P28 Understanding ethnic inequality and barriers to participation in artificial intelligence (AI) image analysis research in dermatology

Abstract As a visual-based specialty, dermatology has the opportunity to use artificial intelligence (AI) to aid diagnosis and guide precision management of skin diseases. To employ effective and safe AI, algorithms should be trained and validated on data sets that are representative of patient demographics in clinical practice. In our recent systematic review of AI image analysis studies in inflammatory skin conditions, ethnicity and Fitzpatrick skin type were reported in only four (6%) and 10 (16%) of 64 included studies, respectively. A systematic review of skin cancer image data sets similarly demonstrated poor ethnicity and Fitzpatrick skin type reporting, and under-representation of darker skin types (Wen D, Khan SM, Xu AJ et al. Characteristics of publicly available skin cancer image datasets: a systematic review. Lancet Digit Health 2022;4:e64–e74). Such bias in image acquisition limits the accuracy of AI tools when applied to real-world populations. We aimed to assess the representation of ethnic groups in our AI research and understand barriers to participation. We established a prospective observational cohort study in a national specialized psoriasis service based in South London, which seeks to develop AI algorithms for image-based assessment of psoriasis severity. Participants have photographs of their skin taken (both self-taken on participants’ own devices and professionally acquired in a studio) for AI image analysis. Images are securely stored in an ethically-approved research database in accordance with UK General Data Protection Regulations. Willingness to participate and reasons for declining were recorded. Of 557 patients approached between September 2021 and December 2022, 151 (27.1%) declined to participate in the AI study. The mean (SD) age of those willing and declining to participate was 47.4 (16.6) and 50.7 (16.2) years, respectively. Of those willing to participate, 57.5% were male vs. 53.6% who declined. Twenty-five per cent (n = 77/308) of individuals of white ethnicity declined to participate vs. 32% (n = 20/63) Indian/Pakistan/Bangladeshi, 33% (n = 6/18) black/black British, 50% (n = 5/10) Chinese/Japanese/Korean/Indochinese and 50% (1/2) of mixed ethnicity. Common reasons for declining participation included a disinclination to have photos taken, psoriasis affecting intimate sites, time constraints, medical reasons (e.g. poor mobility) and concerns regarding data security. Ninety-eight of 406 individuals who were willing to participate met the study-specific inclusion criteria (e.g. plaque psoriasis, Psoriasis Area and Severity Index > 3) and were recruited. Of those recruited, 82% (n = 80) were of white ethnicity, 9% (n = 9) Indian/Pakistan/Bangladeshi and 8% (n = 8) were black/black British. Seventeen (17%) had Fitzpatrick skin type V/VI. These data highlight ethnic inequalities in participation of AI image-analysis research. Further dissection of barriers to participation may inform strategies for ensuring more diverse and representative datasets, to maximize the potential of AI in healthcare.