Scratching failure off the list, Korean researchers develop AI to classify skin irritations with 98% accuracy

South Korean researchers have developed an AI model using YOLO v5 technology that can accurately classify skin irritation from patch tests, achieving a classification accuracy of 0.983. The breakthrough addresses longstanding challenges in dermatological diagnostics by providing consistent, objective assessments that could reduce variability between human evaluators and accelerate clinical decision-making.

What you should know: The AI model represents a significant advancement in automated dermatological assessment, moving beyond traditional convolutional neural networks to object detection algorithms.

• Researchers from Sungkyunkwan University, a South Korean research institution, trained the model on 83,629 images collected from patch test participants between 2020-2023.
• The system achieved an F1 score of 0.982, indicating strong balance between precision and recall.
• Individual grade performance showed areas under the curve of 0.914 for score 0, 0.838 for score 1, and 0.865 for score 2.

How it works: The YOLO v5 (You Only Look Once version 5) algorithm analyzes standardized clinical photographs to classify skin reactions on a 5-point scale.

• Patch tests were applied to participants’ upper backs using Van der Bend chambers for 24 hours, with skin responses evaluated 1 hour and 24 hours after removal.
• All images were captured at a fixed resolution of 4160 × 2768 pixels to detect subtle reddening changes.
• Four independent evaluators assessed each image, with a senior reviewer providing final scores in cases of disagreement.

In plain English: Think of this AI like a highly trained dermatologist that never gets tired or has off days. It looks at photos of skin patch tests (small areas where potential allergens are applied to test for reactions) and determines how irritated the skin is on a scale from 0 (no reaction) to 4 (severe reaction with swelling and blisters). The AI was trained by showing it over 83,000 photos along with expert opinions on what each reaction should be rated.

Why this matters: Traditional patch test evaluation suffers from significant inter-rater variability, which this AI model could help eliminate.

• The model demonstrated particularly high sensitivity for detecting score 0 reactions at 0.997, crucial for identifying non-reactive cases.
• Automated classification could enable faster clinical assessments and more consistent diagnostic outcomes across different healthcare settings.

What they’re saying: The research team emphasized the model’s potential to transform dermatological practice through enhanced objectivity.

• “The AI-based erythema reading model developed in this study demonstrates significant potential to enhance the efficiency of evaluations while minimising inter-rater variability, thereby enabling more objective and consistent assessments,” the researchers wrote.
• “Moreover, the integration of the proposed future improvements is expected to further increase the accuracy and reliability of patch test reaction grading, ultimately broadening the model’s applicability across a variety of clinical environments.”

The big picture: This development represents part of a broader trend toward AI-assisted diagnostic tools in dermatology, where machine learning algorithms are increasingly being deployed to automate skin reaction detection and improve clinical workflows.