Multimodal Pneumoconiosis Screening on a Real-World Vietnamese Occupational Health Dataset with Incomplete Metadata and Class Imbalance

Pneumoconiosis remains a major occupational lung disease among workers exposed to silica, coal, and other in-organic dusts. Although chest X-ray screening is widely used in clinical practice, diagnostic performance is often affected by image variability and the limited availability of contextual occupational information. This study presents a multimodal framework that combines chest X-ray images with clinical and occupational metadata for automated pneumoconiosis detection. The proposed approach integrates a ResNet-18 image encoder with a metadata-processing multilayer perceptron through an early-fusion strategy. Experiments were conducted on a dataset of 1,971 subjects, including 1,086 confirmed pneumoconiosis cases and 885 healthy controls. Two configurations were evaluated: an image-only model and a multimodal model incorporating non-imaging information. The multimodal framework achieved the best validation performance, reaching an accuracy of 92.66%, an F1-score of 93.21%, and an AUC-ROC of 97.35%. Compared with the image-only baseline, adding metadata produced a small change in overall performance and shifted the precision–recall balance toward higher precision. These validation figures should be interpreted as an upper bound because the negative class is partly drawn from an external population and the pattern of missing metadata is correlated with the diagnostic label. The results suggest that combining radiographic findings with occupational and clinical information is a promising but data-quality-sensitive direction for pneumoconiosis screening in real-world settings.

@article{Phuong2026,
  title     = {Multimodal Pneumoconiosis Screening on a Real-World Vietnamese Occupational Health Dataset with Incomplete Metadata and Class Imbalance},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {7},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Luong Thi Bich Phuong and Nguyen Hoang Anh and Au Xuan Manh and Truong Thanh Nam and Tran Tien Cong and Nguyen Trong Khanh},
  doi       = {10.14569/IJACSA.2026.0170777},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170777}
}