When it comes to AI, people often talk about discrimination. And rightly so, because most models have built-in biases. But humans also have conscious and unconscious biases and discriminate.
According to a study by researchers at the Massachusetts Institute of Technology (MIT), doctors are less likely to diagnose skin diseases in people with darker skin.
The study found that dermatologists correctly categorized only 34 percent of diseases in images of darker skin, compared to about 38 percent in images of lighter skin.
General practitioners were less accurate overall but showed a similar decrease in diagnostic accuracy with darker skin.
The study involved more than 1,000 dermatologists and primary care physicians recruited through the social networking site Sermo for Doctors.
Participants were shown 364 images from dermatology textbooks and other sources showing 46 skin diseases on different skin tones.
Study participants were asked to indicate their top three predictions for the possible disease of each image and whether they would refer the patient for a biopsy.
AI greatly improved diagnostic accuracy, but exacerbated the gap between lighter and darker skin tones for generalists.
Both dermatologists and general practitioners benefited from AI assistance. Dermatologists improved their accuracy up to 60 percent and general practitioners up to 47 percent.
But the latter were able to increase their diagnostic ability with AI more for light skin, although the AI algorithm works equally well for light and dark skin, according to the researchers.
Doctors were also good at identifying and rejecting misdiagnoses made by the AI when their own diagnosis was correct.
One factor that may contribute to differences in diagnostic accuracy for different skin tones is that images in dermatology textbooks and training materials tend to show lighter skin tones.
Doctors may have less experience treating patients with darker skin tones and therefore have poorer results with certain groups of people.
The MIT researchers hope their findings will lead to more training and content in textbooks about patients with darker skin. They could also help guide the introduction of AI-assisted programs in dermatology, which MIT says are currently being developed by many companies.
In addition to diagnosis, AI in healthcare can be used for clinical care, patient-directed use, clerical and administrative tasks, medical and nursing education, and scientific research and drug development.
However, there are risks associated with AI in medicine. These include the potential for AI models to provide incorrect, inaccurate, biased, or incomplete information. Risks can also arise from training AI models on low-quality or biased data, and from automation bias, where healthcare professionals miss errors or inappropriately delegate decisions to AI.