AI helps physicians diagnose skin disease, but not for darker skin tones

Results from a new study suggest that AI assistance can significantly increase physicians’ diagnostic accuracy for skin diseases, although it fails to improve accuracy with darker skin tones.Specialists and generalists alike should seek additional training in diagnosing skin diseases in darker skin, the authors of the study concluded.The researchers recruited 1,118 dermatologists and primary care physicians to review 364 images of 46 skin diseases. The images represented all skin tones and the majority represented eight diseases: atopic dermatitis, cutaneous T-cell lymphoma, dermatomyositis, lichen planus, Lyme disease, pityriasis rosea, pityriasis rubra pilaris, and secondary syphilis. These often present differently in darker and lighter skin.Dermatologists got 38% of the diagnoses correct and primary care physicians got 19% correct. Skin diseases in dark skin (a Fitzpatrick Skin Type of 5 or 6) were diagnosed significantly less accurately than those in light skin (a Fitzpatrick Skin Type of 1 to 4). The only skin disease that dermatologists diagnosed more accurately in dark skin than light skin was lichen planus. Those primary care physicians who saw more patients with darker skin had a higher percentage of accuracy in diagnosing skin diseases in dark skin.Read: Updated international guidelines for management of atopic dermatitis (eczema)Using a deep learning system (DLS) programmed on more than 30,000 dermatology images significantly increased diagnostic accuracy. Physicians’ diagnoses improved by 33% and provided few incorrect diagnoses. When using the deep learning system, dermatologists’ accuracy for the eight main skin diseases rose from 27% to 36% and primary care physicians’ accuracy increased from 13% to 27%. The key to these physician-AI collaborations is if physicians can correctly ignore or incorporate AI suggestions, which is connected to physicians’ expertise, the authors said. Overall, physicians using the AI made few incorrect diagnoses. "They’re pretty good at not taking AI advice when the AI is wrong and the physicians are right. That’s something that is useful to know," lead author Matthew Groh, assistant professor at Northwestern University Kellogg School of Management in Evanston, Illinois, told Technology Networks’ Diagnostics.The AI didn’t help with proper diagnosis when it came to darker skin tones. Physicians were four percentage points less accurate on darker skin tones than lighter skin tones, and AI assistance actually dropped primary care physicians’ accuracy by five percentage points, a significant difference, with darker skin tones. Read: Should there be routine screening for skin cancer in healthy adults?“Although physician–machine partnerships improve overall diagnostic accuracy, we have found that the DLS-based decision support exacerbates non-specialists’ diagnostic accuracy disparities for light and dark skin,” the authors wrote. The researchers noted that the results can help guide future AI programs to aid in dermatology diagnoses. They added that doctors should get more training in diagnosing skin diseases in darker skin “to avoid the potential for systematic misdiagnoses in clinical settings that may mirror the systematic differences found in diagnosing light and dark skin in this experiment.”The paper was published in Nature Medicine.Read: Psychodermatology: a new under-recognized sub-specialty in dermatology

AI helps physicians diagnose skin disease, but not for darker skin tones

Register for free access or login.

Related Topics