A Comparative Analysis of Large Language Model Accuracy for Image-Based Hair Disease Identification in Diverse Skin Tones

Abstract

Purpose

The rapid integration of artificial intelligence (AI) in dermatology shows promise for support of clinical practice and democratization of diagnosis access. Significant limitations and ethical concerns persist, however. Despite growing research into AI's effectiveness in identifying skin conditions, fewer studies have explored its ability to accurately diagnose hair disorders.

Methods

This study explores the capacity of the large language model (LLM) ChatGPT to correctly identify alopecia areata, androgenetic alopecia, traction alopecia, and central centrifugal cicatricial alopecia across a range of skin tones. Utilizing the Monk Skin Tone Scale, images of hair disorders were sorted into lighter (Monk Scale 1-5) and darker (Monk Scale 6-10) categories. Images were sourced from publicly accessible databases.

Results

Our analysis revealed significant differences in diagnosis rates. ChatGPT was more likely to correctly identify disease in lighter skin, notably for alopecia areata (p<.001) and androgenetic alopecia (p=.003). This trend was also seen in overall diagnosis rates (p<.001). Interestingly, the program repeatedly incorrectly identified 24.48% of all hair conditions in dark skin as traction alopecia. Additionally, while initially this study sought to explore ChatGPT's ability to diagnose common nail disorders across skin tones this could not be completed due to the insufficient availability of images depicting nail disorders in darker skin.

Conclusion

These findings highlight some of the limitations of LLMs in accurate diagnosis of diseases of the hair and nails. It emphasizes potential implications for the performance of artificial intelligence trained on dermatologic databases with limited representation.