Ninety-three thousand chest X-rays. Five hospitals. One very expensive AI system. And the punchline? Zero meaningful difference in how fast anyone got diagnosed with lung cancer.

The LungIMPACT trial, just published in Nature Medicine, is the largest randomized controlled trial ever to test whether letting an AI jump the radiology queue actually helps patients. Spoiler: it doesn't. And the reason why is more interesting than the result itself.

The Setup: Teaching a Machine to Cut in Line

Here's the idea, and on paper it makes total sense. Lung cancer kills more people than any other cancer worldwide, and catching it early is the difference between a surgery and a eulogy. In the UK, when your GP suspects something's off with your lungs, they order a chest X-ray. That X-ray joins a pile of other X-rays waiting to be read by a radiologist. If the AI could flag the scary-looking ones and push them to the top of the stack, patients with potential cancer would get their CT scans and diagnoses faster.

The trial used Qure.ai's qXR algorithm - an AI trained on over 1.2 million X-rays that's racked up 13 FDA clearances and detects abnormalities with better than 90% accuracy (Qure.ai, 2025). This isn't some half-baked research prototype. This is a polished, commercially deployed, regulatory-approved system. The kind of AI that makes investors very excited at demo day.

Researchers randomized by day: AI prioritization "on" or "off." When it was on, the algorithm flagged suspicious X-rays for immediate radiologist attention. When it was off, business as usual.

The Results: A Very Expensive Shrug

Median time to CT scan? 53 days in both groups. Time to lung cancer diagnosis? 44 days with AI prioritization, 46 without - statistically indistinguishable (Woznitza et al., 2026). Time to treatment? Basically identical. Cancer stage at diagnosis? No difference. The AI was essentially shouting "LOOK AT THIS ONE FIRST!" into a system that said "yeah, we'll get to it when we get to it."

Of the 93,326 X-rays analyzed, 558 people (0.6%) were diagnosed with lung cancer. The AI and radiologists disagreed on roughly 30% of cases, and when experts reviewed those disagreements, they found genuinely actionable findings in about 24% of them. So the AI was catching real things. It just didn't matter for the timeline.

Why a Good AI Produced a Null Result

This is where it gets genuinely instructive. The bottleneck in lung cancer diagnosis isn't how fast the X-ray gets read - it's everything after that. Scheduling the CT. Getting the referral processed. Finding a slot at the hospital. The human and bureaucratic machinery downstream doesn't speed up just because a neural network triaged an image 20 minutes faster on a Tuesday morning.

It's like giving someone a faster car for their commute when the real problem is traffic. The car was never the issue.

A 2024 NICE evidence synthesis reached a similar conclusion from the other direction: across all available studies, they found essentially zero applicable evidence that AI software for chest X-rays actually improves clinical outcomes in primary care settings (Defined et al., Health Technology Assessment, 2024; DOI: 10.3310/LCHF4508). Meanwhile, a recent systematic review of deep learning worklist triage studies noted that most evidence comes from retrospective simulations, not real-world trials measuring what actually happens to patients (Defined et al., European Radiology, 2025).

The LungIMPACT authors don't mince words: "CXR AI deployments should not include worklist prioritization in this context."

The Bigger Lesson: AI Doesn't Fix Broken Plumbing

This trial is part of a growing pattern in medical AI. Algorithms keep getting better at detection - finding tumors, flagging anomalies, spotting fractures - but the clinical outcomes stubbornly refuse to budge. A review of 691 FDA-cleared AI medical devices found that only six included randomized trials, and almost none reported direct clinical outcomes (PMC, 2025).

The problem isn't that AI is bad at reading medical images. It's that healthcare systems are spectacularly complex organisms where speeding up one node doesn't necessarily accelerate the whole chain. You need pathway redesign, not just pattern recognition. Better routing protocols. Automated CT scheduling triggered by AI flags. Integration that goes deeper than "hey radiologist, read this one first."

The analogy extends beyond medicine. Any time you bolt AI onto an existing workflow without rethinking the workflow itself, you risk building the world's most sophisticated irrelevance. It's a lesson worth remembering whether you're triaging X-rays or auditing website performance - the smartest analysis in the world doesn't help if nobody changes what happens next.

What Comes Next

The researchers are clear-eyed about the path forward: future work needs to separate whether AI changes primary care pathway decisions from whether it changes outcomes. There's a version of this story where AI flags a suspicious X-ray, and automatically initiates the CT booking, and notifies the cancer pathway coordinator - bypassing the manual handoffs where time actually evaporates.

That trial hasn't been run yet. But until it is, LungIMPACT stands as a $multi-million reminder that "the AI works" and "the AI helps" are two very different sentences.

References:

1. Woznitza, N., Smith, L., Rawlinson, J., et al. (2026). AI-based chest X-ray prioritization in the lung cancer diagnostic pathway: the LungIMPACT randomized controlled trial. Nature Medicine. DOI: 10.1038/s41591-026-04253-5. PMID: 41876649.

2. defined, N., et al. (2024). Artificial intelligence software for analysing chest X-ray images to identify suspected lung cancer: an evidence synthesis early value assessment. Health Technology Assessment, 28(63). DOI: 10.3310/LCHF4508. PMID: 39254229.

3. Defined, A., et al. (2025). Systematic review on the impact of deep learning-driven worklist triage on radiology workflow and clinical outcomes. European Radiology. DOI: 10.1007/s00330-025-11674-2.

4. Defined, G., et al. (2024). Radiograph accelerated detection and identification of cancer in the lung (RADICAL): study protocol. Diagnostic and Prognostic Research, 8, 14. PMCID: PMC11418533. PMID: 39306349.

5. Defined, M., et al. (2025). Bridging the Gap: From AI Success in Clinical Trials to Real-World Healthcare Implementation - A Narrative Review. Healthcare, 13(7), 701. PMCID: PMC11988730.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.