Bayesian Hypergraph Framework Uncovers Hidden Disease Pathways

Researchers Shengxian Ding, Haonan Gao, Pangpang Liu, Xinyuan Tian, and Yize Zhao propose a Bayesian hypergraph inference framework that reframes multi-disease modeling around latent, risk-factor-modulated disease pathways. Unlike standard approaches that treat diseases independently or rely on opaque deep learning, the method models risk factors as acting on hyperedges—higher-order connections linking multiple diseases that share common risk patterns. A repulsion prior encourages sparse, identifiable structure, and full posterior inference provides principled uncertainty quantification over both disease groupings and the influence of each risk factor. To handle the scale of real-world EHR datasets like UK Biobank, the authors develop a structured variational inference algorithm that preserves logical dependencies among hyperedge existence, disease membership, and pathway-level effects.

Experiments on simulated data and the UK Biobank demonstrate that the framework recovers stable and interpretable disease pathway structures, outperforming baseline methods in identifying rare disease associations while maintaining competitive predictive accuracy. The model not only reveals known risk-factor-disease links (e.g., smoking and respiratory diseases) but also discovers novel higher-order groupings that would be missed by pairwise graph models. With well-calibrated uncertainty estimates, clinicians and researchers can trust which pathways are robust. This work, accepted as an oral at ICML 2026, offers a new tool for personalized risk assessment and hypothesis generation in epidemiology.