PaperlayerView exact revision on arXiv

Source-linked AI summary

How Compliant is Sepsis Treatment? An Expert-Guided Neuro-symbolic Pipeline for Generating Clinical Compliance Insights

Himanshu Tripathi, Kaushik Roy, Subash Neupane, Shahram Rahimi

arXiv:2608.13617v1cs.AIcs.SC

TL;DR

Assessing sepsis-care compliance is difficult because clinical records are unstructured and neither symbolic rules nor neural models alone suffice. The paper combines expert-guided semantic normalization with fuzzy reasoning, finding that immediate Hour-1 interventions were the dominant failure mode.

  • Problem

    Sepsis compliance assessment lacks a safe, interpretable approach that handles inconsistent clinical data while evaluating adherence to evidence-based protocols.

  • Method

    An expert-guided neuro-symbolic pipeline normalizes messy clinical text and applies expert-validated fuzzy rules to produce graded sepsis-compliance scores.

  • Results

    Median compliance was 0.00 for the Hour-1 bundle, compared with 0.67 for resuscitation and 1.00 for treatment response.

  • Takeaways & Limitations

    The pipeline provides interpretable, graded sepsis-compliance assessments and can extend to other protocols when guidelines and expert-defined boundaries are available.

  • Takeaways & Limitations

    Mortality was outside the study’s scope, which focused on measuring alignment with sepsis-care protocols.

Abstract

from arXiv · show

Verifying whether clinical care follows evidence-based protocols is a natural neuro-symbolic problem, yet the safety-critical setting defeats either paradigm alone. We present an expert-guided pipeline that constrains a large language model strictly to semantic normalization, mapping messy drug and microbiology strings onto a fixed clinical vocabulary, while a Sugeno fuzzy inference system reasons over the normalized events. The fuzzy layer encodes eight Surviving Sepsis Campaign bundle rules and replaces binary judgments with graded scores in [0,1]. Applied to 2,438 MIMIC-IV v3.1 sepsis episodes, it surfaces antibiotic timing as the most critical breakdown (mean 0.24, 13% within one hour), Hour-1 underperformance (mean 36.7%), a 51% elevated-lactate drop-off, and descriptive differences in ICU stay across compliance groups (3.8 versus 5.1 days).

1. Introduction

Sepsis care requires timely adherence to evidence-based SSC bundles, yet conventional rule-based and neural approaches struggle with semantic variation, interpretability, and safety guarantees. The proposed expert-guided neuro-symbolic pipeline combines LLM semantic normalization with an expert-validated fuzzy system to generate graded compliance scores.

  • Motivation: Timely adherence to evidence-based Surviving Sepsis Campaign bundles is critical for patient survival, while MIMIC-IV enables large-scale protocol-compliance evaluation.The introduction identifies sepsis as a leading cause of in-hospital mortality.
  • Motivation: Traditional rule-based systems fail on semantic variations, whereas pure neural networks produce uninterpretable risk scores that cannot guarantee safety-protocol adherence.The passage frames these limitations as a critical gap in compliance assessment.
  • Contribution: The expert-guided neuro-symbolic pipeline uses MedGemma for semantic normalization and an expert-validated Fuzzy Inference System for graded rather than binary compliance scores.The pipeline addresses limitations of existing paradigms by separating normalization from expert-guided reasoning.

2. Literature Survey

Sepsis care depends on timely adherence to Surviving Sepsis Campaign bundles, yet compliance remains suboptimal. Existing clinical NLP and reasoning approaches face semantic ambiguity, hallucination risk, vocabulary brittleness, or limited interpretability, motivating constrained neuro-symbolic designs.

  • Clinical motivation: Sepsis remains a leading cause of ICU mortality, making timely adherence to evidence-based protocols critical for patient survival.The Surviving Sepsis Campaign bundles are described as the gold standard for reducing mortality.
  • Clinical motivation: Surviving Sepsis Campaign bundle compliance remains suboptimal across healthcare settings, increasing the importance of reliable assessment.Electronic Health Records such as MIMIC-IV provide opportunities for evaluating compliance at scale.
  • Methodological gap: Large language models help resolve semantic ambiguity in free-text clinical data but introduce hallucination risk, making autonomous deployment unsafe for critical interventions such as antibiotic administration.Relevant challenges include negation, uncertainty, and assertion detection in clinical notes.
  • Proposed direction: Hybrid architectures constrain LLMs to semantic normalization while delegating decisions to deterministic fuzzy logic safety boundaries.This pairing addresses the limitations of autonomous LLM reasoning in safety-critical clinical environments.
  • Methodological gap: Traditional symbolic systems suffer vocabulary brittleness, whereas pure neural networks remain uninterpretable black boxes for complex sepsis inputs.The cited examples include “Zosyn” and 64 mmHg blood pressure.

3. Pipeline Architecture

The pipeline combines centralized semantic normalization, expert validation, and Sugeno fuzzy reasoning to generate graded compliance scores for sepsis protocols. Its architecture preserves timestamps, validates classifications with independent embeddings, and supports domain-general transfer through expert-defined fuzzy decision boundaries.

  • Architecture: The domain-general pipeline integrates semantic normalization, expert validation, and fuzzy reasoning, instantiated on MIMIC-IV sepsis data with expert-defined decision boundaries.The architecture transfers to any protocol expressible as fuzzy rules when a domain expert sets the boundaries.
  • Cohort construction: 2,438 sepsis episodes were sampled from 17,926 patients, 22,363 hospitalizations, and 17,152 ICU stays.Random sampling used seed 55 and selected 2,000 patients.
  • Semantic normalization: 1,691 unique drug strings and 650 microbiology combinations were classified centrally, then mapped across episodes using regex exclusion rules and LLM normalization.The system handles trade names, generic names, abbreviations, tall-man spellings, and varied specimen labels.
  • Validation: 94.26% of drug strings and 60.62% of microbiology combinations received agreeing classifier labels, while blood-culture detection reached 95.7% (622/650).The classifiers agreed on 1,594 of 1,691 drug strings and 394 of 650 microbiology combinations.
  • Fuzzy reasoning: Eight SSC rules are scored by a Sugeno fuzzy system using expert-recommended half-Gaussian memberships, preserving graded treatment effects across timing, value, and conditional-trigger rules.For example, treatment at 90 minutes yields µ = 0.61 and at 120 minutes µ = 0.13; mandatory Rules 1 to 3 receive µ = 0 when absent.

4. Quantitative Findings with Expert Discussion

The findings identify early sepsis management—especially antibiotic timing—as the dominant compliance weakness, while downstream bundle actions remain sparse and compliance varies across clinical phases. Descriptive results also show classifier agreement is high, ICU stays differ across compliance groups, and repeated admissions yield little improvement.

  • Classifier Agreement: 94.26% of drug classifications agree between regex and MedGemma, so agreeing episodes produce identical deterministic fuzzy compliance scores.Only 80 episodes (3.3%) diverged, according to the supplied passage.
  • Bundle Rules: µ = 0.24 and 13% meeting the threshold make antibiotic timing (R2) the lowest-scoring rule under the predefined one-hour operationalization.The threshold is cohort-level and does not stratify timing by septic-shock status.
  • Compliance Cascade: 43% of episodes activate fluid resuscitation (R5), while 34% activate lactate clearance, reflecting sparsity in downstream conditional rules.The pipeline conservatively treats both recorded cases as non-compliance, consistent with SSC guidelines.
  • Clinical Outcomes: 2.95 days versus 4.74 days is the median ICU-stay contrast for antibiotics given within 30–60 minutes versus after six hours.Across overall compliance groups, low compliance corresponds to a median stay of 5.1 days, whereas the high-compliance group averages 3.8 days.
  • Clinical Phases: Median compliance rises from 0.00 in Hour-1 to 0.67 in resuscitation and 1.00 in treatment response, identifying immediate interventions as the dominant failure mode.Among 2,000 patients, 14% experienced recurrent sepsis, with compliance increasing only from 0.362 to 0.375 (+1.3%) between episodes.

5. Conclusion

The paper presents an expert-guided neuro-symbolic pipeline that combines semantic normalization, fuzzy reasoning, and domain expertise to produce interpretable, graded sepsis-compliance assessments. It identifies systemic Hour-1 compliance failures and descriptive ICU length-of-stay differences across compliance and antibiotic-timing groups.

  • 5. Conclusion: The pipeline combines semantic normalization, fuzzy reasoning, and domain expertise for interpretable, graded sepsis-care compliance assessments.Its design integrates neuro-symbolic processing with clinical expertise.
  • 5. Conclusion: Systemic Hour-1 compliance failures emerged alongside descriptive ICU length-of-stay differences across compliance and antibiotic-timing groups.These findings were reported in the Results and Comparison analysis.

Appendix A. Data Availability and Ethical Statement

The study uses the de-identified MIMIC-IV v3.1 critical care database from Beth Israel Deaconess Medical Center, available through PhysioNet under credentialed access. The database spans 2008–2022 and covers approximately 250,000 patients across 500,000 hospitalizations.

  • MIMIC-IV v3.1 is a de-identified critical care database sourced from Beth Israel Deaconess Medical Center electronic health records.
  • 2008–2022 is the database coverage period, encompassing approximately 250,000 patients across 500,000 hospitalizations.
  • Credentialed access through PhysioNet requires completion of a recognized human subjects research training program.

Appendix B. Eight Rules From SSC · Appendix C. Notation and Symbols for Problem Statement · Appendix D. Problem Statement

The appendices define eight SSC compliance rules, document notation for the problem statement, and motivate an expert-guided neuro-symbolic pipeline that normalizes clinical language before fuzzy compliance reasoning. The approach produces graded, actionable scores across 2,438 MIMIC-IV sepsis episodes while constraining the language model to semantic normalization.

  • Appendix B. Eight Rules From SSC: The SSC assessment is governed by eight rules covering blood cultures, antibiotics, lactate monitoring, fluids, vasopressors, MAP, and lactate clearance.The rules preserve pathogen identification, initiate broad-spectrum coverage, assess and reassess hypoperfusion, restore circulation, support perfusion pressure, target MAP, and verify metabolic response.
  • Appendix B. Eight Rules From SSC: The rule set includes lactate reassessment above 2 mmol/L, 30 mL/kg crystalloid for hypotension or lactate ≥4 mmol/L, vasopressors to maintain MAP ≥65 mm Hg, and an initial MAP target of 65 mm Hg.These thresholds operationalize reassessment, fluid resuscitation, vasopressor use, and perfusion-pressure targeting.
  • Appendix B. Eight Rules From SSC: The 10% lactate Rule 8 clearance threshold derives from Jones et al. (2010), not directly from the SSC bundle.This note distinguishes the operational threshold from the underlying guideline recommendation to decrease serum lactate.
  • Appendix C. Notation and Symbols for Problem Statement: Table 1 defines the notation and symbols used to formulate the problem statement.The notation appendix provides the formal symbolic context for the subsequent pipeline formulation.
  • Appendix D. Problem Statement: The SSC rules specify time-sensitive actions including antibiotic administration, fluid resuscitation, lactate measurement, and blood cultures.Interchangeable trade names, generic names, abbreviations, and heterogeneous microbiology events make systematic evaluation difficult.
  • Appendix D. Problem Statement: The proposed pipeline constrains the language model to semantic normalization because its grounding probability is otherwise undefined and hallucination risk is unquantified.The model maps variable clinical expressions without autonomous decision-making.
  • Appendix D. Problem Statement: The fuzzy inference system applies expert-validated decision boundaries to episodes and produces graded compliance scores for actionable healthcare insights across 2,438 MIMIC-IV sepsis episodes.This combines neural normalization with transparent symbolic reasoning rather than relying on autonomous language-model decisions.

Appendix E. Data Overview … Appendix I. Additional Graphs

The appendices document the MIMIC-IV data foundation, expert-guided fuzzy-rule design, classifier agreement, and cohort-level compliance patterns. Across 2,438 sepsis episodes, the figures emphasize low overall and early-treatment compliance, conditional intervention burden, documentation gaps, and ICU-stay associations.

  • Appendix E. Data Overview: MIMIC-IV v3.1 spans 2008–2022 and contains approximately 250,000 patients across 500,000+ hospitalizations in hospital and ICU modules.The modules share identifiers that support integration across administrative, clinical, and intensive care data layers.
  • Appendix F. Domain Expert Consultation: Subject Matter Experts shaped the pipeline by recommending Gaussian membership functions, immunocompromised-status organism reclassification, route-specific exclusions, and rule priority ordering.Only IV/IM antibiotics within time windows count toward SSC compliance; the priority order is Rule 3: Lactate > Rule 2: Antibiotics > Rule 5: Fluids > Rule 6: Vasopressors > remaining rules.
  • Appendix G. Fuzzy Membership Function Parameter: Expert-validated normalization rules and fuzzy parameters produce graded compliance assessments through a Sugeno fuzzy inference system.The eight SSC rules use expert-established boundaries and decay rates, with conditional rules activating only when their clinical triggers are met.
  • Appendix H. Drug classification Confusion Matrix: Across 1,691 unique drug strings, regex and MedGemma agreed on 1,594 cases, yielding Po = 0.9426 and Cohen’s κ = 0.65.Their complementary errors included antibiotics, vasopressors, and IV fluids captured by one classifier but missed by the other.
  • Appendix I. Additional Graphs: 36.7% mean overall compliance was observed across 2,438 sepsis episodes, with a median of 37.5% and concentration in lower performance tiers.The distribution reflects execution challenges across the complete Surviving Sepsis Campaign bundle.
  • Appendix I. Additional Graphs: 13% of episodes met the strict 1-hour SSC antibiotic target, indicating a severe operational bottleneck in initial antibiotic administration.The figure overlays antibiotic timing with the fuzzy compliance membership function.
  • Appendix I. Additional Graphs: 43% of episodes triggered fluid resuscitation, while missing documentation affected antibiotic timing in 31.6% and initial lactate measurements in 24.4%.Fluid resuscitation was the most frequently activated conditional intervention, and sparsity was greatest during early care.
  • Appendix I. Additional Graphs: Phase 1 median compliance was 0.00 versus 1.00 in Phase 3, while recurrent patients showed median compliance increasing from 0.362 in the first episode to 0.375 in the second.The phase comparison highlights greater difficulty executing rapid initial interventions than downstream resuscitation and response measures.

Appendix J. Comparison Analysis

The comparison analysis contrasts SSC bundle compliance gap papers and highlights per-episode actionable insights supported by domain-expert validation. It also describes SME consultation within the pipeline for setting boundaries, prioritizing rules, and resolving edge cases.

  • Table 4 compares papers addressing SSC bundle compliance gaps.
  • Per-episode scores enable individual and population-level quality improvement beyond cohort-level analysis.
  • SME consultation supports boundary setting, rule priority ordering, and edge case resolution mid-pipeline.

Appendix K. Combined Results

Appendix K presents comprehensive results from an expert-guided neuro-symbolic pipeline for sepsis compliance assessment, using data from 2,438 sepsis episodes across 2,000 patients in MIMIC-IV v3.1.

  • Appendix K. Combined Results: The appendix reports comprehensive results for an expert-guided neuro-symbolic pipeline assessing sepsis compliance.The results are presented in Table 5.
  • Appendix K. Combined Results: 2,438 sepsis episodes comprise the analyzed dataset.The episodes span 2,000 patients.
  • Appendix K. Combined Results: The data come from MIMIC-IV v3.1.The dataset includes 2,438 sepsis episodes across 2,000 patients.

Appendix L. Research Question Summary

The expert-guided neuro-symbolic pipeline combines hybrid semantic normalization with Sugeno fuzzy reasoning to produce reliable, graded sepsis-compliance assessments. It identifies antibiotic timing and elevated-lactate measurement as major weaknesses and associates higher compliance with shorter ICU stays.

  • RQ1: 80 episodes (3.3%) produced divergent dual-path compliance scores, including 53 with differences exceeding 5%, showing that neither symbolic nor neural components alone were reliable across all episodes.Regex verified 220 drug strings, while MedGemma resolved remaining lexical variation and fuzzy inference enforced SSC safety boundaries.
  • RQ3: µ = 0.24 for antibiotic timing (R2) marked the most critical failure, while MAP recovery (R7) reached µ = 0.97 as near-acceptable.An expert attributed low R2 scores to pre-ICU antibiotic administration and cautioned that conditional-rule performance may reflect survivorship bias.
  • RQ4: 51% patient drop-off occurred at the elevated lactate threshold, while antibiotic timing and lactate measurement were the most data-sparse Hour-1 interventions.The compliance cascade exposed where documented adherence most sharply declined across the sepsis cohort.
  • RQ4: 3.8 days versus 5.1 days was the median ICU stay for higher- versus low-compliance episodes, while antibiotic-timing stays rose from 2.95 days at 30–60 minutes to 4.74 days beyond six hours.These descriptive cohort patterns connect bundle compliance and antibiotic timing with ICU length of stay.
Loading 2608.13617v1…