Thirteen hours sooner to the report, not one day sooner to the diagnosis

Case in one sentence

A randomized controlled trial across five NHS trusts, covering 93,326 chest X-rays from 86,945 patients, tested whether letting an AI push suspicious radiographs to the front of the reporting queue shortened the lung cancer diagnostic pathway, and found that it made the report arrive about thirteen hours sooner and the diagnosis arrive no sooner at all.

Executive summary

England’s National Optimal Lung Cancer Pathway says that a patient with a suspicious chest X-ray should have a CT scan within 72 hours, preferably the same day.¹ The trial in this case measured a median of 53 days from X-ray to CT, in both of its arms.¹ Holding that number in mind is the only way the result makes sense.

The intervention was worklist prioritization. Every chest X-ray requested by primary care at five NHS trusts was scored by qXR, a CE-certified commercial algorithm from Qure.ai.¹ On randomly allocated days, images the AI thought abnormal were pushed to the top of the reporting queue with an active notification. On the other days they were not. The AI ran on every image in both arms, and reporters could see its markings either way.¹ So the trial did not ask whether AI helps. It asked whether prioritizing the queue helps, which is a narrower and more answerable question.

Prioritization worked, in the sense that it did what it was designed to do. The median time from X-ray to report fell significantly, from 47 hours to 34.1 hours.¹

Nothing downstream moved:

• Time to CT: 53 days with prioritization, 53 days without. Ratio of geometric means 0.97, 95% CI 0.93–1.02, P = 0.31.¹
• Time to lung cancer diagnosis: 44 days against 46. Ratio 0.98, 95% CI 0.83–1.16, P = 0.84.¹
• Urgent referral, treatment start, and stage at diagnosis: no differences.¹
• No significant difference at any of the five sites, or in any quarter of the trial.¹

The trial was adequately powered for this. It needed 265 cancers per arm and got 269 and 289, out of 558 lung cancers in 86,945 patients.¹

The authors’ conclusion is unusually direct for a null result:

AI prioritization of CXR requested by UK primary care has no significant impact on the lung cancer pathway. Therefore, CXR AI deployments should not include worklist prioritization in this context.¹

What makes the finding interpretable rather than merely disappointing is that an overlapping team had already run the adjacent experiment. In radioX, a randomized trial at a single London hospital, having a radiographer report the X-ray immediately, with direct triage to same-day CT, cut median time to lung cancer diagnosis from 63 days to 32.² Same pathway, same outcome measure, same country. The human intervention halved it. The algorithmic reordering of a queue did not touch it.

Research question

When an AI is inserted into a clinical pathway and the step it accelerates is not the step that is slow, what happens?

Organization and operating context

Five NHS trusts in England took part: University Hospitals of Leicester performed 54% of the study’s X-rays, Nottingham University Hospitals 15%, a trust the paper names as “East Sussex and North Essex Foundation Trust” 11%, University College London Hospital 11%, and University Hospitals of Birmingham 9%.¹ Nottingham University Hospitals NHS Trust was the sponsor.³ The population was ordinary: mean age 59, 46% male, everyone attending for a chest X-ray requested by their GP.¹

The constraint the trial ran into is systemic and the paper states it plainly. As of April 2025, median reporting time for a primary-care chest X-ray was 2 days and the waiting time for a chest CT was 15 days, against a pathway standard of 72 hours from abnormal X-ray to CT.¹ The authors attribute the gap to fewer CT scanners per capita than comparable countries and a chronic shortage of radiologists and radiographers.¹

Two numbers from the trial show how far practice sits from the standard. Of 13,347 CT scans, 172 happened on the same day as the X-ray — the NOLCP’s preferred option — and 477 happened within 72 hours, its maximum.¹ The pathway the AI was inserted into was missing its own target by a factor of roughly seventeen on the median.

The product was qXR v4.0, a class IIb CE-certified deep learning tool detecting 29 classes of abnormality, already in routine use at some NHS hospitals.¹ Images went to the vendor’s cloud and returned to the local PACS with markings. The algorithm was frozen for the trial’s duration.¹

The situation before AI

Before this trial, the evidence base for chest X-ray AI in England was thin enough that NICE declined to recommend any product, concluding there was insufficient evidence to say anything beyond that products should be carefully evaluated — and naming LungIMPACT as the only study it considered potentially capable of answering the clinical question.¹

What did exist was radioX: the same outcome, in the same national pathway, changed decisively by a change in who reads the image and when. Immediate radiographer reporting cut time to diagnosis from 63 to 32 days, P = 0.03, in 8,682 X-rays yielding 49 cancers.² That trial also found radiographer reports as accurate as local radiologist reports, with blinded thoracic radiologists equally likely to side with either in discordant cases.²

So the prior was favourable, and the trial’s own power calculation was built on it: a conservative 10-day reduction against radioX’s 63-day baseline.¹ The team expected the AI to do something.

The AI intervention or event

21 March 2023. The trial is registered as ISRCTN78987039, sponsored by Nottingham University Hospitals NHS Trust, with two primary outcomes: time from chest X-ray to lung cancer diagnosis, and time from chest X-ray to CT.³ Ethics approval 23/EE/0014 had been granted a month earlier.³¹

17 July 2023 to 31 December 2024. 97,731 chest X-rays are performed across the five trusts. Monday-to-Friday whole-day sessions are block-randomized 1:1 by day and site, so that staffing and capacity variation is spread across both arms.¹ On prioritization days, any qXR-suspected-abnormal case raises an active notification on the reporting worklist. On other days it does not. In both arms, radiographers may flag abnormalities at their discretion and may arrange same-day CT under local procedure — a deliberate choice, so that the trial tested prioritization rather than a pathway change.¹

Data cleaning. 93,326 X-rays from 86,945 patients enter the analysis; 45,987 of them (49.3%) on prioritization days.¹

Follow-up to 5 June 2025. 13,347 patients have a valid CT; 558 are diagnosed with lung cancer.¹

Discordance review, both arms. Every disagreement between the AI and the radiology report is reviewed: 26,505 of 28,261 discordant reports, 94% of discordances and 28.4% of randomized X-rays, by local thoracic radiologists or reporting radiographers.¹ The paper notes this is peer review at a scale seldom, if ever, attempted.

24 March 2026. The result is published in Nature Medicine.¹

Outcomes and economics

The economics are registered and unpublished, which is itself worth recording. Cost-effectiveness is secondary outcome 7 in the trial registration, to be measured as difference in cost per patient diagnosed, per percentage increase in early-stage diagnosis, and potentially per quality-adjusted life year.³ The paper states that a detailed health economics evaluation will be published separately, and characterises the costs of the prioritization function as “both considerable and avoidable, given the results reported in this study” — without giving a figure.¹

For scale, the paper reports that a Scottish before-and-after service evaluation of a different chest X-ray AI increased cost by £3.59 per X-ray.¹ That figure is LungIMPACT’s characterisation of a study not read for this case.

So the best-designed evaluation in this library measured its benefit to a confidence interval and has not yet published its cost. The benefit half arrived first, because this time somebody ran the experiment.

Causal assessment and competing explanations

This case is labelled causal, and it is one of the few in this library that earns the label. Sessions were randomized, allocation was by day and site to balance staffing, the analysis was performed by a statistician independent of the investigators, the outcomes were registered before recruitment and reported as registered, and the trial was powered for the effect it sought.¹³ A null result from that design is evidence of absence, not absence of evidence.

The interesting causal question is not whether prioritization worked but why it did not, and the paper is careful to separate the candidates.

“The AI is not accurate enough.” The authors reject this, and their reasoning is strong: the vast majority of cancers occurred where the AI was abnormal, so prioritization was tested on a large number of true cases, and “even modest differences in accuracy are unlikely to change the primary outcome.”¹ A better algorithm would reorder the same queue.

“The step it accelerated was not the bottleneck.” This is the explanation the evidence supports. Report turnaround improved by about thirteen hours; the pathway’s median is 53 days to CT. The authors put it as a capacity problem: trusts lack the ability to both report immediately and organise downstream tests and appointments, even though the pathway recommends it.¹

“The trial should have mandated a pathway change in the intervention arm.” The authors consider this and decline it, on the grounds that mandating immediate radiologist review before the patient leaves is a pathway change with an effect of its own and “has little to do with AI-based prioritization.”¹ That is the right call for isolating the variable, and it is also the finding: the AI’s contribution is separable from the pathway change, and it is the pathway change that carries the effect.

The comparison with radioX makes this concrete. Immediate human reporting with direct triage to same-day CT halved time to diagnosis.² Immediate algorithmic prioritization, without the triage pathway, did nothing. The active ingredient in “immediate reporting” appears to be the immediacy of the downstream action, not the ordering of the queue.

One post hoc finding cuts the other way and must be reported. Among patients later diagnosed with cancer, those whose X-ray was called normal by both the radiologist and the AI took a median of 177 days to diagnosis, against 38 days when both called it abnormal — seven times longer to CT and three times longer to diagnosis, both P < 0.001.¹ More pointedly, 53 cancers sat in the group where the radiologist said normal and the AI said abnormal, with a median 106 days to diagnosis, while only 22 sat in the reverse group.¹ That asymmetry hints that the AI’s detection may catch something human reporting misses, even though its prioritization changes nothing. It is post hoc, it is not what the trial randomized, and the authors call for detailed evaluation of those cases rather than claiming the finding. This case does the same.

Failures, limitations, and governance

• The bottleneck was never the reading queue. A 53-day median to CT against a 72-hour standard is a capacity problem, and no reordering of a worklist addresses it.
• The trial tested prioritization, not AI. The AI ran in both arms, so this case says nothing about whether having qXR present is better than not having it. The authors state this limitation and explain why a non-consenting design required it.¹
• One product, one version. qXR v4.0, frozen for the trial. The authors note that other algorithms will have different characteristics and that this matters less for prioritization than it would for accuracy.¹
• False positives are a real load. Across 254,349 discordant findings reviews, 11.6% were false positives against expert review, and the paper raises “vigilance fatigue” and the cry-wolf effect as concerns, noting this is part of why NICE has recommended no CXR AI product.¹ Reporter feedback was that most false positives are dismissed quickly, which is a self-reported mitigation rather than a measured one.
• False negatives are few but not harmless. In the nodule category, 343 AI false negatives contained 20 subsequent lung cancers, 5% of the cancers in that category.¹
• The cost half is unpublished. Registered as secondary outcome 7, deferred to a separate paper.³¹
• The registry and the paper disagree on scope. The registration lists seven participating trusts; the paper reports five, with United Lincolnshire and Mid Yorkshire absent.³¹ Registered target enrolment was 150,000; 97,731 X-rays were performed, and the paper’s own power calculation judged 100,000 sufficient without mentioning the registered target.³¹ Registered recruitment start was 1 August 2023; the paper reports 17 July 2023.³¹
• A site name appears to be wrong in the paper. It reports “East Sussex and North Essex Foundation Trust”; the registry lists East Suffolk and North Essex NHS Foundation Trust.¹³ Recorded, not resolved.
• The AI’s code is unavailable. The paper says so and gives the reason — it is a commercial product — which is a disclosure, not a remedy.¹

What this case demonstrates

1. Accelerating a step does not accelerate a pathway. The measured gain was thirteen hours inside a process whose median is 53 days.
2. A feature can work exactly as designed and deliver nothing. Prioritization prioritized; the queue moved; the patient did not.
3. Randomization can settle a question that observational deployment reports cannot, and the two can point in opposite directions about the same technology in the same health service.
4. The active ingredient may be organisational rather than algorithmic. The intervention that halved time to diagnosis in the same pathway was a person reading the image immediately and a CT scanner being available.
5. Pre-registration is checkable, and checking it is cheap. The two primary outcomes reported here are the two registered, which a reader can verify in a minute against an independently hosted record.
6. A vendor can be kept out of the evaluation of its own product. Public funding, an independent statistician, and a published statement that the vendor had no influence on design or analysis.
7. A well-powered null is a result, and this one is stated as an instruction: do not deploy this function in this context.
8. Detection and prioritization are different claims about the same product, and a trial of one is not a trial of the other.

What this case does not demonstrate

• It does not show that chest X-ray AI is useless. Only prioritization was randomized; the AI was present in both arms, and the post hoc discordance findings suggest its detection may have value the trial did not test.
• It does not show that the South West London deployment’s reported improvement was wrong. That deployment changed its pathway at the same time, which is a different intervention from the one randomized here.
• It does not establish cost. The economics were registered and have not been published.
• It does not generalise to every algorithm. One product, one version, one country, one pathway.
• It does not establish that the 53-day median is typical of the NHS in general, only of these five trusts over this period.
• It does not resolve whether AI-flagged, radiologist-dismissed findings should be acted on. 53 cancers sat in that group; the authors call for evaluation rather than a rule.
• It does not tell you what prioritization costs to run, which is precisely what an operator deciding against it would want.

Evidence assessment

Grade A — a pre-registered, publicly funded, vendor-excluded randomized trial, read in full, with its registration checked against it.

This is the best-evidenced case in the library, and it is worth being explicit about why, because the library’s grade A cases have until now come from a different kind of source entirely — regulators, courts, prosecutors and audited filings, bodies that convene when something has gone wrong. This one comes from a trial that would have run whatever it found.

The design does the work. Randomization by session removes the confounding that defeats before-and-after service evaluations. The AI in both arms isolates the variable. The outcomes were registered in March 2023 and the two reported are the two registered — I checked, through the registry’s own API, and there is no substitution, no reordering, no quiet promotion of a secondary outcome that happened to reach significance.³ The analysis was done by a statistician independent of the investigators, the funder is a public innovation programme, and the paper states that the vendor had no influence on design or analysis.¹ The trial is powered for the effect it sought and reports its confidence intervals.

The accountability is correspondingly good: a named ethics committee and reference, a named sponsor with a published route to anonymised data, CONSORT adherence, published statistical code, and an honest statement that the algorithm’s code is proprietary and unavailable.

Three things keep this at A rather than beyond it, none of them a defect in the trial.

It is not replicated. No second randomized trial of chest X-ray AI prioritization exists. A single trial, however good, is one trial. The radioX comparison is not corroboration — same first author, different intervention — and is used here only as the contrast that makes the null interpretable.²

It tests prioritization, not AI. A reader who wants to know whether qXR should be installed at all will not find the answer, and the authors say so.

Two of its five sources here are abstracts. The South West London studies were read only as abstracts, because RSNA and NEJM AI both refuse this sandbox. Every claim drawn from them is labelled, and the one that matters most — the fall from 6.0 to 3.6 days — is also reported inside LungIMPACT’s discussion, which I did read in full, so it does not rest on an abstract alone.¹⁴

That last point deserves its own note, because it is where this case could have gone wrong. The tempting story is that a rigorous trial debunked an enthusiastic deployment report. It did not, and saying so would be a misreading of both. The South West London trust deployed AI triage and built a radiographer-led same-day CT pathway around it; LungIMPACT randomized the prioritization function alone and deliberately left the pathway untouched. These are different interventions. LungIMPACT’s authors make precisely this criticism of that study by name, adding that it used suspected rather than confirmed lung cancer as its reference standard.¹ The honest synthesis is not that one study is wrong, but that the two together point at the pathway rather than the algorithm — which is what the trial’s own closing recommendation says.

A note on the outcome label, which is an editor’s call. This case is filed negative rather than inconclusive or mixed. Inconclusive would be wrong: the trial was powered, the intervals are tight, and the question was settled. Mixed is arguable, because report turnaround did improve significantly and the post hoc detection signal is real. It is filed negative because the deployment under test — worklist prioritization — delivered none of the clinical benefit it was installed for, added cost the authors call avoidable, and ends with its own investigators recommending against it.

Material claims

Direct quotations

AI prioritization of CXR requested by UK primary care has no significant impact on the lung cancer pathway. Therefore, CXR AI deployments should not include worklist prioritization in this context.

— Woznitza and colleagues, stating the trial’s conclusion as an instruction¹ · locator: Abstract, final sentences

Future research should differentiate between primary pathway changes and the direct impact of AI.

— Woznitza and colleagues, on the confound that defeats before-and-after evaluations of the same technology¹ · locator: Abstract, closing sentence

A significant reduction in the median time from CXR acquisition to report was observed, from 47 h to 34.1 h; however, no significant differences were detected in any of the timings measured as primary or secondary outcomes

— Woznitza and colleagues, the proxy moving and the outcome not¹ · locator: Discussion, paragraph beginning “A significant reduction”

It has been shown conclusively that AI prioritization had no impact on time to CT or time to lung cancer diagnosis.

— Woznitza and colleagues¹ · locator: Discussion, paragraph beginning “The LungIMPACT study has examined”

do not meet the NOLCP standards (72 h from abnormal CXR to CT)

— Woznitza and colleagues, on national reporting and CT waiting times as of April 2025¹ · locator: Discussion, second paragraph

there was extensive pathway redesign within the ‘after’ pathway, which the authors recognized as a significant limitation in attributing improvements to AI

— Woznitza and colleagues, on the before-and-after study that reported a fall from 6 to 3.6 days¹ · locator: Discussion, paragraph reviewing comparable studies

Qure.ai Technologies had no influence on the study design or analysis.

— Woznitza and colleagues, in the acknowledgements¹ · locator: Acknowledgements

Immediate reporting of CXRs from primary care reduces time to diagnosis of lung cancer by half, likely due to rapid progress to CT.

— Woznitza and colleagues, the conclusion of the earlier randomized trial of the human intervention² · locator: radioX abstract, Conclusions

Revision notes

• 2026-09-15 — Initial publication at Grade A, the library’s fifth and the first whose evidence comes from a trial rather than from a regulator, a court, a prosecutor or an audited filing. Found while searching for a shadow-mode case; recorded in the previous session’s handoff as the stronger story left on the table, and written next because the access problem turned out to be one-sided in the right direction: the randomized trial is fully readable at nature.com while the deployment reports it critiques are not. Sources span five families. The trial itself was read end to end; its pre-registration was retrieved through the ISRCTN XML API — the registry’s own web page asks users not to scrape it and directs them there — and compared against the paper, confirming that the two primary outcomes reported are the two registered, and surfacing three discrepancies of scope that are recorded rather than resolved. The comparator trial, radioX, was read as an abstract carrying its full design and result. The two South West London deployment reports were read as abstracts only, because RSNA and NEJM AI both refuse this sandbox, and every claim taken from them is labelled Attributed accordingly. The case deliberately does not adopt the tempting reading that a rigorous trial debunked an enthusiastic deployment report: the two tested different interventions, and saying otherwise would repeat the confound the trial’s authors warn about. causal is the correct strength here and is used sparingly in this library. The outcome is filed negative rather than mixed or inconclusive, argued in the evidence assessment. Related to AAI-2026-020 as the case where a proxy improved and the outcome went unmeasured, against this one where the outcome was measured and did not follow; and to AAI-2026-006 as a second healthcare case in which a proxy moved and the clinical outcome behind it did not.