Skin complaints are among the most common reasons patients visit their GP, yet most GPs receive only limited dermatology training. When a case is uncertain, the safe default is a specialist referral. That instinct is right for the individual patient and expensive in aggregate, adding to dermatology waiting lists already under pressure. An earlier primary care study (Escalé-Besa et al., 2023) found that AI support cut specialist referrals by 34% in that setting.
This study asked a narrower question: does the same hold for UK GPs, and does their accuracy improve alongside it? Ten experienced UK GPs reviewed the same 20 real patient cases. Five worked unaided. Five saw Autoderm’s top-5 ranked condition suggestions beside each case. Both groups were scored against a dermatologist reference standard on the condition they placed first, their top-3 differentials, their management recommendations, and the time they spent.
The design isolates a single variable, which is whether ranked suggestions were present. It does not measure the AI on its own. It measures what a GP does with it. Autoderm’s clinical evidence record
Top-1 accuracy rose from 48% without AI to 69% with AI, an improvement of 21 percentage points.
Top-3 accuracy rose from 74% to 81%.
With AI assistance, GPs identified 81% of cases correctly within their top 3 differential diagnoses.
Dermatology referrals fell from 37 to 22 cases, a reduction of 15 cases (40%).
Average time per case was significantly lower in the AI-assisted group.
All GPs in the AI-assisted group said the suggestions helped them consider additional differential diagnoses, and all said they would use the tool in similar settings.
We then created a second version of the same questionnaire. It contained the identical cases and questions, with one addition: five possible answers generated by the AI, ranked in order.
Ten UK-based GPs with between 8 and 25 years of clinical experience took part. Five received the questionnaire with AI assistance (Group B, the study group) and five received it without (Group A, the control group). Every GP answered based on their own knowledge and experience; those in Group B could review the AI’s ranked suggestions and take them into consideration when forming their diagnoses.
At the end of the survey, both groups rated the difficulty of the task. Group B additionally gave feedback on how helpful the AI assistance was and whether they would use it in practice.
The correct diagnosis for each case was established by board-certified dermatologists via the First Derm platform. This served as the reference standard against which both groups were scored on diagnostic accuracy, management recommendations and time per case. A reader study with medical students
Bar chart comparing correct top-1 answers across 20 cases, with and without AI support
Bar chart comparing correct answers within the top three differentials across 20 cases, with and without AI support
Column chart summarising top-1 and top-3 correct answer rates, unaided compared with AI-supported
Bar chart comparing correct management recommendations per case, with and without AI support
Dermatology referrals
Bar chart comparing dermatology referrals per case, with and without AI support
Bar chart comparing average seconds spent per case, with and without AI support
The control group, working without AI, rated the overall difficulty of the task higher than the AI-assisted group did.
The AI did not replace the GPs’ clinical judgement. It supported it, prompting differentials the GPs might not otherwise have considered while leaving every decision with the clinician. That division of responsibility is not incidental to the study design. It reflects Autoderm’s intended purpose and regulatory status.
