P009 Safety and service impact of dermatologist second read in artificial intelligence-assisted teledermatology

Abstract

Artificial intelligence (AI)-supported teledermatology is increasingly used to triage skin lesions, aiming to improve clinical efficiency and access to services. Some platforms combine AI-based risk stratification with a dermatologist second read as a safety net, supporting discharge of low-risk lesions while escalating cases of concern. However, the downstream service implications of escalation following dermatologist second read remain unclear. The aim of this study was to evaluate the impact of a dermatologist second read in cases where AI classified lesions as benign and recommended discharge. We conducted a retrospective review of cases in an AI-assisted teledermatology pathway over a 1-month period (October 2025). We included lesions triaged as benign with a recommendation for discharge by AI, but where the dermatologist second read recommended onward referral for local dermatology input. Data were extracted from the AI platform. We recorded management outcomes, downstream appointments and procedures, and confirmed diagnoses where available. Forty cases with discordance between AI and the second-read dermatologist were identified. Of these cases, 23 (58%) were benign, 8 (20%) were premalignant and 1 (3%) was malignant (melanoma in situ). Outcomes were pending for eight lesions. Discordant cases generated 22 face-to-face (F2F) appointments, 13 minor operations and 1 day-case excision. Following F2F review, 18 cases were discharged; no additional procedures were generated at F2F review. Among lesions classified as benign by AI but escalated on second read, a clinically important minority were premalignant or malignant (9 of 40, 23%), supporting a second-read safety net within AI-assisted teledermatology pathways. However, discordant cases generated significant downstream activity. All procedural decisions were initiated at the teledermatology triage stage, with no additional procedures generated after F2F review. Limitations of this study include the small cohort size and incomplete outcomes. Future work should identify modifiable drivers of escalation (e.g. image quality) to reduce avoidable activity while maintaining safety.