BT05 Comparing the number-needed-to-biopsy ratio for melanoma diagnosis between teledermatology, an artificial intelligence device with teledermatology, and face-to-face models of care
Radhika Bali,Jenny Chung,Lucy Thomas,Khawar Hussain,Louise Fearfield
DOI: https://doi.org/10.1093/bjd/ljae090.403
IF: 11.113
2024-06-28
British Journal of Dermatology
Abstract:Abstract Traditional face-to-face (FTF) models of care are shifting towards use of hybrid FTF–teledermatology (TD) models, with certain UK centres adopting the additional use of artificial intelligence (AI) medical devices in suspected 2-week-wait (2WW) skin cancer referrals. The number-needed-to-biopsy (NNB) ratio is calculated by dividing the total number of biopsies by the number of biopsied (excised) skin cancers. NNB is increasingly used as a metric for diagnostic accuracy of malignant melanoma (MM) detection, with low ratios suggesting high specificity and vice versa. Our TD service has incorporated AI technology (DERM) using a marked class IIA medical device from April 2022. We therefore aimed to compare the NNB ratio for MM detection assessed via TD before and after introduction of the AI tool with FTF models of care. Retrospective analysis was performed on 50 consecutive patients in each group with suspected MM, assessed by consultant dermatologists. In the pre-AI group (31 female, 19 male; median age 51 years, range 23–94), six patients had confirmed MM (NNB 8.3). Subtypes included 4 of 6 superficial spreading melanoma (SSM), 1 of 6 nodular, and 1 of 6 in situ melanoma. The median Breslow thickness (BT) was 0.6 mm (range 0.6–7.5). Comparatively, in the post-AI cohort (36 female, 14 male; median age 53.5 years, range 22–97), seven lesions were confirmed as MM (NNB 7.1). Four of 7 were SSM, 1 of 7 lentiginous, 1 of 7 regressed and 1 of 7 in situ melanoma, with median BT 0.7 mm (range 0.4–17.1). Concordance between AI and reviewer diagnosis was 70% (35 of 50) in total and 86% (6 of 7) for confirmed MM. The accuracy of the AI compared with the histological diagnosis was 28% (14 of 50) overall, with a tendency to overdiagnose benign lesions as malignant, and 86% (6 of 7) for the MM cohort. FTF analysis (25 female, 25 male; median age 50.5 years, range 21–87) revealed six confirmed MM (NNB 8.3). Two of 6 were SSM, 2 of 6 nodular and 2 of 6 in situ melanoma, with median BT 2.8 mm (range 1–15). Our results show similar NNB ratios in both the pre-AI (8.3) and post-AI (7.1) TD cohorts, and the FTF group (8.3), which is reassuring and comparable with an NNB of 9.6, reported from a meta-analysis of 26 studies worldwide (Petty AJ, Ackerson B, Garza R et al. Meta-analysis of number needed to treat for diagnosis of melanoma by clinical setting. J Am Acad Dermatol 2020; 82: 1158–65). Decreased median BT was observed in the TD cohorts compared with FTF analysis, signalling possible earlier detection of MM. While TD, in conjunction with AI tools, may offer innovative and practical approaches to manage high volumes of 2WW referrals through extraction of benign lesions, specificity may initially be compromised. Ultimately, larger real-world datasets are required to further validate the performance of AI to assess superiority in diagnostic accuracy.
dermatology