Educational Artificial Intelligence Software to Support Assessment of Atopic Dermatitis Severity

Accurate assessment of the severity of atopic dermatitis is crucial for guiding treatment. However, the Eczema Area and Severity Index (EASI), a widely used clinical assessment tool, relies on subjective visual scoring, which can lead to inter-rater variability. This study aimed to evaluate the performance of convolutional neural network-based artificial intelligence software in assessing atopic dermatitis severity from uncropped, unmasked skin images acquired under uncontrolled conditions. In this prospective, multicenter observational study involving 16 Japanese institutions, 1016 patients with atopic dermatitis provided 3676 smartphone-captured skin images. A ConvNeXt-based artificial intelligence model was trained to predict the severity scores for four established signs of the EASI-erythema, papulation/edema, excoriation, and lichenification-and these were evaluated using receiver operating characteristic analysis against dermatologist assessments. For an EASI component score of ≥ 2, the area under the receiver operating characteristic curve ranged from 0.825 to 0.860 across signs, with sensitivities of 0.745-0.873 and specificities of 0.675-0.764. For a component severity score of 3, the area under the receiver operating characteristic curve exceeded 0.900, with sensitivity up to 1.000 and specificity up to 0.906. The area under the receiver operating characteristic curve for a severity score of ≥ 1 remained consistently > 0.700. Among the four evaluated signs, erythema demonstrated the highest accuracy, whereas papulation/edema showed the lowest. This ConvNeXt-based artificial intelligence software demonstrated robust performance in evaluating atopic dermatitis severity using real-world images. As an educational tool, it may support standardized training and reduce inter-rater variability in EASI sign scoring. It is not intended for direct clinical decision-making at this stage.