Missed fractures are the most common diagnostic error in UK emergency departments. In January 2025, NICE recommended four AI tools to help clinicians catch them. The recommendation is conditional. The tools can be used now, but only under monitoring while the evidence is still being collected.

The dilemma

Between 3% and 10% of fractures are missed on first reading in UK urgent care, making this the most common diagnostic error in emergency departments. AI tools that flag potential fractures on X-rays were arriving fast, with promising early results. But the standard NHS evaluation pathway would have taken years to deliver a definitive verdict. Waiting risked preventable patient harm. Moving too quickly meant deploying tools that had not been properly tested in real-world conditions..

The regulatory gap

A pathway that let clinicians use promising AI tools while the evidence was still being collected. The pathway needed to set out clearly what would count as proof the tools worked, and what would trigger their withdrawal from use

The approach

In January 2025, the National Institute for Health and Care Excellence (NICE) published guidance HTE20, recommending four AI tools for use in NHS urgent care: BoneView, RBfracture, Rayvolve and TechCare Alert. The recommendation came through its Early Value Assessment (EVA) framework. The EVA lets tools be used in monitored conditions while developers collect the evidence needed to prove they work, and NICE then reviews the guidance once that data is in.

Results achieved

Early modelling cited by NICE shows clinicians using these tools detect meaningfully more fractures without raising false positives. With BoneView, clinician sensitivity rises from 61% to 74%; with Rayvolve, from 79% to 94%, at a cost of around £1 per scan All four tools are CE and UKCA marked and must meet NHS England's Digital Technology Assessment Criteria (DTAC).

The evidence is still being collected. NICE itself has acknowledged the modelling likely understates the benefits of catching more fractures and overstates the costs of false positives. This is a structured experiment, not a final verdict.

The transferable lesson

Conditional access only works if there is a real possibility of withdrawal. NICE's guidance makes that possibility concrete: it sets out the evidence developers must collect, the date it will be reviewed, and the conditions under which the recommendation would be withdrawn. That turns "early adoption" from a leap of faith into a testable claim, and is what distinguishes adaptive regulation in practice from adaptive regulation in name only.