Research map / Large language models / Retrieval and knowledge
Reasoning-Enhanced Healthcare Predictions with Knowledge Graph Community Retrieval
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 KARE, which builds a multi-source medical knowledge graph, retrieves hierarchical community summaries to augment EHR patient contexts, and fine-tunes Mistral-7B to produce reasoning chains and predictions for mortality and readmission on MIMIC-III/IV.
Problem
LLMs used for clinical outcome prediction hallucinate and lack fine-grained medical knowledge, and standard RAG retrieves semantically similar but clinically irrelevant information, hurting mortality and readmission prediction.
Approach
Construct a concept-specific multi-source KG (UMLS, PubMed triples, LLM-generated relations) anchored to EHR co-occurrences, cluster and hierarchically community-detect it with Leiden, summarize communities, then dynamically select summaries via a relevance score with node-hit decay (DGRA) and fine-tune a small LLM to jointly output reasoning and labels.
Key results
- KARE (fine-tuned Mistral-7B) reaches 64.6 macro F1 and 24.7 sensitivity on MIMIC-III mortality, versus 59.6 macro F1 for the best baseline RAM-EHR.
- On MIMIC-III readmission, KARE achieves 73.9 accuracy and 73.7 macro F1, outperforming the best baseline GraphCare at 65.4 accuracy.
- On MIMIC-IV mortality, KARE achieves 94.1 accuracy and 90.4 macro F1, compared with 91.5 accuracy and 80.3 macro F1 for GraphCare.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use KARE instead of the strongest reported baseline, RAM-EHR | MIMIC-III mortality prediction · MIMIC-III mortality | Macro F1: 59.6 -> 64.6 | Helped |
| Use KARE instead of GraphCare | MIMIC-III readmission prediction · MIMIC-III readmission | Macro F1: 64.1 -> 73.7 | Helped |
| Use KARE instead of GraphCare | MIMIC-IV mortality prediction · MIMIC-IV mortality | Macro F1: 80.3 -> 90.4 | Helped |
| Use KARE instead of GraphCare | MIMIC-IV readmission prediction · MIMIC-IV readmission | Macro F1: 65.5 -> 73.8 | Helped |
| Add reasoning to the fine-tuned model (ablation) | KARE on MIMIC-III readmission · MIMIC-III readmission | Macro F1: 57.6 -> 64.7 | Helped |
| Add retrieved knowledge and reasoning (ablation) | KARE on MIMIC-III readmission · MIMIC-III readmission | Macro F1: 57.6 -> 72.6 | Helped |
| Add similar patients to retrieved knowledge and reasoning (ablation) | KARE on MIMIC-III readmission · MIMIC-III readmission | Macro F1: 72.6 -> 73.7 | Helped |
| Add reasoning to the fine-tuned model (ablation) | KARE on MIMIC-IV readmission · MIMIC-IV readmission | Macro F1: 46.7 -> 62.1 | Helped |
| Add retrieved knowledge and reasoning (ablation) | KARE on MIMIC-IV readmission · MIMIC-IV readmission | Macro F1: 46.7 -> 71.9 | Helped |
| Add similar patients to retrieved knowledge and reasoning (ablation) | KARE on MIMIC-IV readmission · MIMIC-IV readmission | Macro F1: 71.9 -> 73.8 | Helped |
| Add similar patients to retrieved knowledge and reasoning (ablation) | KARE on MIMIC-III mortality · MIMIC-III mortality | Macro F1: 64.6 -> 61.3 | Hurt |
| Use separate multitask objectives for reasoning and label prediction instead of concatenating both outputs (ablation) | KARE on MIMIC-IV mortality · MIMIC-IV mortality | Macro F1: 86.5 -> 90.4 | Helped |
Surprising findings
- Naive classic RAG sometimes degrades zero-shot and fine-tuned performance (e.g., MIMIC-III readmission zero-shot macro F1 drops relative to context-only in some settings).
- Including similar patients degrades performance on highly imbalanced MIMIC-III mortality.
Limitations
- Evaluated only on MIMIC-III and MIMIC-IV from a single US hospital system, limiting generalizability.
- KG relies on LLM-generated outputs that may contain hallucinations and biases.
- English-only biomedical literature and EHR data.
Experimental setup
Datasets: MIMIC-III MIMIC-IV UMLS PubMed
Benchmarks: MIMIC-III MIMIC-IV
Metrics: accuracy macro F1 sensitivity specificity
Baselines: GRU Transformer RETAIN GRAM GraphCare RAM-EHR EMERGE EHR-CoAgent
Builds on: GraphRAG GraphCare EHR-CoAgent Leiden algorithm PyHealth
Closest papers by meaning
- GraphCare: Enhancing Healthcare Predictions with Personalized Knowledge Graphs (ICLR 2024 · similarity 0.89)
- KGARevion: An AI Agent for Knowledge-Intensive Biomedical QA (ICLR 2025 · similarity 0.87)
- Memorize and Rank: Elevating Large Language Models for Clinical Diagnosis Prediction (AAAI 2025 · similarity 0.87)
- LLaVA Needs More Knowledge: Retrieval Augmented Natural Language Generation with Knowledge Graph for Explaining Thoracic Pathologies (AAAI 2025 · similarity 0.87)
- Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.86)
- BoxLM: Unifying Structures and Semantics of Medical Concepts for Diagnosis Prediction in Healthcare (ICML 2025 · similarity 0.86)
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.