Using artificial intelligence on dermatology conditions in Uganda: a case for diversity in training data sets for...

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Informatics

Background:

• Universal Artificial Intelligence (AI) in healthcare requires representative data from diverse geographies.
• Current AI algorithms may exhibit bias due to underrepresentation of certain populations.

Purpose of the Study:

• To evaluate the diagnostic performance of the AI-powered Skin Image Search algorithm on Fitzpatrick 6 skin type (dark skin).
• To assess the accuracy of AI in diagnosing dermatological conditions in individuals with dark skin.

Main Methods:

• Retrospective analysis of 123 dermatological images from a Ugandan telehealth company.
• Evaluation of AI diagnostic accuracy, disease diagnosis, and body part using R on R studio.
• AI predictability graded on a scale of 0-5 to assess incorrect diagnosis rates.

Main Results:

• The AI algorithm achieved a low diagnostic accuracy of 17% for dark skin types, compared to 69.9% for Caucasian skin.
• Dermatitis showed the highest individual diagnostic accuracy at 80% among the evaluated conditions.
• Predictability correctness varied, with lower percentages for higher grades (e.g., 1.6% for grade 5).

Conclusions:

• A significant need exists for diverse image datasets to train dermatology AI algorithms.
• Increased data diversity is essential to improve AI accuracy across different skin types and geographic locations.
• Addressing data gaps is critical for developing unbiased and equitable AI healthcare solutions.