CRU-Reason: Knowledge Graph-Grounded and Interpretable Clinical Prediction with Clinical Rationale Units¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/lzhmarkk/CRU-Reason
Area: Medical LLM
Keywords: clinical prediction, knowledge graph, clinical rationale units, large language models, electronic health records
TL;DR¶
CRU-Reason integrates machine learning priors and constrained knowledge graph exploration to construct disease-centered Clinical Rationale Units (CRUs), guiding large language models to conduct fine-grained, decoupled clinical reasoning and delivering a 9.4% average improvement over strong baselines in drug recommendation and diagnosis prediction.
Background & Motivation¶
Healthcare prediction using electronic health records (EHRs) serves as a cornerstone of modern clinical decision support systems (CDSS), supporting high-stakes tasks such as drug recommendation, next-visit disease diagnosis, and mortality estimation. Traditional machine learning (ML) architectures—including gated recurrent units (GRUs), RETAIN, and stage-aware networks like StageNet—achieve competitive discriminative performance by learning temporal statistical patterns across sequential patient visits. However, these data-driven models fundamentally operate as opaque black boxes, failing to provide human-interpretable, evidence-grounded rationales, which severely impedes their trustworthiness and adoption in clinical workflows.
To enhance interpretability and medical precision, recent efforts incorporate biomedical knowledge graphs (KGs) to ground large language models (LLMs) with structured clinical facts. Nevertheless, prevailing KG-augmented methods typically perform monolithic reasoning over entire longitudinal records: they concatenate a patient's multifaceted visit history with broadly retrieved KG subgraphs directly into a single prompt. This global aggregation introduces substantial semantic noise, mixing irrelevant clinical contexts from disparate organ systems and obscuring disease-specific attribution, which often leads to vague, hallucinated, or medically unsupported justifications.
The key insight of this work is that clinical concepts within EHRs naturally exhibit localized community structures governed by shared pathological mechanisms and therapeutic objectives (e.g., uncontrolled type 2 diabetes clustering with hyperglycemia, insulin resistance, HbA1c tests, and hypoglycemic agents). Core idea: decompose multi-morbid EHR records into disease-centered localized communities—termed Clinical Rationale Units (CRUs)—through constrained knowledge graph exploration and ML statistical priors, enabling LLMs to perform fine-grained, traceable reasoning over decoupled clinical sub-problems.
Method¶
Overall Architecture¶
CRU-Reason executes an end-to-end pipeline consisting of candidate discovery, constrained unit construction, decoupled reasoning, and final cross-unit aggregation. Given historical patient visits \(X = \{x_1, \dots, x_T\}\) containing disease codes \(D\), procedure codes \(P\), and medication codes \(M\), the framework first discovers dual candidate sets via a pre-trained ML model and global KG neighborhood expansion. Next, starting from unvisited disease seeds, a constrained breadth-first search (BFS) governed by expert metapaths and LLM semantic relevance scoring partitions the visit subgraph into disjoint Clinical Rationale Units (CRUs). An LLM then summarizes the localized context of each CRU, evaluates candidates under ML-derived few-shot guidance, and aggregates scores across units to generate final predictions alongside evidence-grounded rationales.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Patient Historical EHR Records<br/>(Diseases D, Procedures P, Medications M)"] --> B["Stage 1: Candidate Discovery<br/>ML Statistical Priors + KG Expansion"]
B --> C["Stage 2: Constrained BFS Graph Exploration<br/>Metapath Constraints + Semantic Edge Scoring"]
C --> D["Stage 3: CRU Context Summarization<br/>Local Subgraph Abstraction + Clinical Synthesis"]
D --> E["Stage 4: Prior-Guided CRU-Wise Reasoning<br/>Dual-Track Evaluation + Evidence Rationale Generation"]
E --> F["Stage 5: Cross-Unit Aggregation & Decision<br/>Score Averaging + Structured Clinical Predictions"]
Key Designs¶
1. Dual-Track Candidate Discovery: Balancing Task-Specific Precision and Relational Coverage
Directly prompting general-purpose LLMs over massive medical code spaces inevitably induces severe hallucination due to EHR concept sparsity. To mitigate this, CRU-Reason derives two complementary candidate pools for each patient record \(X_{\le T}\): a high-precision, task-specific prior \(E_{\text{ML}}\) produced by a pre-trained sequential ML backbone \(f_{\text{ML}}\) (e.g., StageNet), and a broad relational candidate pool \(E_{\text{KG}}\) formed by extracting the \(L\)-hop neighborhood around patient concepts from the biomedical KG \(G\):
This dual-source strategy ensures that the subsequent reasoning modules are primed with both statistically robust, task-aligned candidates and medically valid long-tail connections.
2. Constrained KG Exploration for CRU Construction: Decoupling Multi-Morbidity into Localized Subgraphs
Standard monolithic graph retrieval indiscriminately merges heterogeneous diseases and treatments into a noisy context. To isolate disease-specific dependencies, CRU-Reason constructs localized subgraphs \(H_{t,i} = (E_{\text{ctx}}^{t,i}, E_{\text{ML}}^{t,i}, E_{\text{KG}}^{t,i}, R_{t,i})\) centered around individual diseases using constrained breadth-first search (BFS). The traversal enforces two complementary filtering rules: - Metapath Constraint: Exploration strictly follows expert-curated clinical trajectories (e.g., \(D \to P \to M\), denoting disease \(\to\) procedure \(\to\) medication), preventing arbitrary hops and preserving causal therapeutic logic. - LLM Relevance Pruning: Each traversed edge is assessed by an LLM via clinical relevance prompts, pruning semantically disconnected neighbors that exceed cross-system boundaries (e.g., validating shared cardiovascular etiology between hypertension and coronary disease).
The BFS expands from each disease seed until reaching max exploration depth, producing a connected CRU; the algorithm then iterates to the next unvisited disease until all conditions are mapped into disjoint disease-centric communities.
3. Progressive CRU-Wise Reasoning: Knowledge Grounding and Exemplar-Driven Score Aggregation
Within each isolated CRU, reasoning operates in a two-tier fashion. First, the LLM digests the context subgraph \(G(E_{\text{ctx}})\) to generate a coherent clinical summary \(S\) highlighting primary conditions and therapeutic goals. Second, candidate evaluation is conducted: for data-driven candidates \(e \in E_{\text{ML}}\), the model pairs the candidate with its direct context subgraph to output a discrete score \(\hat{y}(e) \in \{0, 1, 2, 3, 4\}\) and rationale \(o_{\text{ML}}(e)\).
For knowledge-derived candidates \(E_{\text{KG}}\), which often contain task-agnostic noise, the framework leverages historical co-occurrence frequency \(\phi(e, e')\) with ML candidates to retrieve top-\(k\) verified ML rationales as few-shot exemplars \(B(e)\):
This transfers disciplined clinical reasoning patterns from high-confidence ML candidates to broader KG entities. Finally, for targets appearing across multiple CRUs, scores are averaged, and candidates exceeding the calibrated threshold are selected alongside their attribution-traceable clinical rationales.
A Worked Example¶
Consider a complex patient admitted with bronchus cancer, chronic obstructive pulmonary disease (COPD), respiratory failure, and status post-lobectomy: 1. CRU Partitioning: Constrained BFS builds two distinct units. CRU #1 clusters bronchus cancer, secondary acute respiratory failure, and mechanical ventilation; CRU #2 clusters post-surgical recovery and secondary prophylactic management. 2. Target Evaluation: When evaluating inhalants, CRU #1 connects "bronchus cancer \(\to\) complicates \(\to\) respiratory failure" and "ventilation \(\to\) requires \(\to\) inhalants", prompting the LLM to assign an optimal score (4/5) justified by guideline-directed COPD exacerbation protocols. For corticosteroids, CRU #1 recognizes its acute airway anti-inflammatory efficacy (4/5), while CRU #2 correctly down-weights it as secondary palliative care (2.5/5). 3. Consolidated Output: Inhalants are recommended with explicit attribution to respiratory failure support, entirely eliminating erroneous attribution to unrelated surgical antibiotics.
Key Experimental Results¶
Main Results¶
Evaluations are conducted on two public benchmarks, MIMIC-III (v1.4) and MIMIC-IV (v2.0), across drug recommendation and diagnosis prediction using Macro F1, Jaccard similarity, and mean candidate rank position (Rank, lower is better):
| Task / Dataset | Metrics | CRU-Reason (Ours) | Strongest Baseline (StageNet/TRANS/GPT4.1-m) | Relative Gain |
|---|---|---|---|---|
| MIMIC-III Drug Recommendation | Macro F1 ↑ / Jaccard ↑ / Rank ↓ | 0.657 / 0.506 / 1.639 | 0.635 / 0.485 / 1.896 (TRANS) | +3.5% / +4.3% / +13.6% |
| MIMIC-IV Drug Recommendation | Macro F1 ↑ / Jaccard ↑ / Rank ↓ | 0.586 / 0.434 / 1.645 | 0.569 / 0.419 / 1.853 (StageNet) | +3.0% / +3.6% / +10.4% |
| MIMIC-III Diagnosis Prediction | Macro F1 ↑ / Jaccard ↑ / Rank ↓ | 0.397 / 0.260 / 2.047 | 0.321 / 0.211 / 2.439 (StageNet) | +7.0% / +8.3% / +15.2% |
| MIMIC-IV Diagnosis Prediction | Macro F1 ↑ / Jaccard ↑ / Rank ↓ | 0.370 / 0.243 / 2.037 | 0.327 / 0.219 / 2.803 (GPT4.1-m) | +7.2% / +9.5% / +27.3% |
Across all four settings, CRU-Reason improves upon the strongest evaluated baseline by an average of +9.4%, achieving superior candidate ranking quality (Rank \(\sim 1.6 - 2.0\)).
Ablation Study¶
Ablation experiments on MIMIC-III and MIMIC-IV isolate the impact of graph constraints, unit decomposition, and candidate sources:
| Variant Config | MIMIC-III Drug F1 (Jaccard) | MIMIC-IV Drug F1 (Jaccard) | MIMIC-III Diag F1 (Jaccard) | MIMIC-IV Diag F1 (Jaccard) | Relative Impact & Notes |
|---|---|---|---|---|---|
| Full Model (CRU-Reason) | 0.657 (0.506) | 0.586 (0.434) | 0.397 (0.260) | 0.370 (0.243) | Full architecture |
| w/o CNST (Unconstrained CRU) | 0.648 (0.494) | 0.582 (0.426) | 0.385 (0.252) | 0.345 (0.224) | Omitting metapath & pruning introduces graph noise (-1.5% to -2.0%) |
| w/ LM-CRU (LLM-Generated CRU) | 0.450 (0.305) | 0.481 (0.330) | 0.339 (0.222) | 0.278 (0.178) | Severe collapse (-11.3% to -26.2%), indicating raw LLM graph clustering is noisy |
| w/o CRU (Monolithic Reasoning) | 0.629 (0.473) | 0.560 (0.402) | 0.378 (0.246) | 0.342 (0.222) | Full patient-level reasoning dilutes disease-centric attribution (-3.0% to -6.4%) |
| w/o KG (Removing KG Modules) | 0.598 (0.444) | 0.528 (0.373) | 0.353 (0.235) | 0.317 (0.209) | Deprived of structured factual grounding and expansion (-7.6% to -10.5%) |
| w/o ML (Removing ML Priors) | 0.488 (0.342) | 0.457 (0.314) | 0.361 (0.203) | 0.297 (0.185) | Substantial drop (-13.5% to -25.3%), confirming ML priors are essential filters |
Key Findings¶
- ML Priors Provide Essential Statistical Anchoring: Removing ML candidates (
w/o ML) induces the steepest performance drops across all tasks (-13.5% to -25.3%), proving that generalist LLMs and KGs alone struggle to navigate discrete EHR search spaces without task-trained statistical bounds. - Topological Constraints Are Indispensable: Directly prompting LLMs to cluster patient records (
w/ LM-CRU) degrades F1 by over 26%, while unconstrained BFS (w/o CNST) admits spurious links. True performance stems from pairing deterministic KG graph structure with semantic pruning. - Superior Cross-Dataset Out-of-Distribution Robustness: In cross-dataset evaluation (training on MIMIC-IV and evaluating zero-shot on MIMIC-III, and vice versa), CRU-Reason exhibits minor performance decay (-2.4% to -14.6%), substantially outperforming trained deep models like StageNet (-26.2% to -35.3%) and GraphCare (-23.4% to -38.6%), underscoring the invariant grounding provided by the medical KG.
Highlights & Insights¶
- Disease-Centered Rationale Units: Shifts clinical NLP away from monolithic, kitchen-sink prompt concatenation toward modular, disease-specific sub-problem decomposition, establishing a cleaner framework for multi-morbidity reasoning.
- Symbiotic Fusion of Data-Driven and Knowledge-Driven Paradigms: Bridges the gap between black-box temporal EHR networks and open-ended generative LLMs, demonstrating that ML priors can supervise and calibrate knowledge-based generation.
- Physician-Validated Clinical Trustworthiness: In blind evaluations across 2,000 comparisons with three certified physicians, CRU-Reason achieved up to 81.0% physician win rate over baselines in factual grounding, reasoning granularity, and contraindication safety.
Limitations & Future Work¶
- Dependence on Standardized Discrete Codes: Operates exclusively on pre-coded ICD, CCS, and ATC-3 vocabularies from ICU datasets, lacking direct ingestion of unstructured clinical narratives and progress notes.
- Unverified External KG Completeness: Adopts the KARE biomedical KG without independent factual auditing, leaving open the risk of error propagation from KG inaccuracies.
- Discrete Scoring without Continuous Calibration: Relies on 5-point ordinal ratings and heuristic thresholds rather than well-calibrated posterior probabilities, pointing to conformal prediction as a valuable future direction.
Related Work & Insights¶
- vs KARE / MedRAG: KARE retrieves isolated subgraphs based on corpus statistics; CRU-Reason originates directly from patient diagnoses with constrained path exploration, preventing irrelevant cross-system noise from polluting the prompt context.
- vs GraphCare / RetCare: RetCare and GraphCare either rely heavily on parametric graph networks or holistic patient retrieval; CRU-Reason decouples EHR records into structured rationale units, achieving traceable single-disease attribution.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Introduces Clinical Rationale Units to decouple complex EHRs into disease-centric subgraphs for structured LLM reasoning.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Exhaustive dual-dataset benchmarks, cross-dataset OOD transfers, physician blind evaluations, and comprehensive ablations.
- Writing Quality: ⭐⭐⭐⭐⭐ Rigorous methodology formulation, lucid architecture diagrams, and clear empirical insights.
- Value: ⭐⭐⭐⭐⭐ Effectively resolves the interpretability bottleneck of clinical prediction models, providing a practical blueprint for trustworthy medical AI.