Faster AI chest-X-ray reports did not shorten diagnosis in a major NHS trial

In one set of UK figures, artificial intelligence is helping people suspected of lung cancer get answers faster. In another, a randomized study of more than 93,000 chest X-rays found that AI-based prioritization did not meaningfully shorten the journey to a cancer diagnosis. It is tempting to declare one right and the other wrong. The more useful investigation is to ask exactly which patients, actions and clocks the two sets of figures describe.[1][2]

On 24 October, HM Treasury said AI tools were being used to help analyse more than 210,000 chest X-rays a month—approximately a third of all NHS chest X-rays, according to its statement. The release attributed to the technology an 11.5% increase in suspected lung cancer cases identified within 24 hours, translating into roughly 400 additional patients each month receiving faster clarity. It also cited a reduction in the average wait for follow-up CT scans from 6.3 to 5.6 days. These are government-reported programme figures, not results of a randomized comparison documented in that release.[1]

The trial changed the queue, with AI in both groups

The LungIMPACT trial, published in Nature Medicine in March, tested a particular intervention across five NHS trusts. On randomized days, chest X-rays requested in primary care could be prioritised for immediate radiologist review using AI findings. On control days, they were not prioritised in this way. Crucially, the AI itself was available in both groups; the randomized difference was the prioritization of the reporting worklist, not the wholesale introduction of AI interpretation versus no AI.[2][3]

After specified exclusions, researchers analysed 93,326 chest X-rays from 86,945 patients. There were 45,987 images in sessions when prioritization was switched on and 47,339 with it switched off. Among these, 13,347 subsequent CT scans were identified and 558 people were diagnosed with lung cancer. The study was large, prospective and randomized, offering stronger causal evidence about the worklist decision than an unadjusted before-and-after dashboard.[2]

Why the two sets of figures are not interchangeable

The Treasury’s 6.3-to-5.6-day CT figure and the trial’s 53-day CT median should not be put side by side as if they describe the same clock. The study included all identified valid CT scans within its defined analysis, many of which happened far beyond an urgent follow-up window. The trial separately reported that among CT scans performed within 14 days of the original X-ray, the median time was eight days in each arm. The Treasury did not provide comparable cohort definitions or the full evaluation design within its press release.[1][2]

A higher number of suspected cases identified within 24 hours is a different endpoint from the time to confirmed cancer diagnosis. The Treasury release does not supply the cohort definitions, comparison design or full source tables needed to align its programme figures with the trial. Its statement attributes the improvement to AI; LungIMPACT isolates the effect of prioritisation more directly.[1][2]

A faster X-ray report may expose a bottleneck in CT capacity, urgent referral slots, clinical communication or follow-up. If the downstream service is constrained, sorting a report queue earlier does not automatically add imaging appointments or specialist staff. That is a reason to evaluate clinical workflows end to end, not to dismiss the reporting improvement. An earlier alert might still matter for certain patients even when the population-wide median does not move, but such benefits require appropriately designed evidence.[2][3]

The trial’s practical result is specific: moving suspicious X-rays up a reporting queue was insufficient to speed the downstream pathway in this setting. Its faster reports still reached a system of CT appointments, referrals and clinical reviews. Any claim that medical AI saves patients time needs to say where that time was saved.[2][3]

Report an error or suggest a correction