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, medical knowledge graph, large language models, clinical rationale units, interpretable reasoning
TL;DR¶
CRU-Reason synergizes machine learning sequential statistical priors, medical knowledge graphs, and large language models by decomposing complex electronic health records into disease-centric Clinical Rationale Units (CRUs) via constrained graph exploration, driving fine-grained evidence-grounded clinical reasoning and outperforming top EHR baselines by an average of 9.4% on MIMIC-III and MIMIC-IV.
Background & Motivation¶
Healthcare prediction using electronic health records (EHRs)—encompassing medication recommendation, recurrent diagnosis prediction, and in-hospital mortality estimation—is fundamental to modern clinical decision-support systems. Conventional machine learning and deep learning approaches (such as StageNet, RETAIN, recurrent networks, and graph neural networks) excel at mining statistical regularities and temporal patterns from sequential clinical codes. Despite their competitive predictive accuracy, these models function largely as black boxes, incapable of providing human-readable, knowledge-grounded explanations that clinicians can trace, audit, and trust in high-stakes environments. Recent attempts to incorporate large language models (LLMs) with external medical knowledge graphs (KGs) introduce structured domain relations into reasoning prompts; nevertheless, existing approaches predominantly adhere to monolithic reasoning over entire patient records. In this paradigm, multi-visit trajectories spanning dozens of heterogeneous conditions and interventions are fed into the LLM in a single pass alongside broad graph retrieval contexts.
Such monolithic, patient-wide reasoning introduces critical clinical attribution roadblocks. Real-world patient histories frequently exhibit high complexity and multiorgan comorbidities. Task-agnostic graph retrieval inevitably pulls in noisy, cross-organ, or irrelevant relational triplets that distract reasoning. More critically, when an LLM evaluates a long, tangled record all at once, it struggles to attribute a candidate medication or diagnostic hypothesis to specific causal conditions or physiological complications. This frequently leads to superficial or hallucinated justifications that sound grammatically fluent yet detach from the patient's underlying disease etiology. Conversely, clinical concepts in EHRs exhibit natural localized structure when connected through medical KGs: diseases, pathological mechanisms, diagnostic procedures, and therapeutic drugs cluster around shared physiological goals. Empirical analysis indicates that modeling connected concepts jointly within a KG-defined community increases the empirical probability of correctly selecting indicated medications by 67% and rejecting contraindicated candidates by 74% compared to isolated concept modeling.
Consequently, decomposing tangled longitudinal records into clinically coherent, localized sub-problems is essential to resolve the dilemma between statistical accuracy and fine-grained attribution. Core idea: construct disease-centric Clinical Rationale Units (CRUs) via rule-constrained graph exploration, integrating machine learning statistical priors and structured medical knowledge graphs to empower large language models with fine-grained, traceable clinical reasoning.
Method¶
Overall Architecture¶
CRU-Reason establishes a modular three-stage framework comprising candidate discovery, rule-constrained CRU construction, and CRU-wise LLM reasoning with multi-unit aggregation. Given longitudinal patient records, the framework first retrieves high-precision statistical candidates from a pretrained machine learning model and broader contextual candidates from the medical knowledge graph. Next, taking diagnosed conditions as seed nodes, it executes constrained breadth-first search (BFS) over the visit-specific knowledge subgraph governed by curated metapaths and LLM-assisted semantic relevance pruning, partitioning the complex record into mutually disjoint, disease-centered connected components (CRUs). Finally, the LLM synthesizes a clinical summary for each CRU's local context, evaluates ML-derived candidates using direct topological evidence, and assesses KG-derived candidates via few-shot exemplar transfer from co-occurring ML predictions before aggregating scores across all CRUs.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Patient Longitudinal EHR<br/>(Diseases D, Procedures P, Meds M)"] --> B["Dual-Source Candidate Discovery<br/>ML statistical priors + KG neighborhood"]
B --> C["Constrained CRU Construction<br/>Metapath filtering + LLM relevance pruning BFS"]
C --> D["CRU Context Summarization<br/>Generate local clinical narrative S"]
D --> E["Fine-Grained Dual Reasoning & Aggregation<br/>ML few-shot exemplar transfer + Multi-CRU scoring"]
E --> F["Output Predictions & Explanations<br/>(Calibrated score + Traceable rationale chain)"]
Key Designs¶
1. Dual-source candidate discovery: synergizing sequential statistical priors and global graph connectivity
Directly prompting a general-purpose LLM to predict targets over vast clinical codebooks (containing thousands of ICD/CCS diagnoses or ATC medication classes) induces catastrophic hallucination due to the lack of domain- and task-specific inductive bias. Conversely, relying exclusively on knowledge graph retrieval suffers from task-agnostic noise, as biomedical KGs encode universal medical relations without task-specific optimization. To resolve this bottleneck, candidate generation is decoupled into a data-driven ML candidate set \(E_{\text{ML}}\) and a knowledge-driven graph candidate set \(E_{\text{KG}}\):
Here \(f_{\text{ML}}\) represents a lightweight temporal EHR model (e.g., StageNet) pretrained on historical sequences \(X_{\le T}\) to supply high-precision task priors; \(L\text{-hop}(e, \mathcal{G})\) extracts local topological neighbors from the unified medical knowledge graph \(\mathcal{G}\) (comprising 43,601 entities and 229,991 relations, with default \(L=1\)). This dual-source mechanism exploits the data-driven model's sharp discriminative focus while leveraging the knowledge graph to safeguard candidate recall and cover long-tail associations.
2. Rule-constrained CRU construction: entropy-ordered breadth-first exploration with semantic boundaries
Unrestricted graph traversal across a multi-condition EHR record causes rapid combinatorial explosion and semantic drift, re-entangling disparate organ systems into monolithic graphs. To build clinically self-contained sub-problems, a Clinical Rationale Unit is formalized as a localized community \(H = (E_{\text{ctx}}, E_{\text{ML}}, E_{\text{KG}}, R)\) centered around a primary disease. In each visit \(x_t\), diseases are prioritized as seed nodes by calculating the entropy of their co-occurrence distributions with candidate targets in the training corpus; lower entropy reflects stronger indicative specificity.
Starting from the highest-priority unvisited disease seed, a breadth-first search constructs the CRU under two explicit constraints. First, traversal is strictly governed by expert-curated clinical metapaths (such as \(D \to P \to M\) for disease \(\to\) procedure \(\to\) medication), enforcing biologically and therapeutically plausible transition pathways. Second, an LLM evaluates the clinical relevance of each candidate edge \((u \to r \to v)\) using few-shot prompts to confirm whether both concepts belong to the same pathological process or physiological system, pruning irrelevant relations. Traversing reachable entities up to a predefined depth (3 hops for drug recommendation, 2 hops for diagnosis prediction) isolates one connected component per primary disease, yielding a clean partition of the visit graph into disjoint disease-centric CRUs.
3. Fine-grained CRU-wise dual reasoning and exemplar-guided transfer: modular deduction and score aggregation
Once CRUs are constructed, LLMs perform localized, fine-grained reasoning independently over each unit, eliminating inter-disease interference. Reasoning proceeds across three coordinated tiers:
First, local context summarization extracts the subgraph \(G(E_{\text{ctx}})\) induced by historical concepts in the unit. The LLM synthesizes this structured neighborhood into a concise clinical narrative \(S = \text{LLM}(I_{\text{ctx}}, E_{\text{ctx}}, G(E_{\text{ctx}}))\), detailing the primary condition, active complications, and therapeutic intentions.
Second, the system executes dual-track candidate verification. For high-confidence ML candidates \(e \in E_{\text{ML}}\), the LLM inspects their direct relational subgraph \(G(e, E_{\text{ctx}})\) against narrative \(S\), outputting a discrete likelihood score \(\hat{y}(e) \in \{0, 1, 2, 3, 4\}\) alongside an evidence-grounded natural-language rationale \(o_{\text{ML}}(e)\). For KG-derived candidates \(e \in E_{\text{KG}}\), which lack task-specific training, the framework introduces exemplar-guided reasoning transfer: it retrieves the top-\(k\) ML candidates that historically co-occur most frequently with candidate \(e\), using their validated rationales as dynamic few-shot exemplars \(B(e)\) to ground the evaluation of \(e\).
Finally, candidate scores are gathered across all CRUs. For any candidate appearing across multiple disease units, its predicted scores are averaged into a consolidated rating; candidates matching or exceeding the decision threshold are selected as positive predictions, accompanied by the corresponding disease-specific rationales.
A Worked Example¶
Consider a complex intensive care patient presenting with multiorgan pathology: 1. Input EHR context: Diagnoses include cancer of bronchus, chronic obstructive pulmonary disease (COPD), respiratory failure, and septic shock, with clinical procedures including thoracentesis and invasive mechanical ventilation. 2. Candidate discovery & CRU partitioning: Rather than processing all diagnoses together, constrained BFS splits the visit into two distinct CRUs. CRU #1 clusters bronchus cancer, COPD, respiratory failure, thoracentesis, and ventilation, retrieving adrenergics, inhalants, and corticosteroids. CRU #2 isolates septic shock and systemic bacteremia, linking antiinfective agents. 3. Localized reasoning & attribution: When evaluating corticosteroids, CRU #1 grounds the candidate in bronchial inflammation and acute ventilator support, scoring it 4/5 with the rationale: "Corticosteroids are indicated for acute COPD exacerbations and reduce airway inflammation in respiratory failure." In contrast, in a background unit referencing historical localized lesions, corticosteroids receive a lower priority score (2.5/5) with explicit notes regarding limited utility. Multi-CRU aggregation cleanly separates respiratory management from antimicrobial therapies, preventing cross-disease contamination.
Key Experimental Results¶
Main Results¶
Evaluations were conducted on two widely benchmarked clinical datasets, MIMIC-III (v1.4) and MIMIC-IV (v2.0), across drug recommendation and recurrent diagnosis prediction against temporal ML architectures, standalone and reasoning-augmented LLMs, KG-retrieval LLM baselines, and hybrid ML+KG+LLM methods.
| Dataset | Task | Metric | CRU-Reason (Ours) | Prev. SOTA (Second Best) | Gain |
|---|---|---|---|---|---|
| MIMIC-III | Drug Rec. | Macro F1 ↑ | 0.657 | 0.635 (RETAIN) | +3.5% |
| MIMIC-III | Drug Rec. | Jaccard ↑ | 0.506 | 0.485 (TRANS) | +4.3% |
| MIMIC-III | Drug Rec. | Rank ↓ | 1.639 | 1.896 (TRANS) | +13.6% |
| MIMIC-IV | Drug Rec. | Macro F1 ↑ | 0.586 | 0.569 (StageNet) | +3.0% |
| MIMIC-IV | Drug Rec. | Jaccard ↑ | 0.434 | 0.419 (StageNet) | +3.6% |
| MIMIC-IV | Drug Rec. | Rank ↓ | 1.645 | 1.853 (StageNet) | +10.4% |
| MIMIC-III | Diagnosis Pred. | Macro F1 ↑ | 0.397 | 0.371 (Qwen3-235B) | +7.0% |
| MIMIC-III | Diagnosis Pred. | Jaccard ↑ | 0.260 | 0.240 (Qwen3-235B) | +8.3% |
| MIMIC-III | Diagnosis Pred. | Rank ↓ | 2.047 | 2.439 (StageNet) | +15.2% |
| MIMIC-IV | Diagnosis Pred. | Macro F1 ↑ | 0.370 | 0.345 (Qwen3-235B) | +7.2% |
| MIMIC-IV | Diagnosis Pred. | Jaccard ↑ | 0.243 | 0.222 (Qwen3-235B) | +9.5% |
| MIMIC-IV | Diagnosis Pred. | Rank ↓ | 2.037 | 2.803 (GPT4.1-mini) | +27.3% |
Ablation Study¶
To dissect the specific contributions of localized graph decomposition, exploration constraints, and dual-source modeling, systematic ablations were conducted across MIMIC-III and MIMIC-IV.
| Config | MIMIC-III Drug F1 (∆%) | MIMIC-III Drug Jaccard | MIMIC-III Diag F1 (∆%) | MIMIC-IV Drug F1 (∆%) | MIMIC-IV Diag F1 (∆%) | Note |
|---|---|---|---|---|---|---|
| Full model | 0.657 (-) | 0.506 | 0.397 (-) | 0.586 (-) | 0.370 (-) | Full CRU-Reason architecture |
| w/o CNST | 0.648 (-2.0%) | 0.494 | 0.385 (-1.5%) | 0.582 (-1.6%) | 0.345 (-1.7%) | Traversal without metapaths & LLM edge pruning |
| w/ LM-CRU | 0.450 (-26.2%) | 0.305 | 0.339 (-11.3%) | 0.481 (-15.5%) | 0.278 (-17.0%) | Unconstrained LLM-prompted text clustering |
| w/o CRU | 0.629 (-3.0%) | 0.473 | 0.378 (-3.4%) | 0.560 (-6.4%) | 0.342 (-6.4%) | Monolithic patient-wide reasoning |
| w/o KG | 0.598 (-7.8%) | 0.444 | 0.353 (-7.6%) | 0.528 (-9.0%) | 0.317 (-10.5%) | Eliminating knowledge graph components |
| w/o ML | 0.488 (-20.1%) | 0.342 | 0.361 (-13.5%) | 0.457 (-23.1%) | 0.297 (-25.3%) | Removing ML model candidate priors |
Key Findings¶
- Statistical ML priors provide the indispensable performance foundation: Stripping ML candidate priors (
w/o ML) causes F1 to collapse by 20.1% to 23.1% on drug recommendation and 13.5% to 25.3% on diagnosis prediction. General-purpose LLMs augmented solely with KGs lack temporal sequence calibration over longitudinal visits; ML priors are vital to restrict candidate search spaces. - Topological graph exploration guarantees community coherence: Delegating unit construction entirely to LLMs (
w/ LM-CRU) triggers a massive 26.2% F1 degradation. Free-form language prompting lacks structural topology guarantees, producing noisy clusters that violate clinical ontologies; rule-constrained BFS over KGs provides the necessary structural guardrails. - Localized decomposition yields superior explainability and out-of-distribution robustness: In blind qualitative audits conducted by three board-certified physicians across 2,000 paired decisions (achieving strong inter-rater agreement, Cohen's \(\kappa = 0.82\)), CRU-Reason achieved win rates between 53.5% and 81.0% over MedReason, MedIKAL, and KARE in grounding, granularity, and clinical safety. Under cross-dataset transfer between MIMIC-III and MIMIC-IV, CRU-Reason suffered minimal performance degradation (-2.4% to -14.6%), substantially outperforming purely supervised models whose metrics plummeted by 26% to 38%.
Highlights & Insights¶
- Disease-centric decoupling via Clinical Rationale Units: Instead of stuffing thousands of unstructured tokens and noisy graph triplets into an unwieldy prompt, CRU-Reason modularizes EHRs into disease-centered reasoning kernels, structurally resolving context contamination and attribution ambiguity.
- Harmonious integration of connectionist learning and symbolic reasoning: Leveraging deep sequence models for high-accuracy statistical priors while utilizing knowledge graphs for grounded causal explanations—bridged by dynamic exemplar transfer—establishes a balanced paradigm for medical decision intelligence.
- Generalizable blueprint for clinical agent workflows: The four-stage blueprint—sequential prior filtering \(\to\) constrained subgraph partitioning \(\to\) modular LLM reasoning \(\to\) multi-unit evidence aggregation—naturally transfers to complex clinical AI systems, including multi-morbidity monitoring, ICU complication early warning, and drug-drug interaction auditing.
Limitations & Future Work¶
- Dependence on standardized structured clinical codes: The framework is currently evaluated on curated ICD/CCS diagnosis codes and ATC drug classes from ICU datasets; it does not yet extract concepts directly from unstructured clinical notes, leaving error propagation from entity extraction unaddressed.
- Reliance on external knowledge graph fidelity: The implementation directly adopts the KARE biomedical graph without independent verification of relational freshness or factual accuracy, allowing potential graph errors to propagate into generated clinical rationales.
- Inference overhead and discrete scoring: Evaluating multiple CRUs increases the average number of LLM invocations to 25–57 calls per patient visit. Furthermore, the model produces discrete scores rather than calibrated continuous probabilities, necessitating future research on uncertainty quantification and automated error auditing.
Related Work & Insights¶
- vs Black-box temporal EHR models (StageNet, RETAIN): Deep sequence models provide robust statistical discriminative power but fail to deliver interpretable, medical-grade rationales; CRU-Reason preserves their discriminative strength as candidate priors while supplying human-auditable causal justifications.
- vs Monolithic KG-augmented LLM architectures (GraphCare, KARE, MedReason): Prior methods concatenate retrieved knowledge into global prompts, obscuring disease-specific attribution under multi-condition noise; CRU-Reason enforces disease-centered graph exploration, isolating reasoning into localized CRUs and achieving superior explanation granularity and clinical safety.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Formulates the concept of Clinical Rationale Units (CRUs), elegantly bridging statistical priors and knowledge graph topology for modular clinical reasoning.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorously evaluated across two major clinical benchmarks, dual prediction tasks, cross-backbone variations, cross-dataset distribution shifts, extensive ablations, and physician-blinded qualitative reviews.
- Writing Quality: ⭐⭐⭐⭐⭐ Meticulously motivated, mathematically transparent, and supported by concrete clinical walkthroughs.
- Value: ⭐⭐⭐⭐⭐ Sets a new benchmark for interpretable, trustworthy AI clinical decision support in high-stakes healthcare settings.