A Mobile-Accessible Cost-Effective AI-Aided Self-Assessment Tool for Early Measles Detection in Low-Resource Settings
Author(s)
LIU MING, MSc1, Xinyao YI, MSc1, Wai-kit Ming, MPH, PhD, MD1, Yun-Zhe Chen, MPH2.
1Department of Infectious Diseases and Public Health, JCC of Veterinary Medicine and Life Sciences, City University of Hong Kong, Hong Kong, Hong Kong, 2School of Public Health, LKS Faculty of Medicine, The University of Hong Kong, Hong Kong, Hong Kong.
1Department of Infectious Diseases and Public Health, JCC of Veterinary Medicine and Life Sciences, City University of Hong Kong, Hong Kong, Hong Kong, 2School of Public Health, LKS Faculty of Medicine, The University of Hong Kong, Hong Kong, Hong Kong.
OBJECTIVES: Measles remains a significant global health concern, particularly in low-resource settings where under-vaccination and limited diagnostic capacities hinder early detection. Rapid and accessible identification of measles cases is critical for timely intervention and outbreak control. This study aims to develop a mobile-accessible, deep learning-based self-assessment tool to facilitate early measles detection.
METHODS: A dataset of 554 measles and 41,741 non-measles skin lesion images was compiled from diverse sources, including journal articles, encyclopedias, news articles, social media, and eight databases. Each image was manually annotated for characteristics such as age, gender, origin, skin tone, and body region. A convolutional neural network (CNN) was trained on this dataset and validated across different image characteristics, with additional external validation using four out-of-distribution datasets. The self-assessment tool integrates this model with a structured questionnaire covering symptoms, exposure history, and immunity status.
RESULTS: The model achieved 98.0% accuracy, 99.6% precision, and 96.4% sensitivity, outperforming state-of-the-art models in measles lesion classification. It maintained high true negative rates (TNR) (80.7%-94.2%) across external datasets but exhibited lower true positive rates (TPR) for images of African origins (66.7%), darker skin tones (Fitzpatrick type V: 60.0%), and lesions on the lower extremities (50.0%). Lower TNR was also observed for conditions resembling measles, such as drug eruptions (68.3%) and urticaria hives (73.7%). To enhance diagnostic reliability, the tool employs a decision tree with four risk levels, rash characteristic screening to minimize false positives, and a follow-up questionnaire to reduce false negatives.
CONCLUSIONS: This self-assessment tool has the potential to improve early measles detection and outbreak response, particularly in low-resource settings. Future improvements include expanding the dataset with more diverse skin tones and regional variations to enhance accuracy and applicability.
METHODS: A dataset of 554 measles and 41,741 non-measles skin lesion images was compiled from diverse sources, including journal articles, encyclopedias, news articles, social media, and eight databases. Each image was manually annotated for characteristics such as age, gender, origin, skin tone, and body region. A convolutional neural network (CNN) was trained on this dataset and validated across different image characteristics, with additional external validation using four out-of-distribution datasets. The self-assessment tool integrates this model with a structured questionnaire covering symptoms, exposure history, and immunity status.
RESULTS: The model achieved 98.0% accuracy, 99.6% precision, and 96.4% sensitivity, outperforming state-of-the-art models in measles lesion classification. It maintained high true negative rates (TNR) (80.7%-94.2%) across external datasets but exhibited lower true positive rates (TPR) for images of African origins (66.7%), darker skin tones (Fitzpatrick type V: 60.0%), and lesions on the lower extremities (50.0%). Lower TNR was also observed for conditions resembling measles, such as drug eruptions (68.3%) and urticaria hives (73.7%). To enhance diagnostic reliability, the tool employs a decision tree with four risk levels, rash characteristic screening to minimize false positives, and a follow-up questionnaire to reduce false negatives.
CONCLUSIONS: This self-assessment tool has the potential to improve early measles detection and outbreak response, particularly in low-resource settings. Future improvements include expanding the dataset with more diverse skin tones and regional variations to enhance accuracy and applicability.
Conference/Value in Health Info
2025-09, ISPOR Real-World Evidence Summit 2025, Tokyo, Japan
Value in Health Regional, Volume 49S (September 2025)
Code
RWD199
Topic Subcategory
Distributed Data & Research Networks
Disease
STA: Personalized & Precision Medicine