WEDNESDAY, Oct. 2, 2024 -- An artificial intelligence (AI) model, VisualDx, demonstrates diagnostic bias for images in the Fitzpatrick I to III subgroup, according to a study published online Sept. 15 in SKIN.

Katrina D. Cirone, M.D., from Western University in London, Ontario, Canada, and colleagues examined whether the AI model VisualDx performs differently when classifying conditions across different skin phenotypes. Sixteen common conditions were included in the image dataset. For each condition, there were three subgroups: "Fitzpatrick I to III," "Fitzpatrick IV to VI," and "Processed," which contained images from the Fitzpatrick I to III subgroup altered to resemble skin with rich pigmentation. A differential diagnosis list was obtained by processing images with VisualDx.

The researchers found the highest sensitivity (97 percent) for hidradenitis suppurativa, prurigo nodularis, and tinea versicolor across all subgroups. The lowest sensitivity was seen for atopic dermatitis, postinflammatory hyperpigmentation, and basal cell carcinoma (23, 23, and 27 percent, respectively). For all conditions except acanthosis nigricans, melasma, and melanoma, significantly greater diagnostic sensitivity was noted in the Fitzpatrick I to III subgroup than the Fitzpatrick IV to VI and Processed subgroups. A reduction in sensitivity and specificity was seen for all conditions for the processed images.

"Future training of AI diagnostic tools should be done using more comprehensive databases so that diagnosis can be accurately performed across all skin phenotypes," the authors write.