DS03: A scoping review of commercially available artificial intelligence-based technologies for diagnosis or risk assessment of skin cancer

David Wen,^1,2 April Coombe,³ Maria Charalambides,⁴ Owain Jones,⁵ Nina Wietek,¹ Jac Dinnes³ and Rubeta Matin⁶

¹University of Oxford, Oxford, UK; ²Royal Berkshire NHS Foundation Trust, Reading, UK; ³University of Birmingham, Birmingham, UK; ⁴Sandwell & West Birmingham Hospitals NHS Trust, Birmingham, UK; ⁵University of Cambridge, Cambridge, UK; and ⁶Oxford University Hospitals NHS Foundation Trust, Oxford, UK

Artificial intelligence (AI)-based technologies for skin cancer detection have the potential to support the management of suspicious skin lesions but without appropriate validation and implementation in the correct clinical setting there is a potential for harm through misdiagnosis. This scoping review aimed to identify AI-based technologies currently available in the UK for skin cancer diagnosis or risk assessment, and to evaluate their intended use, cost and regulatory approvals. Technologies were identified through relevant dermatology AI systematic reviews identified using a formal search strategy, and by searching app stores, social media platforms and medical technology websites. Data were collected by accessing source material, including developers’ websites and app store webpages. Overall, 65 technologies were identified with 82% primarily being smartphone or tablet applications (n = 53). Five (8%) were integrated into skin-imaging devices or systems. A fee was charged for 38% (n = 25), which included one-off fees, regular subscription fees and charge per item analysed; 34% (n = 22) were free (including four that had free versions) and 18% (n = 28) provided no pricing information. Sixty-six per cent of technologies (n = 43) were intended for use by patients/public, 20% (n = 13) for clinicians and 11% (n = 7) for use by both patients and clinicians. Technology inputs included macroscopic photographs alone (69%; n = 45), dermoscopic images alone (9%; n = 6), macroscopic and dermoscopic images (3%; n = 2), and clinical information in combination with macroscopic photographs (17%; n = 11). Technology outputs included a diagnostic label (n = 13; 20%), a probability score for predicted diagnoses (n = 24; 37%), listing differential diagnoses (n = 5; 8%) or providing a binary output [e.g. malignant vs. nonmalignant (n = 5; 8%)]. Nine technologies (14%) provided additional treatment recommendation, including advice for further review by a medical professional. The output was unclear for 10 technologies (15%). Regulatory approvals were available for 16 (25%); nine reported CE Class I marks, six reported CE approval (class not stated) and two stated Health Canada approval. This review of current technologies commercially available for diagnostic purposes in the skin cancer setting has highlighted insufficient regulation, with 75% having no clear regulatory oversight. For technologies available to the public that are clearly intended for diagnostic use but lack appropriate regulation, there is a potential risk for harm through misdiagnosis. The British Association of Dermatologists is working with the Medicines and Healthcare products Regulatory Agency and Apple and Google Play app stores to address this safety issue. Future work will focus on evaluating published and unpublished supporting evidence for diagnostic accuracy, safety and cost-effectiveness.

Funding sources: this work was funded by the British Association of Dermatologists.