Artificial intelligence-based mobile application for instant acne severity assessment among adolescents and young adults in Kampala and Wakiso districts, Uganda

Abstract

Background: Acne vulgaris is a common dermatological condition among adolescents and young adults in Uganda. Limited access to dermatologists contributes to delayed assessment, inappropriate self-treatment, post-inflammatory hyperpigmentation, scarring, and psychosocial distress. Existing AI dermatology tools may also underperform on darker skin tones because many training datasets underrepresent Fitzpatrick skin types V–VI. Objective: This study aimed to develop and evaluate an AI-based mobile application for instant acne severity assessment among adolescents and young adults in Kampala and Wakiso districts, Uganda, with emphasis on usability, offline functionality, privacy, and performance across darker skin tones. Methods: The study used a methodological design combining a cross-sectional user needs survey, AI model development, mobile application prototyping, and observational diagnostic accuracy evaluation. Survey data were collected from 200 participants, while dermatologist/key informant input guided image annotation, clinical relevance, and safety requirements. A MobileNetV2 convolutional neural network was trained to classify cropped skin-region images into Clear, Mild Acne, Moderate Acne, Severe Acne, and Non-Skin categories. Model performance was evaluated using accuracy, precision, recall, F1-score, and confusion matrix analysis. Usability was assessed using the System Usability Scale. Results: The final MobileNetV2 model achieved [insert final accuracy] accuracy, 85% precision, 86% recall, and 85% F1-score. The mobile prototype supported image capture/upload, preprocessing, offline TensorFlow Lite inference, confidence score display, non-diagnostic recommendations, privacy controls, and prediction history. Usability testing with 30 participants produced a SUS score of 75.3, indicating above-average usability. Users and experts valued offline use, simplicity, privacy, and referral guidance for severe or uncertain cases. Conclusion: The study demonstrates the potential of an offline-capable AI mobile application to support acne severity assessment in a low-resource Ugandan context. The application should be used as a decision-support and self-monitoring tool rather than a replacement for dermatologist diagnoses. Future work should expand the dataset, conduct formal control and ablation experiments, test against acne-like skin conditions, and undertake larger dermatologist-led clinical validation.