Mobile-accessible deep learning-based self-assessment tool for measles screening in low-resource settings

Objective: Measles remains a major global health challenge, particularly in low-resource settings where undervaccination and limited diagnostic tools hinder early detection and response. This study aimed to develop a mobile-accessible, artificial intelligence-based self-assessment tool for measles screening in such settings.

Methods and analysis: A dataset of 461 measles and 44 050 non-measles skin lesion images was constructed from journal articles, encyclopaedias, news articles, social media and eight datasets. Images were manually annotated by age group, gender, origin, skin tone, body region and rash colour. A deep learning model was trained and validated across these characteristics, with external validation on four out-of-distribution datasets. The self-assessment tool integrates this model with a symptom-based questionnaire covering clinical features, exposure history and immunity status to generate a risk level and score. Risk scores were calculated using an XGBoost classifier trained on real-world clinical series data.

Results: The deep learning model achieved 93.5% accuracy, 93.6% precision and 93.5% sensitivity in detecting measles skin lesions. It demonstrated robust performance across all annotated subgroups, with true negative rates (TNRs) ranging from 79.0% to 92.9% in external datasets. Lower true positive rates were observed for images with Fitzpatrick type IV skin tones (77.8%) and lesions on the upper extremities (80.0%). Lower TNR was noted in children under 5 years of age (67.2%) and for lesions on the torso (68.0%) and neck (71.7%). The tool stratifies measles risk into four levels using a decision-tree framework with follow-up recommendations. The risk score algorithm achieved 91.9% accuracy, 86.4% sensitivity and 95.5% specificity.

Conclusion: Our self-assessment tool offers a scalable, cost-effective approach for early measles screening and outbreak management in low-resource settings. Expanding the dataset to include more images from diverse skin tones and regions could further enhance its accuracy and applicability.

© Author(s) (or their employer(s)) 2026.