PT - JOURNAL ARTICLE AU - Kamulegeya, Louis Henry AU - Okello, Mark AU - Bwanika, John Mark AU - Musinguzi, Davis AU - Lubega, William AU - Rusoke, Davis AU - Nassiwa, Faith AU - Börve, Alexander TI - Using artificial intelligence on dermatology conditions in Uganda: A case for diversity in training data sets for machine learning AID - 10.1101/826057 DP - 2019 Jan 01 TA - bioRxiv PG - 826057 4099 - http://biorxiv.org/content/early/2019/10/31/826057.short 4100 - http://biorxiv.org/content/early/2019/10/31/826057.full AB - Introduction Artificial intelligence (AI) in healthcare has gained momentum with advances in affordable technology that has potential to help in diagnostics, predictive healthcare and personalized medicine. In pursuit of applying universal non-biased AI in healthcare, it is essential that data from different settings (gender, age and ethnicity) is represented. We present findings from beta-testing an AI-powered dermatological algorithm called Skin Image Search, by online dermatology company First Derm on Fitzpatrick 6 skin type (dark skin) dermatological conditions.Methods 123 dermatological images selected from a total of 173 images retrospectively extracted from the electronic database of a Ugandan telehealth company, The Medical Concierge Group (TMCG) after getting their consent. Details of age, gender and dermatological clinical diagnosis were analyzed using R on R studio software to assess the diagnostic accuracy of the AI app along disease diagnosis and body part. Predictability levels of the AI app was graded on a scale of 0 to 5, where 0-no prediction made and 1-5 demonstrating reducing correct prediction.Results 76 (62%) of the dermatological images were from females and 47 (38%) from males. The 5 most reported body parts were; genitals (20%), trunk (20%), lower limb (14.6%), face (12%) and upper limb (12%) with the AI app predicting a diagnosis in 62% of image body parts uploaded. Overall diagnostic accuracy of the AI app was low at 17% (21 out of 123 predictable images) with varying predictability levels correctness i.e. 1-8.9%, 2-2.4%, 3-2.4%, 4-1.6%, 5-1.6% with performance along individual diagnosis highest with dermatitis (80%).Conclusion There is a need for diversity in the image datasets used when training dermatology algorithms for AI applications in clinical decision support as a means to increase accuracy and thus offer correct treatment across skin types and geographies.