Computational approaches to multimodal data integration in rheumatoid arthritis: from data landscape to clinical translation

Multimodal data integration approaches represent a breakthrough advancement in artificial intelligence applications for rheumatoid arthritis (RA) research, enabling comprehensive analysis of heterogeneous, high-dimensional datasets spanning genomics, transcriptomics, proteomics, medical imaging and longitudinal electronic health records (EHRs). This comprehensive review systematically examines the development and application of computational fusion strategies in RA-early, late and intermediate fusion, as well as graph-based integration-across key tasks including early diagnosis, disease activity quantification, treatment response prediction, patient stratification and drug repurposing. We highlight the distinctive capabilities of these frameworks in end-to-end representation learning, cross-modal knowledge transfer and interpretable decision support. Furthermore, we thoroughly discuss major challenges-data heterogeneity and harmonization, structured missingness, model interpretability, computational scalability and reproducibility-and explore emerging solutions such as adaptive fusion architectures, federated and privacy-preserving learning, self-supervised pretraining and model compression. In conclusion, this paper aims to objectively evaluate the current landscape of multimodal computational methods in RA, delineate their advantages and limitations, and offer strategic recommendations to guide future research toward clinically translatable precision rheumatology.