Research map / Graph learning / Knowledge and temporal graphs
HI-DR: Exploiting Health Status-Aware Attention and an EHR Graph+ for Effective Medication Recommendation
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.
Proposes HI-DR, a medication recommender that selects similar-health-status visits from any patient via a pre-trained encoder (health status-aware attention) and uses a directed EHR graph+ weighted by medication-level co-prescription degree, evaluated on MIMIC-III and MIMIC-IV.
Problem
Medication recommendation for a patient's current visit needs to use past medications in similar health status and capture asymmetric co-prescription relations, while keeping adverse drug-drug interaction rates low.
Approach
Health status-aware attention selects top-N similar visits from all patients (using patient representations from a pre-trained VITA model) instead of only the patient's own past visits, and an EHR graph+ uses directed edges weighted by medication-level co-prescription degree, modeled with DiGCN.
Key results
- On MIMIC-III, HI-DR reaches Jaccard 0.6281 vs 0.5292 for the best competitor MoleRec (up to 18.69% gain).
- On MIMIC-IV, HI-DR reaches Jaccard 0.6119 vs 0.5218 for VITA (up to 17.26% gain).
- HI-DR's DDI rate is 0.0802 on MIMIC-III (below the ground-truth 0.0808) and 0.0912 on MIMIC-IV (comparable to 0.0909).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use health status-aware attention to select similar visits regardless of whose visits they are (ablation) | HI-DR · MIMIC-III | Jaccard: 0.5298 -> 0.6281 | Helped |
| Use health status-aware attention to select similar visits regardless of whose visits they are (ablation) | HI-DR · MIMIC-III | Jaccard: 0.5368 -> 0.6281 | Helped |
| Use health status-aware attention to select similar visits regardless of whose visits they are (ablation) | HI-DR · MIMIC-III | Jaccard: 0.5338 -> 0.6281 | Helped |
| Select similar visits using pre-trained patient representations rather than diagnoses/procedures alone (ablation) | HI-DR · MIMIC-III | Jaccard: 0.5361 -> 0.6281 | Helped |
| Add the EHR graph+ rather than remove EHR-graph modeling (ablation) | HI-DR · MIMIC-III | DDI rate: 0.0845 -> 0.0802 | Helped |
| Use the EHR graph+ instead of the original EHR graph (ablation) | HI-DR · MIMIC-III | DDI rate: 0.0855 -> 0.0802 | Helped |
| Use medication-level directed co-prescription weights instead of pair-level undirected weights (ablation) | HI-DR · MIMIC-III | DDI rate: 0.0835 -> 0.0802 | Helped |
| Use the EHR graph+ with DiGCN rather than Graph Attention Network on the original EHR graph (ablation) | HI-DR · MIMIC-III | Jaccard: 0.5199 -> 0.6281 | Helped |
| Use HI-DR instead of the strongest competitor | Medication recommendation · MIMIC-III | Jaccard: 0.5292 -> 0.6281 | Helped |
| Use HI-DR instead of the strongest competitor | Medication recommendation · MIMIC-IV | Jaccard: 0.5218 -> 0.6119 | Helped |
Surprising findings
- The health status-aware attention selects none of the patient's own past visits for about 92.84% of patients.
- Using the original EHR graph (-G+) yields a higher DDI rate than not using any EHR graph (-G).
Limitations
- Relies on a pre-trained model Φ (VITA) for similar-visit selection, adding a multi-stage training pipeline.
- Accuracy gains come with DDI rates that must be kept near the ground-truth level; safety is not improved beyond ground truth on MIMIC-IV.
- Evaluated only on MIMIC-III and MIMIC-IV; some competitors (MERITS, REFINE, DAPSNet, PROMISE) could not be run.
Experimental setup
Datasets: MIMIC-III MIMIC-IV
Benchmarks: MIMIC-III MIMIC-IV
Metrics: Jaccard PRAUC F1 DDI rate
Baselines: Nearest RETAIN LEAP GAMENet MRSC SafeDrug DrugRec COGNet MoleRec VITA
Builds on: VITA COGNet MRSC DAPSNet MERITS
Closest papers by meaning
- VITA: ‘Carefully Chosen and Weighted Less’ Is Better in Medication Recommendation (AAAI 2024 · similarity 0.88)
- GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs (ICLR 2024 · similarity 0.82)
- SAFER: A Calibrated Risk-Aware Multimodal Recommendation Model for Dynamic Treatment Regimes (ICML 2025 · similarity 0.81)
- Fine-grained List-wise Alignment for Generative Medication Recommendation (NeurIPS 2025 · similarity 0.81)
- EHRCon: Dataset for Checking Consistency between Unstructured Notes and Structured Tables in Electronic Health Records (NeurIPS 2024 · similarity 0.80)
- Enhancing Healthcare Recommendations: A Privacy-Protective and Interpretable Cross-Domain Framework (AAAI 2025 · similarity 0.80)
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.