AI-driven Multimodal Representation Learning for Latent Mediation Structure Discovery of Socioeconomic Disadvantage, Psychosocial Factors, and Cardiometabolic Multimorbidity: Insights from the All of Us Research Program

A multimodal mediation framework integrates socioeconomic, psychosocial, clinical, genomic, and behavioral data from 20,804 participants to map disease pathways. Using modality-specific variational autoencoders, the model performs mediation analysis in latent space, concentrating indirect associations between social disadvantage and multimorbidity within a small subset of latent dimensions.