A new study from the Massachusetts Institute of Technology (MIT) suggests that doctors are less accurate when diagnosing skin conditions on darker skin tones based solely on images. The study, which involved more than 1,000 dermatologists and general doctors, revealed that only 34% of images displaying darker skin were accurately diagnosed by dermatologists, compared to 38% for lighter skin tones.
The researchers also noticed a similar drop in diagnosis accuracy among general practitioners. However, the team found that using artificial intelligence (AI) can improve the accuracy of these diagnoses – although, it was more beneficial for patients with lighter skin.
This research is groundbreaking, as it’s the first study to show diagnostic disparities across different skin tones. Other studies have noted that dermatology textbooks primarily feature images of lighter skin tones, which might be part of the discrepancy, alongside the possibility that some doctors may not have much experience treating patients with darker skin.
The researchers used 364 images of 46 different skin diseases, showcasing various skin shades, to assess the diagnostic accuracy of doctors. The subjects were asked to predict what disease each image might represent and whether they would refer the patient for a biopsy, while general practitioners were also asked if they would recommend the patient to a dermatologist.
Unsurprisingly, dermatologists had higher accuracy rates at 38%, compared with 19% among general practitioners. Both groups’ accuracy dropped by roughly four percentage points when diagnosing skin diseases from images of darker skin – a statistically meaningful fall.
After evaluating the doctors’ performances, the researchers offered additional images for them to analyse with the aid of an AI algorithm they had developed. The AI, trained on around 30,000 images, had an accuracy rate of about 47%. Another version of the algorithm, with an artificially inflated success rate of 84%, helped to assess whether the model’s accuracy would influence doctors’ chances of taking its recommendations.
Both versions of the AI enhanced the accuracy for dermatologists and general practitioners to about 60% and 47% respectively. However, the researchers observed that the AI-aided diagnosis was more beneficial for images of lighter skin than for darker skin. This information could encourage medical schools and textbooks to incorporate more training on darker skin and serve as guidance for deploying AI assistance programmes for dermatology.
