AI framework reveals psychosocial bridge from disadvantage to cardiometabolic multimorbidity

Researchers Cong Cao and Shuangge Ma have introduced an AI-driven multimodal mediation framework that uncovers hidden pathways linking social disadvantage to cardiometabolic multimorbidity. Published on arXiv, the study integrates data from the NIH's All of Us Research Program across six domains: socioeconomic, psychosocial, clinical, laboratory, behavioral, and genomic. Using modality-specific variational autoencoders, the system compresses each data type into latent representations, then performs mediation analysis in that compressed space. This approach bypasses the noise and complexity of high-dimensional real-world health data, allowing the model to identify which intermediary factors truly carry the effect of disadvantage onto disease burden.

The final cohort included 20,804 participants with complete multimodal data. Across 800 exposure-mediator-outcome combinations, mediation signals clustered in just a few latent dimensions. The strongest indirect effect linked a socioeconomic disadvantage dimension to a cardiometabolic multimorbidity dimension via a psychosocial vulnerability dimension, with an NIE of 0.002517. That psychosocial profile reflected poorer mental health, loneliness, low social well-being, and low health literacy, while the outcome dimension was associated with hypertension, diabetes, hyperlipidemia, obesity, chronic kidney disease, and heart disease. Bootstrap analyses confirmed the pathway's stability. The framework demonstrates a scalable, generalizable way to interrogate complex relationships in multimodal biomedical datasets, pointing toward more targeted interventions for social determinants of health.