Understanding when and how knowledge graphs add value to EHR-based clinical prediction across Parkinson's disease, multiple sclerosis, and chronic pain

In a study examining the value of knowledge graphs (KGs) for enhancing machine-learning models trained on electronic health records (EHR) data, researchers systematically evaluated KG integration across three distinct diseases: multiple sclerosis, Parkinson's disease, and chronic pain. The study compared five modelling strategies, ranging from EHR-only features to graph diffusion over various subgraphs, including concept-level, cohort-derived, disease-anchored, and full SPOKE subgraphs.

The researchers found that cohort-derived subgraphs, accounting for roughly 45% of SPOKE's nodes and one-fifth of its edges, either matched or slightly surpassed the full graph's performance while significantly reducing computational cost across all three diseases. However, the predictive benefit of KG integration varied significantly depending on the specific disease. KG integration improved classification in Parkinson's disease but had no measurable effect in multiple sclerosis and reduced performance in chronic pain.

Notably, the predictive benefit of KG integration was not correlated with the amount of biological information incorporated into the subgraphs. The disease with the largest shift toward molecular features did not show any improvement, while the sole beneficiary of KG integration showed only a modest shift, and the disease with no molecular shift was penalized.

This suggests that KG integration functions less as a universal accuracy booster and more as a data- and disease-dependent transformation of the patient representation. The study highlights cohort-derived subgraphs as a computationally efficient alternative to full-scale knowledge graphs and emphasizes the need for disease-specific KG integration strategies.