GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs

TL;DR

GraphCare tackles the challenge of personalized healthcare predictions by generating patient-specific knowledge graphs from external knowledge sources, including LLM-derived triples and established biomedical KGs, and by processing them with a Bi-attention AugmenTed (BAT) GNN. The method first creates concept-level KGs, then composes a patient-level KG with temporal visit data, and finally applies BAT to produce robust predictions across mortality, readmission, LOS, and drug recommendation tasks. Empirical results on MIMIC-III and MIMIC-IV show clear gains over strong baselines, with particular strength in data-scarce settings and clear interpretability via KG-based importance scores. The work highlights the practical potential of integrating open-world knowledge into clinical prediction pipelines while addressing ethical considerations and privacy concerns.

Abstract

Clinical predictive models often rely on patients' electronic health records (EHR), but integrating medical knowledge to enhance predictions and decision-making is challenging. This is because personalized predictions require personalized knowledge graphs (KGs), which are difficult to generate from patient EHR data. To address this, we propose \textsc{GraphCare}, an open-world framework that uses external KGs to improve EHR-based predictions. Our method extracts knowledge from large language models (LLMs) and external biomedical KGs to build patient-specific KGs, which are then used to train our proposed Bi-attention AugmenTed (BAT) graph neural network (GNN) for healthcare predictions. On two public datasets, MIMIC-III and MIMIC-IV, \textsc{GraphCare} surpasses baselines in four vital healthcare prediction tasks: mortality, readmission, length of stay (LOS), and drug recommendation. On MIMIC-III, it boosts AUROC by 17.6\% and 6.6\% for mortality and readmission, and F1-score by 7.9\% and 10.8\% for LOS and drug recommendation, respectively. Notably, \textsc{GraphCare} demonstrates a substantial edge in scenarios with limited data availability. Our findings highlight the potential of using external KGs in healthcare prediction tasks and demonstrate the promise of \textsc{GraphCare} in generating personalized KGs for promoting personalized medicine.

Figures (10)

• Figure 1: Overview of GraphCare. Above: Given a patient record consisting of conditions, procedures and medications, we generate a concept-specific KG for each medical concept, by knowledge probing from a LLM and subgraph sampling from an existing KG; and we perform node and edge clustering among all graphs (§\ref{['subsec:ehr_graph']}). Below: For each patient, we compose a patient-specific graph by combining the concept-specific KGs associated with them and make the graph temporal with sequential data across patient's visits (§\ref{['subsec:personal_kg']}). To utilize the patient graph for predictions, we employ a bi-attention augmented graph neural network (GNN) model, which highlights essential visits and nodes with attention weights (§\ref{['subsec:bi_att_gnn']}). With three types of patient representations (patient-node, patient-graph, and joint embeddings), GraphCare is capable of handling a variety of healthcare predictions (§\ref{['subsec:health_pred']}).
• Figure 2: Performance by EHR training data sizes. Values on the x-axis indicate $\%$ of the entire training data. The dotted lines separate three ranges: [0.1, 1], [1, 10] and [10, 100] (%).
• Figure 3: Performance by different KG sizes. We test on three distinct KGs: GPT-KG, UMLS-KG, and GPT-UMLS-KG. For each, we sample sub-KGs using varying ratios: $[0.0, 0.1, 0.3, 0.5,$$0.7, 0.9, 1.0]$ while ensuring the nodes corresponding to EHR medical concepts remain consistent across samples. The distributions are based on 30 runs on the MIMIC-III with different random seeds.
• Figure 4: Example showing a patient's personalized KG with importance scores (Appendix \ref{['ap:importance_score']}) visualized. For better presentation, we hide the nodes of drugs. The red node represents the patient node. Nodes with higher scores are darker and larger. Edges with higher scores are darker and thicker. Subgraph (a) shows a central area overview of this personalized KG, and other subgraphs show more details with a focused node highlighted.
• Figure 5: High-level View of GraphCare for Clarification on Ethical Considerations.GraphCare consists of two general stages: data preparation and local model training. During data preparation, the LLM solely extracts knowledge graphs associated with medical concepts, without accessing any patient's data. At the local model training stage, personalized knowledge graphs for patients are constructed using the knowledge graphs corresponding to medical concepts found in the patient's EHR, without any engagement of the LLM. A local graph storage serves as both the repository for the procured medical concept-wise KGs and the mechanism for querying KGs for personalized KG compositions.