Abstract
Skin conditions affect an estimated 1.9 billion people worldwide. A shortageof dermatologists causes long wait times and leads patients to seekdermatologic care from general practitioners. However, the diagnostic accuracyof general practitioners has been reported to be only 0.24-0.70 (compared to0.77-0.96 for dermatologists), resulting in referral errors, delays in care,and errors in diagnosis and treatment. In this paper, we developed a deeplearning system (DLS) to provide a differential diagnosis of skin conditionsfor clinical cases (skin photographs and associated medical histories). The DLSdistinguishes between 26 skin conditions that represent roughly 80% of thevolume of skin conditions seen in primary care. The DLS was developed andvalidated using de-identified cases from a teledermatology practice serving 17clinical sites via a temporal split: the first 14,021 cases for development andthe last 3,756 cases for validation. On the validation set, where a panel ofthree board-certified dermatologists defined the reference standard for everycase, the DLS achieved 0.71 and 0.93 top-1 and top-3 accuracies respectively.For a random subset of the validation set (n=963 cases), 18 clinicians reviewedthe cases for comparison. On this subset, the DLS achieved a 0.67 top-1accuracy, non-inferior to board-certified dermatologists (0.63, p<0.001), andhigher than primary care physicians (PCPs, 0.45) and nurse practitioners (NPs,0.41). The top-3 accuracy showed a similar trend: 0.90 DLS, 0.75dermatologists, 0.60 PCPs, and 0.55 NPs. These results highlight the potentialof the DLS to augment general practitioners to accurately diagnose skinconditions by suggesting differential diagnoses that may not have beenconsidered. Future work will be needed to prospectively assess the clinicalimpact of using this tool in actual clinical workflows.