AI triage shifts from detection to worklist prioritisation

The first generation of radiology AI products was sold on detection performance: sensitivity for a nodule, a fracture, or a haemorrhage. Buyers are now more interested in what happens to the worklist. A tool that reorders a reading queue so that suspected intracranial haemorrhage is seen first changes departmental flow in a way that a standalone detector does not, and the effect is measurable in turnaround time.

That shift changes the technical requirements. Worklist integration demands a stable interface to the modality worklist and the reporting system, deterministic behaviour when the algorithm is unavailable, and a clear rule for what happens to a study that the model cannot score. Vendors that treat integration as a project deliverable rather than a product feature struggle in deployment, particularly across sites with different reporting systems.

Radiologists remain sceptical of anything that hides a study. The accepted design keeps every examination visible while marking priority, and logs the reason for the ordering so that a reader can audit it after the fact. Departments that piloted full automation generally rolled it back, usually after a case where a low-priority label delayed a finding that later turned out to be significant.

Evidence expectations are also maturing. Studies that show improved turnaround for a specific pathway are more persuasive to a clinical committee than a standalone area-under-curve figure, particularly where the department can measure the effect on length of stay for the relevant patient group. Vendors are responding by running local evaluations with the hospital's own historical data before a procurement decision is taken.