A multimodal deep learning framework for automated stunting detection using visual and anthropometric data

Background Early childhood stunting remains a significant global public health issue, causing long-term adverse impacts on physical growth, cognitive development, and overall productivity. Conventional anthropometric screening methods often suffer from measurement errors, limited scalability, and inconsistent accessibility. Objective This study aims to develop an automated and scalable stunting detection framework by integrating visual and anthropometric data using deep learning techniques. Methods A multimodal deep learning architecture was proposed, combining frontal full-body child images and anthropometric data to predict stunting status. The ARAN dataset was utilized, with labels determined based on the World Health Organization (WHO) Height-for-Age Zscores (HAZ). Visual representations were extracted using a pre-trained ResNet50 network and fused with numerical anthropometric inputs. To address class imbalance, the RandomOverSampler method was applied during model training. Results The proposed model achieved an accuracy of 99.66%, precision of 0.9405, recall of 0.9753, and an F1-score of 0.96 on the validation set. The model exhibited a notably low false negative rate, indicating strong reliability for screening applications. Conclusion The findings highlight the potential of multimodal data fusion to enhance the precision and scalability of stunting detection systems. This work establishes a foundation for AI-assisted health monitoring and equitable early childhood growth assessment.