Abstract
INTRODUCTION: Atopic dermatitis (AD) presents significant diagnostic challenges in darker skin types due to the masking of classical signs like erythema by pigmentation and the lack of diagnostic criteria tailored for this population. The use of machine learning (ML) as a noninvasive diagnostic tool in this context remains underexplored. OBJECTIVES: This study aimed to develop and evaluate the accuracy of an ML model in diagnosing AD in darker skin types and distinguishing it from healthy skin. METHODS: A cross-sectional study using images of AD cases in dark-skinned patients was conducted. An EfficientNet-based convolutional neural network (CNN) model was trained on 1403 labelled images (687 atopic and 716 healthy skin images). Performance metrics including accuracy, precision, recall, and F1 score were assessed. Ethical approval was obtained, and datasets were anonymized for confidentiality. RESULTS: The dataset comprised images from 50 AD patients (age range 1-93 years, median 18.0, IQR: 7.0-32.0) and 60 healthy volunteers (age range 18-50 years, median 27.5, IQR: 23.0-33.5), with a slight female predominance in both groups (56.0% and 53.3%, respectively). The model achieved an accuracy of 91.1%, precision of 83.7%, recall of 97.0%, and an F1 score of 89.9%, correctly classifying 103/106 AD images and 129/149 healthy skin images. External application programming interfaces (APIs) were integrated into the design to further enhance accuracy and generalizability. CONCLUSION: This study highlights the potential of ML as a diagnostic tool for AD in darker skin types. Further optimization and large-scale validation could enhance diagnostic accuracy, improving clinical outcomes for populations with darker skin. Iterative refinement and user feedback will ensure its continued efficacy and reliability in clinical practice.