A workflow gain should not be turned into a larger clinical promise than the evidence supports. The study took place in one hospital and used images acquired over five months. It included a diverse population and analysed age, sex, and ethnicity subgroups, but other hospitals may have different equipment, disease prevalence, and operational workflows.

An 82% sensitivity in the urgent group also means the system does not identify every urgent case. That is why human reading and safety procedures remain essential. High specificity is not a diagnosis either: a prioritized X-ray still needs to be interpreted alongside the patient’s history and, where appropriate, other tests.

Nor does this study by itself show lower mortality, fewer admissions, or a specific reduction in costs. It measures classification and turnaround time, both valuable but intermediate outcomes. The practical question for a hospital is not, “Can AI replace the radiologist?” It is, “Can it help the radiologist see sooner the study that should not wait?”

A transferable lesson

The CGH case illustrates a mature medical-AI application: it does not promise an automated consultation, but removes friction at one defined point in the process. A professional still reviews the image; AI adds another ordering layer that can make a possible emergency more visible.

For patients and clinical teams, the potential benefit is easy to grasp: a critical test may reach the person making the decision sooner. For those implementing it, the requirement is equally straightforward and more demanding: validate the system in their own population, measure false negatives, and monitor whether automatic prioritization improves care without concealing mistakes.