Artificial Intelligence Model Effective in Classification of Skin Erythema

A recently developed artificial intelligence (AI) model known as YOLO v5 is effective in classifying skin irritation following patch tests, new data suggest, allowing for swift and accurate evaluations.¹

The data regarding this new AI model for classification of skin erythema from patch tests were authored by investigators such as Seoyoung Kim, from the Department of Integrative Biotechnology at Sungkyunkwan University in South Korea. Kim et al noted that recent studies have examined AI’s implementation as an assistive diagnostic tool in dermatology.

Advances in machine learning (ML) and deep learning (DL) have also led to certain studies highlighting the feasibility of utilizing ML to automate skin reaction detection.²

“These investigations highlight the revolutionary potential of AI-powered diagnostics in dermatology, particularly for improving the accuracy, consistency and accessibility of epicutaneous patch testing,” Kim and coauthors wrote.¹ “In this study, we developed a model that overcomes the limitations of the convolutional neural network (CNN) to classify skin irritation from images using the object detection algorithm YOLO v5 (You Only Looking Once version 5).”

Trial Design Details

The investigative team gathered patch test data from trial participants between 2020 – 2023, assessing patients who had been recruited in Seoul and the Gyeonggi Province who had also voluntarily enrolled in their analysis. Individuals deemed eligible for enrollment in the trial were adults ≥18 years who were shown to have healthy skin and were able to provide written informed consent and commit to follow-up visits as well as all other study requirements.

The patch testing required for Kim and colleagues’ research was carried out using Van der Bend chambers, which consisted of hypoallergenic, medical-grade non-woven polyester and contained chromatographic filter paper. The investigators would apply cosmetics or cosmetic ingredients to the upper back of study subjects and leave these in place for 24 hours. Participants’ skin responses were then evaluated 1 hour and 24 hours after removal. Reactions were graded on a 5-point scale and were adapted from the Frosch and Kligman method and CTFA safety guidelines:

0: A lack of visible reaction

1: Mild erythema, either scattered or diffuse

2: Moderate, uniform erythema

3: Pronounced erythema with edema

4: Pronounced erythema with edema and vesicles

The team, for the purposes of documenting responses, gathered standardized clinical photographs under controlled conditions, regardless of whether a reaction occurred. All images were captured by Kim and colleagues at a fixed resolution of 4160 × 2768 pixels to allow for the detection of any subtle erythematous shifts. Preprocessing of images was performed to prepare them for analysis with the YOLOv5x deep learning algorithm.

Each of the collected images was assessed independently by 4 evaluators. In cases of discordance, the investigators instructed a senior reviewer to provide the final score. Annotations were later mapped to corresponding bounding boxes within the dataset.

Findings on AI Model for Skin Irritation

Kim and coauthors’ training dataset was made up of 83,629 images, with 1,312 and 1,536 images being utilized for evaluation and validation, respectively. Overall, the AI model attained a classification accuracy of 0.983 at both the 24 and 48-hour mark.¹ The model’s F1 score, indicating a balance between precision and recall, was shown to be 0.982.

In terms of individual grades, the areas under the curve (AUCs) were shown to be 0.914 for score 0, 0.838 for score 1, and 0.865 for score 2. Sensitivity for detecting score 0 was particularly high at 0.997. Such conclusions would indicate that this AI model supports and classifies skin irritation effectively, thus allowing for quicker and more accurate skin assessments.

“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,” they wrote.¹ “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.”

References

1. Kim S, Hwang H, Oh M, Han J, Park S, Lee S, Kim G, Cho S, Lee DH, Cho JY. Evaluation of Artificial Intelligence-Assisted Diagnosis of Skin Erythema in a Patch Test. Contact Dermatitis. 2025 Aug 27. doi: 10.1111/cod.70011. Epub ahead of print. PMID: 40859876.

2. WH Chan, R Srivastava, N Damaraju, et al. “Automated Detection of Skin Reactions in Epicutaneous Patch Testing Using Machine Learning,” British Journal of Dermatology 185, no. 2 (2021): 456–458.