BT11 Artificial intelligence and ethnically diverse skin: potential, pitfalls and a model for reducing risk

Abstract There is increasing demand for artificial intelligence (AI) tools to manage skin cancer; however, there is minimal diversity in images used for AI training and validation (Wen D, Khan SM, Xu AJ et al. Characteristics of publicly available skin cancer image datasets: a systematic review. Lancet Digit Health 2022; 4: e64–74). Limited validation in ethnically diverse skin, in tandem with no human ‘sense check’ on photographic input and analysis by AI, increases the risk of misdiagnosis of skin cancers. We assessed AI bias and developed a cost-effective protocol to increase diagnostic accuracy in patients with ethnically diverse skin. A dermatology nurse collected clinical history and used videodermoscopy to capture optimized macro and dermoscopic images of lesions. Data were from consented patients with self-identified ethnically diverse skin, being assessed via the 2-week-wait skin cancer teledermatology pathway over 7 months. This was combined with clinical history. Histological diagnosis (where available) and consultant diagnosis were recorded for each lesion, used as gold standard to compare Moleanalyzer pro AI assessments of lesions. This ‘real-time AI’ is run at the bedside and clearly demarcates the lesion being analysed. We analysed 69 lesions from 64 patients (45 female, 19 male) with an age range of 18–81 years (mean 49.9). The AI malignancy threshold (MT) is 0–1 (1 being highest risk of malignancy). With MT on the Moleanalyzer pro AI at 0.2, clinical–AI concordance was 75.4%, while AI sensitivity and specificity were 83.3% and 74.6%, respectively. Published results for Moleanalyzer pro AI in people with Fitzpatrick type II and III skin at MT 0.5 and above show sensitivity 81.6% and specificity 88.9%. In some cases in our study, AI picked out the incorrect darker area and not the index lesion in ethnically diverse skin. In addition, only rare subtypes of skin cancers were found in our patient population. We found that when using dermatological AI diagnostic tools, a nurse-led photography protocol is optimal. Incorporating clinical history alongside ‘real-time AI’ at the bedside, and a human ‘sense check’ on AI diagnosis, reduces risk. When using AI in an ethnically diverse population, a reduced MT must be considered to achieve a level of accuracy comparable with published data. With an effective protocol, AI can provide extensive savings and reduced waiting times. Each diagnostic AI tool developed for skin cancer should be validated with ethnically diverse skin images to reduce health inequality and potential harm to patients. We propose this is introduced into formal quality standards for AI used in skin cancer diagnosis.