A new study by researchers from MIT revealed that doctors are less accurate in diagnosing skin diseases when the patient has darker skin. The study involved over 1,000 dermatologists and general practitioners, with dermatologists accurately diagnosing about 38 percent of their caseload based on images. However, for darker skin, the accuracy dropped to 34 percent. General practitioners, less accurate overall, demonstrated a similar trend.
The team detected that the use of an artificial intelligence (AI) algorithm could improve the rate of accurate diagnoses. However, the improvements were more significant for lighter-skinned patients. This is the first study to display physician diagnostic disparities related to skin tone.
One contributing factor could be that dermatology textbooks and training materials mainly feature images of lighter skin tones. Doctors with limited experience in treating darker-skinned patients could also contribute to this discrepancy.
The study further explored whether AI algorithms could improve diagnostic abilities on darker shades of skin. They used an array of 364 images representing 46 skin diseases with varying skin shades to assess diagnostic accuracy. The experiment included atopic dermatitis, Lyme disease, secondary syphilis, and cutaneous T-cell lymphoma, conditions that present differently on different skin tones.
The study participants were shown ten images each and asked to provide their top three diagnoses for each image, as well as whether they would recommend a biopsy or refer the patient to a dermatologist.
Examining accuracy rates, the researchers found unsurprisingly that dermatologists had the highest, classifying 38 percent of images correctly compared to 19 percent for general practitioners. Both groups saw about a four percent drop in accuracy when diagnosing skin conditions in darker skin—a statistically significant decrease.
After studying individual performance, the team introduced an AI algorithm to assist diagnoses. This algorithm was trained on approximately 30,000 images, and it boosted diagnostic accuracy to 47 percent. The team also designed an artificially enhanced version of the algorithm with an 84 percent success rate. Using these algorithms amplified accuracy for both dermatologists (up to 60 percent) and general practitioners (up to 47 percent).
The study illustrated the usefulness of AI assistance and verified that physicians tend to ignore AI suggestions deemed as incorrect while relying more on AI with higher accuracy rates. However, general practitioners showed more improvement on images of lighter skin than darker skin when using AI assistance.
The findings could promote the inclusion of more training material on patients with darker skin in medical schools and textbooks, and guide the development of AI assistance programs in dermatology.
