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

Himanshu Tripathi; Kaushik Roy; Subash Neupane; Shahram Rahimi

cs.AIAug 12, 2026

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

Authors: Himanshu Tripathi, Kaushik Roy, Subash Neupane, Shahram Rahimi

Organizations: Department of Computer Science, The University of Alabama

Abstract

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).

Figures & tables

Figure 1: The end-to-end workflow of the Expert-Guided Neurosymbolic pipeline, illustrating the transformation of sepsis specific raw clinical data into healthcare insights via LLM semantic normalization, expert knowledge, and fuzzy logic.

Figure 2: Comparison of sepsis detection paradigms handling complex inputs like ”Zosyn” and 64 mmHg blood pressure. Traditional symbolic systems fail from vocabulary brittleness , while pure neural networks fail as uninterpretable black boxes. The proposed Neuro-Symbolic approach succeeds by pairing LLM semantic normalization with deterministic fuzzy logic safety boundaries.

Figure 3: The neuro-symbolic pipeline normalizes clinical text using a dual regex and LLM approach. Following embedding-based validation and domain expert refinement of decision boundaries, a Fuzzy Inference System generates graded compliance scores.

Figure 4: Membership functions for all eight SSC bundle rules. Timing-based interventions (Rules 1-3, 6) use Boolean or right-side half-Gaussian decay, while value-based targets (Rules 7-8) utilize left-side half-Gaussians. Rule 4 applies a window penalty, Rule 5 combines volume and time in a 2D contour, and purple badges mark conditional triggers.

Figure 5: A scatter plot comparing Regex and MedGemma compliance scores, highlighting episodes with divergent classifications.

Figure 6: Bar and radar charts displaying average fuzzy compliance scores across all eight bundle rules.

Figure 7: A compliance cascade illustrating patient drop-offs across the sequential sepsis management bundle steps.

Figure 8: Descriptive comparison of ICU length of stay across overall bundle-compliance groups and antibiotic-timing categories.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Definition
$v$	Documented drug variant in clinical records
$V$	Predefined synonym set in rule-based system
$\mathcal{L}$	Large Language Model for semantic normalization
$\text{input}_{m}$	Clinical prompt supplied to $\mathcal{L}$
$S_{k}$	Sepsis episode $k$
$\mathbb{P}$	$\mathbb{P}(E)$ denotes probability of event $E$

Appendix

Table 1: Notation and Symbols used to formulate the Problem Statement

Figure 9: MIMIC-IV Database Structure and Relational Organization. The database comprises two modules: MIMIC-IV/hosp maintaining hospital-level records indexed by subject_id , hadm_id , and pharmacy_id , and MIMIC-IV/icu (9 tables, bottom portion) maintaining ICU-specific records indexed by stay_id . Primary linking identifiers ( subject_id , hadm_id , stay_id ) enable integration across administrative, clinical, and intensive care data layers. The pipeline extracts data from emar , prescriptions , microbiologyevents , labevents , inputevents , chartevents , and icustays .

Rule	Target	MF Type	Parameters	Rule Weight
Phase 1: Hour-1 Bundle
R1	Culture before abx	Boolean step	$\mu{=}1$ if $t_{\text{cx}}{<}t_{\text{abx}}$ , else 0	0.5
R2	Abx $\leq$ 1 hr	Right HG	$c{=}60$ min, $\sigma{=}30$	0.9
R3	Lactate $\leq$ 1 hr	Right HG	$c{=}60$ min, $\sigma{=}30$	1.0
R4 †	Re-lactate 2–4 hr	Window HG	$c_{1}{=}120$ , $c_{2}{=}240$ min	0.5
	lac ${}_{1}{\geq}2.0$		$\sigma_{e}{=}30$ , $\sigma_{l}{=}60$

Appendix

Table 2: Membership function parameters for the eight SSC bundle compliance rules. Boundaries (

c

) and decay rates (

\sigma

) were established through domain expert consultation. Rules marked † are conditional and activate only when the clinical trigger is met.

Property	Value
Inference type	Sugeno first-order
Defuzzification	$\sum_{r\in R_{i}}R_{r}z_{r}\,/\,\sum_{r\in R_{i}}R_{r}$
Conditional untriggered	$\mu=1.0$ , excluded from $R_{i}$
Missing mandatory (R1–R3)	$\mu=0$
Culture only, no abx (R1)	$\mu=0.5$
No weight available (R5)	$\mu_{5v}=0.5$ penalty

Appendix

Table 3: Sugeno fuzzy system properties and missing data handling. Priority weights encode the SME-established ordering

Rule_{3}{>}Rule_{2}{>}Rule_{5}{>}Rule_{6}{>}Rule_{1}{=}Rule_{4}{=}Rule_{7}{=}Rule_{8}

Figure 10: Drug classification confusion matrix between regex-based and MedGemma classifiers across 1,691 unique drug strings. Diagonal cells (orange borders) represent agreement ( $P_{o}=1{,}594/1{,}691=0.9426$ ). Off-diagonal entries reveal complementary error profiles: regex misses 32 antibiotics, 4 vasopressors, and 9 IV fluids that MedGemma captures, while MedGemma misses 15 antibiotics, 22 vasopressors, and 15 IV fluids that regex captures. Cohen’s $\kappa=0.65$ (substantial agreement). All off-diagonal cells within the clinical-category submatrix (antibiotic, vasopressor, IV fluid) are zero, confirming that disagreements occur exclusively between a clinical category and other , never across clinical categories.

Figure 11: Overall distribution of fuzzy compliance scores across the cohort of 2,438 sepsis episodes. The histogram is color-coded by compliance quartile, revealing a heavy concentration of episodes in the lower performance tiers. The mean overall compliance score is 36.7% (median 37.5%), highlighting significant systemic challenges in executing the complete set of Surviving Sepsis Campaign (SSC) bundle guidelines in clinical practice.

Figure 12: Distribution of time from sepsis onset to initial antibiotic administration, overlaid with the fuzzy compliance membership function

\mu(x)

(orange line). A severe operational bottleneck is evident, as only 13% of the observed episodes met the strict 1-hour SSC target (red dashed line).

Figure 13: Proportion of episodes triggering the conditional rules (R4 through R8) of the SSC bundle. This stacked chart visualizes clinical acuity; fluid resuscitation (R5) was the most frequently activated conditional intervention (43% of episodes), reflecting a high prevalence of shock or severe hyperlactatemia in the cohort.

Figure 14: Prevalence of missing clinical documentation across the eight SSC bundle components. Data sparsity is most severe during the critical early stages of care, specifically regarding the exact timing of antibiotic administration (R2, 31.6%) and initial lactate measurements (R3, 24.4%).

Figure 15: Box plots illustrating the distribution of fuzzy compliance scores categorized by clinical care phase: Phase 1 (Hour-1 Bundle), Phase 2 (Resuscitation), and Phase 3 (Response). The data demonstrates a significant increase in median compliance from Phase 1 (0.00) to Phase 3 (1.00). This high variance highlights the operational difficulty of executing rapid, immediate interventions in the initial acute phase versus achieving compliance in downstream resuscitation and response metrics.

Figure 16: This scatter plot evaluates the correlation between overall fuzzy compliance scores and Intensive Care Unit Length of Stay (ICU Length of Stay (LOS)). The orange trend line demonstrates that higher bundle compliance leads to a reduction in ICU time . Specifically, the analysis shows a trend of -0.006 days of ICU stay saved for every 1% increase in compliance , based on a correlation of

r=-0.026

across

n=2,362

episodes.

Figure 17: Violin plots displaying ICU Length of Stay distributions, grouped by the specific bundle rule that yielded the lowest compliance score per episode. Failures in early interventions such as delayed antibiotics (R2, median 7.0 days) and delayed initial lactate (R3, median 5.0 days) are strongly associated with prolonged ICU stays.

Figure 18: The left panel illustrates the distribution of sepsis episodes per patient across the study cohort (

N=2,000

patients, 2,438 total episodes). While the majority of patients (86%) experienced a single episode, 14% presented with recurrent sepsis. The right panel evaluates whether clinical compliance improves in subsequent episodes for recurrent patients. A marginal improvement in median compliance is observed, increasing from 0.362 in the first episode to 0.375 in the second and third+ episodes, representing a +1.3% gain in median performance between the first two clinical encounters.

Parameter	Gao et al. (2005)	Damiani et al. (2015)	You et al. (2022)	Our Work
Type of Study	Prospective observational study	Systematic review & meta-analysis	Retrospective observational cohort	SME guided computational pipeline
SSC Bundle Gaps Identified	Non-compliance with 6-hr and 24-hr bundles linked to higher hospital mortality	Suboptimal compliance across healthcare settings despite performance improvement programs	Bundle adherence varies significantly by ED admission time (off-hours worse)	Hour-1 bundle mean compliance of only 36.7%; antibiotic timing (R2) worst at 0.24 mean fuzzy score
Compliance Measurement Method	Binary — compliant vs. non-compliant per bundle window	Binary — pooled compliance rates across 50+ studies	Binary — adhered/not-adhered per bundle element	Graded fuzzy scores (0–1) per each of 8 SSC rules via Sugeno FIS
Granularity of Gap Analysis	Bundle-level (6-hr vs. 24-hr)	Program-level across institutions	Admission-time-level (day vs. night shift)	Per-rule, per-episode, per-phase (Hour-1, Resuscitation, Response) across 2,438 episodes
Automated at Scale	Manual prospective data collection	Manual aggregation of existing studies	Manual retrospective review	Partially automated on (need to have SME)
Handles Unstructured EHR Text	Not addressed	Not addressed	Not addressed	MedGemma resolves trade names, abbreviations, and microbiology ambiguity

Appendix

Table 4: Comparison of SSC Bundle Compliance Gap Papers

STUDY COHORT		MISSING DATA
Metric	Value	Metric	Value
Total Sepsis Patients	2,000	R2: Antibiotic Timing	31.6%
Total Sepsis Episodes	2,438	R3: Lactate Timing	24.4%
Recurrent Sepsis (14%)	284	R1: Blood Culture	8.8%
Unique Drug Strings	1,691	R5: Fluids	6.7%
NORMALIZATION		CONDITIONAL ACTIVATION

Appendix

Table 5: Comprehensive Results from the Expert-Guided Neuro-Symbolic Pipeline for Sepsis Compliance Assessment. Data derived from 2,438 sepsis episodes across 2,000 patients from MIMIC-IV v3.1.

Explore similar work

Agentifying Patient Dynamics within LLMs through Interacting with Clinical World Model
May 14, 2026Minghao Wu, Yuting Yan, Zhenyang Cai +9Agentic RL Clinical Decision Support
The Chinese University of Hong Kong, Shenzhen · Beijing Hospital
Clinically Interpretable Sepsis Early Warning via LLM-Guided Simulation of Temporal Physiological Dynamics
Apr 22, 2026Weizhi Nie, Zhen Qu, Weijie Wang +4Interpretable ML LLM-Based Time Series Forecasting
School of Electrical and Information Engineering, Tianjin University, Tianjin, 300072, Tianjin, China · Department of Information Engineering and Computer Science, University of Trento, Trento, Trentino, Italy · Department of Orthopedics, Affiliated Kunshan Hospital of Jiangsu University, Kunshan, Jiangsu, China +3
SepsisLens: Structure-Preserving Sequence Modelling for Decomposable Early Sepsis Warning
Oct 6, 2026Yikun Ou, Wei LiInterpretable ML Clinical Prediction
School of Computer Science, University of Sydney Sydney, Australia

May 14, 2026, cs.AIJ/K move · Enter open · S save

Agentifying Patient Dynamics within LLMs through Interacting with Clinical World Model

Minghao Wu, Yuting Yan, Zhenyang Cai, Ke Ji+8

The Chinese University of Hong Kong, Shenzhen · Beijing Hospital

Agentic RL Clinical Decision Support LLM Agent Training Medical LLMs

Open paper AskPDF arXiv

Sepsis management in the ICU requires sequential treatment decisions under rapidly evolving patient physiology. Although large language models (LLMs) encode broad clinical knowledge and can reason over guidelines, they are not inherently grounded in action-conditioned patient dynamics. We introduce SepsisAgent, a world model-augmented LLM agent for sepsis treatment recommendation. SepsisAgent uses a learned Clinical World Model to simulate patient responses under candidate fluid--vasopressor interventions, and follows a propose--simulate--refine workflow before committing to a prescription. We first show that world-model access alone yields inconsistent LLM decision performance, motivating agent-specific training. We then train SepsisAgent through a three-stage curriculum: patient-dynamics supervised fine-tuning, propose--simulate--refine behavior cloning, and world-model-based agentic reinforcement learning. On MIMIC-IV sepsis trajectories, SepsisAgent outperforms all traditional RL and LLM-based baselines in off-policy value while achieving the best safety profile under guideline adherence and unsafe-action metrics. Further analysis shows that repeated interaction with the Clinical World Model enables the agent to learn regularities in patient evolution, which remain useful even when simulator access is removed.

Abstract

Explore similar work

Agentifying Patient Dynamics within LLMs through Interacting with Clinical World Model

Clinically Interpretable Sepsis Early Warning via LLM-Guided Simulation of Temporal Physiological Dynamics

SepsisLens: Structure-Preserving Sequence Modelling for Decomposable Early Sepsis Warning