Measure the algorithm in your own hospital

At two in the morning in a regional emergency department, a chest radiograph of a breathless sixty-year-old may wait hours for a radiologist's report. At our hospitals in Valenciennes and Dunkirk in late 2023, routine films sometimes waited two or three days, because three radiologist posts had been vacant for more than a year. An algorithm that reads every chest film within minutes and flags the urgent ones is an obvious answer to that problem. We bought one. Before any clinician saw its output, we spent four months and €310,000 finding out whether it worked on our patients.

That decision was contested inside the group, and I understand why. The tool carried a CE mark. The vendor had a published study of more than 40,000 images. Patients were waiting. Delaying a certified product to test it again can look like bureaucratic caution dressed up as rigour. I want to explain why I think local measurement is not an optional extra but part of the price of using these tools at all, and why the boards of hospital groups should insist on paying it.

What a vendor's study cannot tell us

A CE mark under the EU Medical Device Regulation tells a hospital that the manufacturer has shown the device conforms to the regulation's requirements for its stated intended use, with clinical evidence behind that claim. It is a meaningful assurance and I do not want to belittle it. But it is an assurance about the product, measured on data the manufacturer chose. It says much less about my scanners, my radiographers' positioning habits, my patients' ages and illnesses, or the share of films in my queue that actually show the condition the algorithm looks for.

Each of those matters. Sensitivity and specificity are properties of a test applied to a population, and they move when the population moves. A study rich in obvious pneumothoraces will flatter a tool that struggles with subtle ones. Images drawn mostly from upright films taken in outpatient departments say little about supine films taken on a portable machine beside an intensive-care bed, where tubes, lines and drains cross the lungs. The manufacturer's study was honest. It simply described somewhere else.

Our own measurement showed exactly this. On upright adult films from emergency and outpatient departments, the tool detected pneumothorax with a sensitivity of 0.94, close to the vendor's reported figure. On supine portable films from intensive care, sensitivity fell to 0.68, and drains and catheters triggered so many flags that only about one in five was correct. Neither figure appeared in any document we had been given. Both were plain once we looked at our own images.

Regulation will add requirements over the coming years. Europe's AI Act treats most AI that forms part of a regulated medical device as high-risk, and obligations for the companies that supply such systems and for the organisations that deploy them will arrive in stages. I welcome that. But no regulation will measure my hospital for me. The duty to know how a tool performs on the patients in front of us stays with the people who decide to use it.

Silent running before live use

The method we use is not novel, and it is not expensive by the standards of a hospital's equipment budget. For a defined period the algorithm reads every eligible image, but its output goes only to the data team. Radiologists report as usual, unaware of the tool's opinion. We then compare the two, and a panel of three radiologists adjudicates every disagreement, because a radiologist's report is a good reference but not a perfect one. Our first silent run covered 26,400 films at three hospitals, and the panel reviewed nearly two thousand discordant cases.

The results allowed a decision more precise than yes or no. We went live for upright adult films in emergency and outpatient settings at all six hospitals, and excluded intensive care and children, who were outside the product's intended use but had been arriving in its queue anyway. In 2025 the tool read 182,000 films, and the median time to report a flagged urgent film at the two hospitals with vacancies fell from 31 hours to four. That is the benefit our chief of radiology wanted, delivered in the places where the tool had earned trust.

I should be clear about what silent running cannot do. It measures the algorithm, not the combined performance of a radiologist who can see the algorithm's opinion. Clinicians shown a confident flag may defer to it; clinicians who distrust a tool may dismiss it when it is right. So after go-live we track how often radiologists overrule the flag, and we audit a monthly sample of films the tool passed as normal. Those measures are less tidy than a sensitivity estimate, and they matter more.

The line nobody writes into the tender

The hardest lesson came after deployment. In November 2025 we replaced a radiography unit at Arras. Within three weeks our monthly monitoring showed the tool's specificity falling at that site. We suspended the flag there for five weeks while the vendor and our radiographers investigated. The cause was a change in the new unit's image-processing settings, which the radiographers reversed. Nobody had done anything wrong. A hospital had bought new equipment, as hospitals do, and the algorithm's performance had shifted without anyone deciding that it should.

That episode convinced me that monitoring belongs in the contract. Most hospital tenders for imaging AI specify the licence, integration and support. Few require the supplier to provide performance data by site and by scanner, to warn us before an update that may change behaviour, or to share the work of investigating a drift. Our first validation and our monitoring were paid for from my department's budget, which is the wrong place, because it makes measurement look like an overhead rather than a condition of use. Our next two imaging tenders will set aside a defined share of contract value for local validation and continuing monitoring, and will oblige the supplier to support both.

Boards do not need to understand neural networks to oversee this well. They need to ask four questions about every algorithm that touches patients. On whose patients, and on which equipment, was it measured? Who measures it now, and how often? What would lead us to switch it off, and who has the authority to do so? And where in the budget does that work sit? A management team that answers those questions plainly is using AI. One that points to the CE mark is relying on someone else's hospital.

An algorithm's accuracy is not a fixed property of the product. It is a measurement, taken in a particular place, on particular patients, at a particular time. The only accuracy that matters to our patients is the one we have measured ourselves, and recently.