Melanoma Incidence Trends by Population Groups in South Africa, 1991–2021

In-app teledermatology review occurred in 65.7% (1042/1587) of images and significantly increased overall specificity (86.8% (85.0 to 88.6) (p < 0.001)), at the expense of lower overall sensitivity (75.3% (68.0 to 81.7) (p = 0.023)).Risk classification of included malignant and benign/premalignant skin lesions at each stage of the app assessment, compared to the final clinical or histopathological diagnosis.Numbers represent the number of assessed lesions per category.CNN: Convolutional Neural Network.Combined advice: Risk score after teledermatological revision.Taking into account the 16.6% undetectables, the worst case sensitivity and specificity according to the STARD guidelines were 67.6% (60.3-74.3)and 71.8% (69.6-73.9),respectively.Of the 32 melanomas, nine melanomas were misclassified, and one was undetectable (scalp location), resulting in a false negative rate of 31.3% (10/32).Detectability and diagnostic accuracy varied between smartphone devices, and 71.3% of participants were unable to use the app themselves.A survey of 446 participants showed that sensitivity was considered more important than specificity, with 90% expecting a sensitivity of at least 85%.Conclusions: This is the first large, independent study evaluating the diagnostic performance of a CE-certified AI-based smartphone app for skin cancer detection in a prospective cohort that approximates actual end-users.It underscores the importance of independent clinical validation of deep learning algorithms in healthcare in the intended setting, and advocates for stricter regulation to ensure adequate protection of the public.