PaperlayerView exact revision on arXiv

Source-linked AI summary

Neurosymbolic Alignment for Physiologically-Safe Clinical Language Models

Abdulhady Abas Abdullah, Erik Cambria, Milena Zivkovic

arXiv:2608.24534v1cs.AI

TL;DR

Clinical LLMs can produce factually plausible recommendations that remain physiologically unsafe, while existing mitigations do not directly enforce physiological consistency in multi-step recommendations. Neurosymbolic Alignment derives preference supervision from an HGNN-based Physiological World Model, achieving the highest reported safety scores under controlled evaluation, though external validation remains necessary for deployment transfer.

  • Problem

    Factual correctness alone is insufficient for clinical safety because recommendations can conflict with the implied patient state, and current mitigations do not directly enforce physiological consistency in multi-step recommendations.

  • Method

    Neurosymbolic Alignment derives preference supervision from an HGNN-based Physiological World Model and applies it to training clinical LLMs.

  • Results

    90.8% CSS was achieved under controlled evaluation, alongside 86.4% RSS and 91.6% DID, with physician-evaluated hallucination rate reduced to 5.1%.

  • Takeaways & Limitations

    Structured physiological knowledge can generate effective preference supervision and produce safety gains supported by HGNN-independent and physician-adjudicated endpoints.

  • Takeaways & Limitations

    CSB is synthetic, and synthetic EHR-style noise tests do not replace external validation on de-identified clinical notes or blinded clinician trust studies.

Abstract

from arXiv · show

Clinical LLMs can generate recommendations that are factually plausible yet physiologically unsafe. We investigate whether safety alignment can be improved by grounding preference optimization in structured physiological knowledge rather than text-only supervision. Methods: We propose Neurosymbolic Alignment, a training-time framework that couples a 7B clinical LLM with an HGNN-based Physiological World Model over an 847K-node biomedical knowledge graph. Candidate responses are scored using homeostatic constraints, multi-hop path plausibility, and drug-interaction penalties, and the resulting rankings drive iterative on-policy ORPO updates. Evaluation is performed on the Clinical Safety Benchmark (CSB), a 2,500-scenario benchmark for physiological constraint violations in generative clinical reasoning. Results: Relative to ORPO, the proposed method improves CSS from 69.5% to 90.8% (+21.3 pp), reduces physician-evaluated HR from 14.1% to 5.1% on the blinded subset, and improves DID from 72.8% to 91.6%. These gains are corroborated by an HGNN-independent Rule-Engine Safety Score (RSS: 86.4%, +21.2 pp over ORPO; r=0.97 concordance with CSS). The method also exceeds GPT-4 (5-shot) on all safety metrics despite a 10x parameter disadvantage, and outperforms an inference-time self-correction pipeline (SFT+SelfCorrect) by 11.4 pp CSS. Under synthetic EHR-style noise, 84.2% CSS is retained. Ablation analysis shows that HGNN scoring (-16.2 pp) and iterative training (-11.5 pp) are the dominant contributors. PhysioScore calibration against 200 clinician labels yielded ECE = 0.038 and kappa = 0.91. Conclusion: Training-time physiological grounding produces measurable and independently verifiable safety improvements in open-weight clinical LLMs under controlled evaluation. External validation on real clinical data is required to determine whether these gains transfer to deployment settings

I. INTRODUCTION

Clinical LLMs may produce fluent, factually plausible recommendations that violate physiology, motivating training-time alignment with structured physiological knowledge. The framework uses an explicit world model to generate preferences and evaluates safety with a controlled benchmark.

  • Clinical LLMs can produce medically plausible recommendations that violate physiology, including harmful drug combinations and homeostatic constraints.
  • Factual correctness alone is insufficient because safe recommendations must also fit the implied patient state and context-dependent contraindications.
  • Neurosymbolic Alignment scores candidate responses with homeostatic constraints, multi-hop plausibility, and drug-interaction penalties before iterative on-policy ORPO updates.
  • The HGNN scorer is used during alignment but omitted at inference, simplifying deployment while increasing alignment-time computation.
  • The Clinical Safety Benchmark contains 2,500 synthetic clinical scenarios designed to test physiological constraint adherence under controlled, reproducible conditions.

III. REFERENCE BASELINES AND SCORING

The alignment pipeline generates candidate responses, scores their physiological feasibility on a heterogeneous graph, and uses the resulting rankings for iterative policy updates. The scorer combines homeostatic, relational, and drug-interaction signals into a safety-oriented feasibility score.

  • The LLM samples candidate responses, the HGNN scores them over the Physiological World Model, and Iterative ORPO updates the policy.
  • Candidate responses are entity-linked with scispaCy and UMLS before HGNN feasibility scoring on the biomedical graph.
  • At inference, the final policy runs without the HGNN, while iterative training regenerates candidates and preference pairs from the current policy.
  • The physiological scorer combines homeostatic boundary functions, multi-hop paths up to length 3, and curated drug-interaction relations.
  • A multiplicative score makes unsafe values on any component dominant, so a severe drug-interaction penalty can drive the overall score toward zero.

IV. DATASET CHARACTERIZATION AND EVALUATION PROTOCOL

The evaluation uses multiple medical benchmarks alongside CSB, a synthetic safety benchmark with leakage controls and structured scenario metadata. The protocol also tests graph dependence and novelty handling.

  • The evaluation suite spans clinical reasoning, biomedical evidence synthesis, medical knowledge, drug interactions, and clinical concept extraction.
  • CSB contains 2,500 clinical scenarios specifically designed to test adherence to physiological constraints.
  • The protocol separates train, validation, and test data using clinician-authored tests, temporal novelty, overlap filtering, and withheld graph entities.
  • The temporal split is a novelty stress test rather than a full simulation of routine clinical deployment.
  • When answer-critical graph relations are absent, MA is 72.8% versus 76.1% when present, a 3.3% gap suggesting performance is not explained by direct lookup alone.

C. Baseline Methods

The evaluation compares the proposed approach with supervised, preference-optimization, retrieval, rule-based, self-correction, and proprietary RLHF alternatives using safety, accuracy, and clinical-utility metrics. The study emphasizes open-weight 7–8B models, decoupled safety checks, and calibrated physiological-feasibility scoring.

  • Experimental setting: Primary training experiments use open-weight 7–8B models because Iterative ORPO requires repeated on-policy sampling and full parameter access.Closed proprietary models are used only as inference-time reference comparators.
  • Baseline coverage: The baseline suite spans supervised learning, preference optimization, knowledge enhancement, rule-based guardrails, self-correction, and proprietary RLHF alternatives.These comparisons include retrieval-augmented, static-KG, preference-only, and inference-time methods.
  • Evaluation metrics: The evaluation reports accuracy, safety, and clinical-utility metrics, including CSS, HR, DID, RSS, and PC.CSS measures responses without physiological constraint violations; HR is physician-evaluated, DID measures contraindicated-drug detection, RSS uses a deterministic rule engine, and PC is HGNN-computed.
  • Metric calibration: τ = 0.85 defines a CSS NoViolation response from the HGNN feasibility score, calibrated on 200 clinician-labeled responses with κ = 0.91.Bootstrap analysis gives a 95% threshold interval of [0.83, 0.87], within which CSS varies by ≤1.2 pp.
  • Scorer decoupling: RSS provides a structurally independent safety endpoint because the DrugBank-Rule checker shares no parameters, training signal, or learned representations with the HGNN scorer.Its coverage is narrower than CSS because it checks catalogued contraindications and predefined homeostatic thresholds rather than learned multi-hop interactions.
  • Scorer decoupling: The reported decoupled gains over ORPO are RSS: +21.2 pp, DID: +18.8 pp, and HR: −9.0 pp, with CSS–RSS concordance of r = 0.97 and κ = 0.82.HR is reported only for the blinded physician-evaluated subset.

E. Statistical Analysis

Across controlled comparisons, training-time physiological grounding improves safety and utility over ORPO, retrieval, self-correction, and other baselines, with statistically significant primary comparisons. Residual violations remain, including moderate-to-severe cases and errors concentrated in several failure categories.

  • Statistical testing: All reported primary comparisons remain significant after Bonferroni correction, with p < 10^-4 across 35 comparisons.Experiments use five random seeds and report mean ± standard deviation; Cohen’s d is descriptive.
  • Comparative results: 27.6 pp CSS, 23.1 pp DID, and 7.4 pp MA improvements over SFT+RAG indicate gains beyond retrieval alone.
  • Comparative results: 21.3 pp CSS, 18.8 pp DID, and 9.0 pp HR improvements over ORPO are reported on the corresponding evaluation measures.HR is measured on the physician-evaluated subset; the gains use automatically generated HGNN-derived preferences.
  • Baseline comparisons: 11.4 pp CSS separates the proposed method from SFT+SelfCorrect, while GPT-4 (5-Shot) reaches 80.6% CSS and 78.2% RSS.The proposed 7B model remains stronger on all safety metrics, while GPT-4 retains a 3.9 pp MA advantage.
  • Residual violations: 9.2% of responses remain below the safety threshold, including 3.9% moderate-to-severe violations and 5.3% minor deviations.The residual-error breakdown is non-exclusive when failures have multiple causes.
  • Residual violations: 4.2% sparse graph coverage, 3.1% multi-system interaction complexity, and 2.8% missing temporal physiology account for 70% of unsafe responses.

B. Robustness and Novelty Generalization

The method retains substantial safety under synthetic EHR-style perturbations, while novelty splits and ablations identify remaining generalization challenges and the main contributors to performance.

  • Robustness stress tests: 84.2% CSS is retained under combined synthetic noise, a 6.6 pp degradation from clean-CSB performance.Perturbations combine noisy narratives, 30% entity masking, and ambiguous abbreviations.
  • Robustness stress tests: Ambiguous abbreviations produce the largest single-family degradation across methods, consistent with entity-linking fragility.
  • Novelty generalization: Novel drug combinations and sparse disease conditions remain meaningful residual-error sources under novelty splits.
  • Ablation analysis: Removing HGNN scoring reduces CSS by 16.2 pp, while removing iterative training reduces it by 11.5 pp.These are the two largest ablation degradations.
  • Cross-domain results: Across six domains, average CSS improves by 21.3 pp over ORPO, with the largest gains in geriatric and pediatric scenarios.
  • Iteration dynamics: Performance improves monotonically across iterations, with diminishing returns after iteration 4 and stable multi-objective convergence.

F. Extended Technical Validation

Extended validation combines independent endpoints, calibration, robustness evidence, and deployment-oriented analysis. The results support controlled-benchmark safety gains but leave external transfer, graph maintenance, cost, and runtime verification as boundaries.

  • Evidence independence: Six of twelve key results use HGNN-independent or physician-adjudicated endpoints, supporting triangulation beyond HGNN-derived metrics.The independent measures include RSS, DID, HR, clinician kappa, and CSS–RSS concordance.
  • Implication: Training-time physiological grounding improves measured safety while maintaining diagnostic utility and targets physiologically invalid generations more directly than text-only supervision.
  • External validity: 84.2% CSS under synthetic noise provides a proxy for input variability, not external validation on real clinical data.CSB is synthetic and should be interpreted as a controlled methodological benchmark.
  • Computational cost: 42.6 GPU-hours on 8×A100s are required for iterative ORPO with K=8 over five iterations, roughly four times single-pass ORPO cost.Candidate generation accounts for 70,000 forward passes; HGNN scoring takes 2.0 hours.
  • Knowledge-graph limitations: Static knowledge-graph coverage gaps and temporal staleness contribute to residual failures, with coverage gaps accounting for 4.2% of test-case errors.Updating the graph requires full HGNN retraining and alignment; marginal per-model update cost is 31.6 hours.
  • Runtime safety: Removing the HGNN at inference simplifies deployment but leaves no runtime verifier for violations not anticipated during training; 9.2% residual violations remain.
  • Evidence independence: CSS and RSS correlate at r = 0.97, while three independent endpoints—RSS, DID, and HR—corroborate the CSS conclusions.Agreement is strong but imperfect, with Cohen’s κ = 0.82 at respective binary thresholds.

C. Future Directions

Future work targets knowledge-graph maintenance, efficient alignment, selective verification, broader evaluation, and eventual external validation. The paper frames these directions as responses to current benchmark, compute, and deployment limitations.

  • External validation on real clinical notes, expanded blinded clinician adjudication, and matched-scale PPO/RLHF comparisons are identified as necessary next steps.
  • Smaller and efficient model variants: Alignment distillation and post-alignment quantization are proposed to adapt the 7–8B framework for resource-constrained deployment.
  • Hybrid train-plus-verify architecture: A selective verifier would gate outputs using a lightweight PhysioScore probe and route uncertain responses to HGNN or deterministic rule-engine checks.The full verifier is intended to process only the uncertain tail, keeping average latency near the unverified baseline.
  • The framework achieves strong controlled-benchmark safety while retaining limitations from synthetic data, incomplete graph coverage, missing temporal physiology, and absent matched-scale PPO/RLHF evaluation.

APPENDIX A CHALLENGES OF NEUROSYMBOLIC ALIGNMENT IN

The appendix identifies representation, extraction, rule-encoding, scalar-scoring, and on-policy co-adaptation challenges that can separate graph consistency from clinical safety. These limitations motivate independent evaluation and richer clinical state representations.

  • Knowledge coverage debt arises because rapidly evolving clinical domains can outpace graph curation, leaving the HGNN to reason over outdated snapshots.
  • Canonical entity mapping can erase clinically relevant distinctions, so graph-consistent reasoning may remain clinically misleading.Renal failure can differ by acuity, assay standard, and trajectory toward or away from a threshold.
  • Entity-extraction errors involving negation, shorthand, and institution-specific abbreviations distort the represented clinical state before verification.Ambiguous identifiers such as “ASA” can make consistent downstream reasoning semantically incorrect.
  • Static rules and incomplete background conditions cannot fully represent context-dependent thresholds or defeasible constraints without trading utility against assumption risk.
  • Scalar safety scores can compress clinically distinct errors, while scorer calibration and proxy optimization may favor easy shortcuts over rare consequential failures.On-policy regeneration reduces stale preferences but may allow scorer–policy co-adaptation, reinforcing shared biases.
  • Snapshot reasoning cannot adequately model temporal clinical risk, so locally plausible recommendations may be unsafe when recent patient trajectories matter.

G. Retrieval, Tools, and Responsibility Boundaries

Retrieval and tool-augmented systems introduce provenance, attribution, topology, and deployment-readiness concerns beyond model-only safety scoring. Controlled CSB results therefore require cautious interpretation and broader safeguards.

  • Retrieval and tools: Retrieved evidence can vary in provenance and recency, allowing outdated recommendations to be treated as authoritative when newer guidance is absent.
  • Retrieval and tools: Tool pipelines require end-to-end auditability because failures may originate in model reasoning, APIs, query formulation, or incomplete databases.
  • Graph modeling: HGNN message passing can attenuate rare severe contraindications in dense subgraphs, while hub-based paths may score topology rather than mechanistic relevance.
  • Deployment boundaries: CSB reproducibility and interpretability do not establish deployment readiness because real clinical notes are longer, noisier, more ambiguous, and structurally heterogeneous.
  • Evaluation and metrics: Comparative claims are limited by unavailable alternatives, and metric design must balance unsafe specificity against clinically unhelpful abstention.
  • Patient representation: Patient-specific graph personalization may reproduce inequitable patterns, while missing covariates can make model and verifier errors systematically correlated.
  • Deployment boundaries: Deployment requires acuity-tiered triage, explicit uncertainty outputs, and traceable evidence paths rather than relying on controlled safety metrics alone.

APPENDIX C EXTENDED TECHNICAL VALIDATION (MOVED)

Extended validation examines scoring equations, calibration, graph and entity-linking sensitivity, component choices, parameter robustness, training progression, and compute-efficient configurations. Results support the multiplicative scorer while showing quality and budget trade-offs.

  • The HGNN feasibility score extracts and links entities, computes induced-subgraph embeddings, combines homeostatic and path terms, penalizes drug interactions, and returns a scalar score.
  • Compute sensitivity: Reducing K from 8 to 4 costs 2.1 pp CSS and approximately halves generation time, while K=4 and T=3 achieves 86.1% CSS at roughly one-third of full compute.
  • Knowledge-graph and linking sensitivity: A 10% random edge deletion reduces CSS by 2.2 pp, while 30% deletion reduces it by 9.4 pp; targeting high-severity edges causes a 4.7 pp drop at 10% volume.
  • Additional diagnostics: The extended analyses include PhysioScore calibration, combination-function sensitivity, threshold sensitivity, cross-domain generalization, and training statistics.
  • 90.8% CSS is achieved by the multiplicative PhysioScore combination, where a near-zero component strongly penalizes severe interactions.
  • Parameter sensitivity: Halving steepness values decreases CSS by 3.5 pp, doubling them decreases CSS by 1.6 pp, and ±20% perturbations decrease CSS by only 0.9 pp.
  • Training progression: Training progression shows the steepest CSS and DID improvements early, with diminishing gains after iteration 4 and continued improvement in inverted hallucination rate.

APPENDIX D IMPLEMENTATION DETAILS

The appendix documents reproducibility settings, computational resources, preference-scoring workload, and supporting dataset, benchmark, evaluation, and failure-analysis tables. It also illustrates HGNN-based constraint verification rejecting a contraindicated candidate plan.

  • Implementation: The framework uses open-source LLMs, with hyperparameters summarized in Table XVI.This supports reproducibility and broad applicability claims.
  • Computational Resources: Experiments used 8× NVIDIA A100 80GB GPUs, bfloat16 mixed-precision training, HuggingFace Transformers, and DeepSpeed ZeRO-3.Table XVII details the computational requirements.
  • Reproducibility Details: Entity extraction used scispaCy v0.5.1 and UMLS CUI linking with a 0.85 similarity threshold and up to five candidate CUIs per entity.Drug embeddings used radius-2, 2,048-bit Morgan fingerprints projected to an HGNN dimension of 256.
  • Computational Resources: Preference scoring consumed 0.4 GPU-hours per iteration to evaluate K = 8 candidates for each of 1,750 training queries and retain one preference pair per query.Aggregate memory reporting covers all eight GPUs and includes graph materialization and initial node-embedding computation.
  • Supporting Materials: The appendix includes tables covering dataset statistics, CSB composition, base-model comparisons, failure causes, human-evaluation criteria, and evaluation metrics.These materials are listed in Tables XVIII–XXIII.
  • Constraint Verification: A toy verification example links candidate-plan entities to the physiological graph and rejects the plan for hyperkalemia-related contraindications.Figure 6 presents the step-wise verification process.
Loading 2608.24534v1…