Biomedical knowledge graphs (KGs) have been proposed to bring biological knowledge into machine-learning models trained on electronic health record (EHR) data. The large scale and heterogeneity of many KGs create substantial computational burdens, and it is unclear whether KG benefits generalize across diseases or depend on graph construction. This study systematically evaluated whether and how KG integration changes predictive performance, computational cost, and model behavior across three clinically and biologically distinct conditions: Parkinson’s disease, multiple sclerosis, and chronic pain.
The authors aimed to answer three linked questions: whether graph construction affects performance and cost; whether KG integration improves classification relative to EHR-only models; and what mechanistic changes occur in model behavior after incorporating KG-derived features.
For each disease the investigators compared five modelling strategies. One baseline used EHR-only features. The other four incorporated graph diffusion over different SPOKE-derived subgraphs: a concept-level subgraph, a cohort-derived subgraph, a disease-anchored subgraph, and the full SPOKE graph (Scalable Precision Medicine Open Knowledge Engine).
Graph diffusion was used to propagate patient-level signal across the chosen graph topology, producing transformed feature representations that were then used in classification tasks. The comparison aimed to reveal whether the way the graph is built and the scale of nodes and edges meaningfully affect predictive accuracy or computational requirements.
The study evaluated multiple subgraph construction approaches to balance biological coverage and computational tractability. Notably, cohort-derived subgraphs were built from the SPOKE graph using nodes and edges relevant to the patient cohorts; these subgraphs comprised approximately 45% of SPOKE’s nodes and about one-fifth of its edges. The full SPOKE graph remained the largest reference for comparing biological breadth.
Other subgraph types included concept-level selections and disease-anchored subsets, each representing different strategies for limiting graph scope while preserving relevant biological relationships.
Across all three diseases the cohort-derived subgraphs substantially reduced computational cost relative to the full SPOKE graph. Despite their smaller size, cohort-derived subgraphs matched or modestly exceeded the full graph’s predictive performance in the classification tasks examined. This result indicates that targeted, cohort-relevant subgraphs can preserve or improve accuracy while offering large efficiency gains.
The study therefore positions cohort-derived subgraphs as a practical compromise when integrating KG-derived features into EHR-based models, since they capture a large fraction of relevant graph content without the full-scale computational overhead.
KG integration did not have a uniform effect across conditions. The impact varied by disease:
In Parkinson’s disease, incorporating KG-derived features improved classification performance compared to EHR-only models.
In multiple sclerosis, KG integration produced no measurable change in predictive accuracy relative to EHR-only models.
In chronic pain, adding KG-derived features reduced predictive performance.
These differences show that KG incorporation is not a universal accuracy enhancer; its effect depends on disease-specific data characteristics and how the graph transforms the patient representation.
Analyses of how the graph changed model behavior revealed a common mechanism: graph diffusion tended to redistribute each disease’s existing predictive signal across many correlated features rather than injecting independent biological signal. In other words, diffusion spread and reweighted information already present in the EHR rather than unearthing novel, orthogonal predictors.
Because the graph mainly redistributes existing signal, whether KG integration helps depends on whether that redistribution yields additional information beyond the original EHR features. The authors observed that the magnitude of molecular or biological information incorporated did not predict benefit: the disease with the largest shift toward molecular features showed no improvement, the only disease that benefited exhibited only a modest molecular shift, and the disease with no molecular shift was harmed.
The study draws two principal practical conclusions. First, cohort-derived subgraphs represent a computationally efficient alternative to full-scale knowledge graphs: they substantially reduce nodes and edges while preserving or modestly improving predictive performance in these tasks. Second, integrating KGs functions as a data- and disease-dependent transformation of patient representations rather than a guaranteed method to increase accuracy across settings.
Practically, these findings suggest that teams planning to augment EHR-based prediction with KGs should evaluate KG benefits on a per-disease basis and consider constructing cohort-relevant subgraphs to reduce computational cost. They should also examine whether graph diffusion changes model behavior by redistributing rather than adding signal, and validate that such redistribution supplies new information that meaningfully improves classification for the target condition.
The authors declare no competing interests. The primary source for these findings is the preprint reporting the systematic evaluation across Parkinson’s disease, multiple sclerosis, and chronic pain conducted using SPOKE-derived subgraphs and graph diffusion techniques.