A new study by MIT researchers reveals that doctors are less accurate in diagnosing skin diseases in darker-skinned patients when relying solely on images of the patient’s skin. The research involved over 1,000 dermatologists and general practitioners and showed dermatologists correctly identified about 38% of images they saw, and just 34% of the images displaying darker skin. Accuracy rates decreased similarly among general practitioners when dealing with darker skin.
Assistance from an artificial intelligence (AI) algorithm could boost doctors’ accuracy. However, the research indicates that improvements were greater for diagnosing patients with lighter skin. This study is the first to highlight physician diagnostic disparities across skin tone. Similarly, researchers identified that dermatology textbooks and training materials are primarily featuring images of light skin tones, which could contribute to the diagnostic discrepancy.
To evaluate the doctors’ diagnostic accuracy, the research team compiled an array of 364 images representing 46 skin diseases across various skin shades. The images were from dermatology textbooks and other sources and mainly depicted eight inflammatory skin diseases.
Regardless of physician type, the accuracy decreased by nearly four percentage points when the doctors diagnosed skin conditions with images of darker skin. The research findings emphasize the potential of AI in improving physicians’ decision-making capabilities and therefore, improving patient outcomes.
After further analysis of the images assisted by an AI algorithm, the researchers identified that the algorithm improved the accuracy for the dermatologists (up to 60%) and the general practitioners (up to 47%). They found physicians were more likely to take accurate suggestions from the AI regardless of the skin color, but often dismissed incorrect suggestions. This suggests that physicians use their expertise for ruling out diseases and would not take suggestions for an ailment they’ve already ruled out.
However, general practitioners showed significant improvements on lighter skin images compared to dark skin images when using the AI algorithm. The researchers emphasized the need for more training on diagnosing patients with darker skin in medical schools and textbooks besides leveraging AI assistance. The results are believed to inform the development of AI assistance programs for dermatology, which many companies are presently exploring. The research received funding from the MIT Media Lab Consortium and the Harold Horowitz Student Research Fund.
