A study by MIT researchers has found that doctors are less accurate at diagnosing skin diseases in patients with darker skin tones. Over 1,000 dermatologists and general practitioners took part in the study, which found that dermatologists correctly diagnosed images of skin diseases 38% of the time, with the figure falling to 34% for darker skin. Similarly, general practitioners, who were overall less accurate, were also less accurate with darker skin.
However, the study found that an artificially intelligent (AI) tool could improve the accuracy of diagnoses, although the improvements were more significant with lighter skin tones. The researchers suggested that the poor representation of darker skin tones in dermatology textbooks and training materials may contribute to the discrepancy. The study’s lead author, Matt Groh, who is an assistant professor at Northwestern University, suggested the findings may help inform changes in dermatology education policies.
To evaluate accuracy, the researchers compiled 364 images of 46 skin diseases, which were then presented to study participants who were asked to provide three potential diagnoses and whether they would refer the patient for a biopsy. Dermatology specialists had a higher accuracy rate of 38% compared to 19% for general practitioners. However, both groups’ accuracy fell by four percentage points when diagnosing conditions on darker skin, a statistically significant drop.
After testing the participants unaided, the researchers presented them with images to be analyzed with an AI tool they had developed that had an accuracy rate of 47%. The use of the AI improved the accuracy of the dermatologists to 60% and the general practitioners to 47%. Importantly, it appeared that the doctors were adept at disregarding AI inputs that were incorrect, suggesting they were good at ruling out diseases.
Despite these overall improvements, the AI was found to be less successful at improving the accuracy of the general practitioners when diagnosing conditions on darker skin. The researchers hope their findings will trigger changes in the way darker skin tones are represented in dermatology education, and guide the deployment of AI assistance in the field.
