Research map / Graph learning / Knowledge and temporal graphs
GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds patient-specific knowledge graphs from GPT-4-prompted triples and UMLS subgraph sampling, then trains a Bi-attention Augmented GNN (BAT) for EHR predictions on MIMIC-III/IV.
Problem
EHR-based clinical prediction models lack personalized, comprehensive medical knowledge; existing clinical KGs mainly encode simple hierarchies rather than rich biomedical relations.
Approach
Generate concept-specific KGs by prompting an LLM (GPT-4) and sampling subgraphs from UMLS, cluster nodes/edges, compose per-patient visit-linked KGs, and apply a BAT GNN with node-level and visit-level attention, edge weights, and attention initialization.
Key results
- On MIMIC-III, GraphCare with BAT improves mortality AUROC to 70.3 vs 61.5 for best baseline (StageNet), roughly +17.6% relative gain as claimed in abstract.
- GraphCare with BAT reaches LOS F1 37.5 on MIMIC-III and drug recommendation F1 66.8, exceeding the best baselines.
- With only 0.1% of training data, GraphCare matches LOS accuracy that StageNet reaches with 2.0% of data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GraphCare with BAT instead of the strongest listed baseline | mortality prediction · MIMIC-III | AUROC: 61.5 -> 70.3 | Helped |
| use GraphCare with BAT instead of the strongest listed baseline | mortality prediction · MIMIC-IV | AUROC: 69.3 -> 73.1 | Helped |
| use GraphCare with BAT instead of the strongest listed baseline | readmission prediction · MIMIC-III | AUROC: 66.7 -> 69.7 | Helped |
| use GraphCare with BAT instead of the strongest listed baseline | readmission prediction · MIMIC-IV | AUROC: 66.3 -> 68.5 | Helped |
| use GraphCare with BAT instead of the strongest listed baseline | LOS prediction · MIMIC-III | F1-score: 35.0 -> 37.5 | Helped |
| use GraphCare with BAT instead of the strongest listed baseline | LOS prediction · MIMIC-IV | F1-score: 32.3 -> 34.2 | Helped |
| use GraphCare with BAT instead of the strongest listed baseline | drug recommendation · MIMIC-III | F1-score: 63.7 -> 66.8 | Helped |
| use GraphCare with BAT instead of the strongest listed baseline | drug recommendation · MIMIC-IV | F1-score: 60.4 -> 63.9 | Helped |
| add node-level attention (α) (ablation) | BAT · MIMIC-III drug recommendation | F1-score: 64.6 -> 66.8 | Helped |
| add visit-level attention (β) (ablation) | BAT · MIMIC-IV readmission | AUROC: 67.0 -> 68.5 | Helped |
| add edge weights (w_R) (ablation) | BAT · MIMIC-III readmission | AUROC: 68.4 -> 69.7 | Helped |
| add attention initialization (ablation) | BAT · MIMIC-IV drug recommendation | F1-score: 62.4 -> 63.9 | Helped |
Surprising findings
- A 2-hop UMLS subgraph sampling variant sometimes underperforms baselines without external knowledge, attributed to noise from random sampling.
Limitations
- Relies on GPT-4 outputs that may hallucinate or be biased; mitigated by manual medical expert review only.
- Random multi-hop subgraph sampling from UMLS can introduce noise and isolated nodes.
- Research use only; not validated for clinical deployment.
Experimental setup
Datasets: MIMIC-III MIMIC-IV UMLS-KG GPT-4 generated KG triples
Benchmarks: MIMIC-III MIMIC-IV
Metrics: AUROC AUPRC Kappa Accuracy F1-score Jaccard
Baselines: GRU Transformer RETAIN GRAM Deepr AdaCare GRASP StageNet SafeDrug MICRON GAMENet MoleRec
Builds on: GRAM GAMENet PyHealth GPS graph transformer
Closest papers by meaning
- Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval (ICLR 2025 · similarity 0.89)
- DearLLM: Enhancing Personalized Healthcare via Large Language Models-Deduced Feature Correlations (AAAI 2025 · similarity 0.86)
- Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data (ICLR 2025 · similarity 0.83)
- Risk Horizons: Structured Hypothesis Spaces for Longitudinal Clinical Prediction (NeurIPS 2026 · similarity 0.83)
- Knowledge-Empowered Dynamic Graph Network for Irregularly Sampled Medical Time Series (NeurIPS 2024 · similarity 0.83)
- Toward Better EHR Reasoning in LLMs: Reinforcement Learning with Expert Attention Guidance (AAAI 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.