The dermatology team at St Luke's Hospital in Bradford began using an artificial intelligence system in April to examine suspicious skin lesions. The clinic now sees 32 patients a session, up from 24 — a third more people through the same door.
The system is called DERM, for Deep Ensemble for the Recognition of Malignancy, built by a company called Skin Analytics. A healthcare assistant photographs the mole or lesion three times and uploads the images; the analysis comes back in minutes. Benign cases get advice and are discharged. Suspicious ones go to a specialist in the same visit.
Consultant plastic surgeon Zakir Shariff said the technology had hit "all the bottlenecks — right from waiting times for patients to unnecessary investigations being done and to being seen by a clinician". Unnecessary biopsies are down about 10%.
What "99.9% accurate" is a claim about
Skin Analytics says the AI is 99.9% accurate at ruling out melanoma. That phrasing is precise and deserves to be read precisely, because it is not the same as being 99.9% accurate at finding cancer.
Ruling out is the easier direction. Most lesions people worry about are harmless, so a system can be very good at saying "this is fine" while being considerably less good at catching the rare dangerous one. The two abilities are measured separately and a single headline number usually describes only one of them.
This is not a criticism of the claim, which is narrow and honest as stated. It is a caution about how such numbers travel. A hospital buying a triage tool needs the other figure too — how often the system misses something — and that figure is the one that determines whether a screen is safe to act on.
Bradford's answer to that is in the workflow rather than the statistic: every image is double-checked by a clinician. The AI is not making the decision; it is ordering the queue.
The patient in the story is the system working
Lawrence Patten, 73, was referred urgently over a mole on his back. He has multiple myeloma and a history of skin cancer on his forearm. The AI flagged his mole as suspicious; a specialist who then examined it thought it likely benign. It will be removed and tested to be sure.
That looks like the AI being wrong, and in a narrow sense it was. It is also exactly what a screening tool is supposed to do. In triage you want errors to land on the side of sending someone for a closer look, not on the side of sending them home. Mr Patten's reaction was about the thing that actually changed: "It's not ticking away in my mind, which is one of the biggest things with cancer."
Why this shape of tool fits Bangladesh, and what would break it
Bangladesh has very few dermatologists for its population, and almost all of them are in Dhaka and the divisional cities. The result is not long waiting lists; it is that most suspicious lesions are never shown to anyone qualified at all. A tool that lets a trained assistant photograph a lesion and get a triage answer in minutes addresses that gap more directly than it addresses Bradford's.
Two things would have to be true, though, and only one of them is about technology.
The first is the training data. Dermatology AI has a documented weakness: these systems have historically been trained on predominantly light skin, and perform worse on darker skin — where some cancers also present differently and are caught later. A 99.9% figure earned on one population is not transferable to another, and the question to ask a vendor here is not how accurate the model is but on whose skin it was measured.
The second is what happens after a flag. A triage system only helps if there is somewhere to triage to. Bradford's suspicious cases walk down the corridor to a specialist the same morning. If a flagged patient in a rural upazila has no realistic route to a biopsy, the system has converted an unknown worry into a known one and changed nothing else — which is worse than it sounds, not better.
On the plumbing that decides whether such results reach a patient record at all, see the piece on Bangladeshi hospital software.
