AI10 Multicentre prospective clinical performance analysis of an artificial intelligence as a medical device deployed within UK NHS urgent suspected skin cancer pathways: device performance does not vary with cancer incidence

Abstract

Urgent suspected cancer (USC) referrals continue to rise rapidly, and solutions are needed that scale across populations with different local incidences of skin cancer. This study aimed to evaluate the performance of an artificial intelligence as a medical device (AIaMD) in NHS sites with varying skin cancer incidence. The CE Class III AIaMD is intended for use in the screening, triage and assessment of skin lesions suspicious for skin cancer. Data were collected prospectively between December 2023 and August 2025. Sites were eligible if ≥ 50 high-risk cancers had been confirmed by histology. In total, 15 USC pathways were included. Local general practice conversion rates were used as a proxy for local skin cancer incidence, which ranged from 2.9% to 9.7% across sites. In total, 62 100 lesions were assessed by the AIaMD, of which 50 247 lesions had a final diagnosis, including 2589 ­biopsy-confirmed high-risk cancers. Sensitivity for high-risk skin cancer was above the 95% target in 14 of 15 sites (range 95.7–100%), and the negative predictive value (NPV) for high-risk skin cancer was above the 99% target in 15 of 15 sites (range 99.4–100%). In the site where the sensitivity was below 95%, the upper confidence interval (CI) exceeded this threshold (sensitivity 92.7%, 95% CI 86.1–96.2%), and all melanomas were detected (18 of 18). There was no observable association between conversion rate and NPV (Spearman’s rs = 0.055, two-tailed P = 0.85). The performance of the AIaMD was observed to be consistent across geographically distinct sites with varying cancer incidence. This supports the generalizability of AIaMD across diverse clinical settings. The authors are employed by the AI provider.