Artificial Intelligence for Diagnosing Images of Common Skin Disorders

Area of Science:

• Dermatology
• Artificial Intelligence (AI)
• Machine Learning

Background:

• AI and machine learning are increasingly utilized in dermatology for diagnosis and severity classification.
• Primary care providers and emergency physicians exhibit higher diagnostic error rates for dermatologic conditions compared to dermatologists.
• This study assesses an AI tool's efficacy in identifying common skin disorders from educational images.

Purpose of the Study:

• To evaluate the diagnostic performance of an AI-driven dermatologic image analytics tool.
• To determine the accuracy of AI predictions for common skin disorders using classical medical images.

Main Methods:

• A prospective study analyzed 42 classical dermatologic images using the bellePro AI application.
• The AI tool was trained on over 400,000 dermatologic images and ranked predictions by 'image match scores'.
• Images from a national board review course underwent blinded validation by an academic dermatology department.

Main Results:

• The AI tool achieved a 90.5% positive predictive value, correctly identifying 38 out of 42 skin disorders on the first attempt.
• All 42 conditions were included in the differential diagnosis list across three attempts, indicating comprehensive identification.
• Correct diagnoses had an average image match score of 0.896 (SD = 0.098), with no correct diagnoses scoring below 0.64.

Conclusions:

• The AI-based dermatologic image analytics tool demonstrates effective performance on classical images of common skin disorders.
• This AI tool has the potential to assist healthcare providers in improving patient outcomes.
• It can be particularly valuable in settings with limited access to specialized dermatology resources.