Educational Artificial Intelligence Software to Support Assessment of Atopic Dermatitis Severity

Area of Science:

• Dermatology
• Artificial Intelligence
• Medical Imaging

Background:

• Accurate atopic dermatitis severity assessment is vital for effective treatment.
• The Eczema Area and Severity Index (EASI) is widely used but suffers from subjective scoring and inter-rater variability.
• Objective assessment tools are needed to improve diagnostic consistency.

Purpose of the Study:

• To evaluate the performance of a convolutional neural network (CNN)-based AI software in assessing atopic dermatitis severity.
• To analyze the AI's ability to interpret real-world, uncontrolled smartphone-captured skin images.
• To compare AI-driven EASI sign scoring against dermatologist assessments.

Main Methods:

• A prospective, multicenter observational study involving 1016 patients with atopic dermatitis across 16 Japanese institutions.
• Collection of 3676 smartphone-captured skin images under uncontrolled conditions.
• Training a ConvNeXt-based AI model to predict EASI component scores (erythema, papulation/edema, excoriation, lichenification) and evaluating performance using receiver operating characteristic (ROC) analysis.

Main Results:

• The AI model demonstrated robust performance, with area under the ROC curve (AUC) ranging from 0.825 to 0.860 for EASI component scores ≥2.
• For a component severity score of 3, the AUC exceeded 0.900, with high sensitivity and specificity.
• Erythema assessment showed the highest accuracy, while papulation/edema had the lowest; overall performance for scores ≥1 remained consistently >0.700.

Conclusions:

• The ConvNeXt-based AI software shows significant potential for evaluating atopic dermatitis severity from real-world images.
• This AI tool may serve as an educational aid to standardize EASI sign scoring and reduce inter-rater variability.
• The software is currently intended for educational support, not direct clinical decision-making.