Understanding when and how knowledge graphs add value to EHR-based clinical prediction across Parkinson's disease, multiple sclerosis, and chronic pain
Biomedical knowledge graphs (KGs) have emerged as a promising approach for incorporating biological knowledge into machine-learning models trained on electronic health record (EHR) data. However, their large scale and heterogeneous structure impose substantial computational cost, and it remains unclear whether their benefits generalize across diseases or depend on how the graph is constructed. We systematically evaluated KG integration across three diseases spanning distinct biological and clinical characteristics: multiple sclerosis, Parkinson's disease, and chronic pain. For each disease we compared five modelling strategies, spanning EHR-only features and graph diffusion over concept-level, cohort-derived, disease-anchored, and full SPOKE (Scalable Precision Medicine Open Knowledge Engine) subgraphs, asking whether graph construction affects predictive performance and computational cost, whether KG integration improves classification over EHR-only models, and what the graph does to model behavior. Cohort-derived subgraphs, comprising roughly 45% of SPOKE's nodes and one-fifth of its edges, matched or modestly exceeded the full graph's performance while substantially reducing computational cost in all three diseases. The predictive benefit of KG integration, however, was strongly disease-dependent: it improved classification in Parkinson's disease, had no measurable effect in multiple sclerosis, and reduced performance in chronic pain. These differences were not explained by the amount of biological information incorporated: the disease with the largest shift toward molecular features showed no improvement, the sole beneficiary showed only a modest shift, and the disease with no molecular shift was penalized. Mechanistically, graph diffusion redistributed each disease's existing signal across many correlated features rather than adding independent biological signal, so that any gain depended on whether it added information beyond the EHR. These findings show that cohort-derived subgraphs offer a computationally efficient alternative to full-scale knowledge graphs, and that KG integration functions less as a universal accuracy booster than as a data- and disease-dependent transformation of the patient representation.