BT11 A service evaluation of the NHS skin cancer 2-week wait pathway: comparing traditional face-to-face dermatology with an AI-enabled teledermatology service

Abstract

AI is increasingly being integrated into dermatology to support early skin cancer detection. This service evaluation assessed the safety and diagnostic performance of an AI-assisted teledermatology system vs. a traditional face-to-face (F2F) 2-week-wait dermatology pathway. A retrospective service evaluation was conducted on 539 patients referred with suspected high-risk skin lesions. Two pathways were analysed: (i) F2F dermatologist assessment (n = 271) and (ii) AI-supported teledermatology assessment (n = 268). In the AI pathway, lesions were photographed and analysed solely by AI, while clinical history and lesion information were recorded separately and reviewed only by clinicians. Dermatologists reviewed AI outputs alongside full clinical information to determine excision or discharge. Outcomes including sensitivity, specificity, positive and negative predictive values, excision rate and discharge rate were compared with those from histology. A simulated AI-only pathway, relying exclusively on image-based AI diagnoses, was also evaluated. In the F2F cohort, dermatologist–histology concordance was 80.3%. In the AI-supported teledermatology group, the discharge rate increased (55.2% vs. 51.3%), reflecting potential triage efficiency. The hybrid AI–clinician model achieved high sensitivity (89.3%) but low specificity (13.8%). Failure of image acquisition occurred in 17% of AI assessments. The simulated AI-only pathway demonstrated high sensitivity (89.3%) but low specificity (13.8%), with positive and negative predictive values of 50.0% and 57.1%, respectively. AI-supported teledermatology can safely augment clinician decision making and improve efficiency in high-risk skin lesion pathways. However, standalone AI triage is limited by low specificity and technical constraints. The diagnostic discordance observed in the AI-only simulation may reflect the absence of clinical context in image-only assessment; incorporating clinical information alongside image analysis could improve diagnostic accuracy. A hybrid AI–clinician model currently represents the most reliable and pragmatic approach for patient management.