Investigating the Performance of VisualDx on Common Dermatologic Conditions in Skin of Color

Abstract

Background: Artificial intelligence (AI) has been used to create diagnostic models, such as VisualDx, to assist in rapidly diagnosing skin conditions. AI diagnostic models are typically trained on image databases of dermatologic conditions, which are known to underrepresent patients with richly pigmented skin. Objectives: We investigated whether VisualDx performed differently when classifying conditions across different skin phenotypes and whether images of conditions processed to resemble richly pigmented skin impacts diagnostic accuracy. Methods: Our image dataset consisted of sixteen common conditions. For each condition, three subgroups were curated: “Fitzpatrick I-III”, “Fitzpatrick IV-VI”, and “Processed”. The “Processed” subgroup contained images from the “Fitzpatrick I-III” subgroup altered to resemble richly pigmented skin. Images were processed by VisualDx to obtain a differential diagnosis list and diagnostic performance was analysed. Results: Across all subgroups, the highest sensitivity (97%) was seen in hidradenitis suppurativa, prurigo nodularis, and tinea versicolor. Atopic dermatitis, post-inflammatory hyperpigmentation, and basal cell carcinoma demonstrated the lowest sensitivity (23%, 23%, and 27%, respectively). Significantly greater diagnostic sensitivity was noted for all conditions in the “Fitzpatrick I-III” subgroup (p < 0.001) except acanthosis nigricans, melasma, and melanoma compared to the “Fitzpatrick IV-VI” and “Processed” subgroups. For all conditions, a reduction in sensitivity and specificity was observed in processed images (p < 0.001). Conclusion: Overall, VisualDx demonstrated diagnostic bias for images in the “Fitzpatrick I-III” subgroup, and colour-editing reduced diagnostic accuracy. These results suggest comprehensive databases should be used for future training of AI diagnostic tools to improve performance in all skin phototypes.