🤖 AI

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

This paper presents an expert-guided neuro-symbolic pipeline that combines large language models for semantic normalization with a Sugeno fuzzy inference system to evaluate clinical compliance with Surviving Sepsis Campaign protocols, revealing significant gaps in antibiotic timing and lactate monitoring across 2,438 sepsis episodes in the MIMIC-IV dataset.

Original authors: Himanshu Tripathi, Kaushik Roy, Subash Neupane, Shahram Rahimi

Published 2026-08-17

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Himanshu Tripathi, Kaushik Roy, Subash Neupane, Shahram Rahimi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to follow a complex recipe to bake a perfect cake, but the ingredients list is written in a chaotic mix of languages, slang, and handwritten notes. One person might write "flour," another "AP flour," and a third "white dust." If you try to read this list with a strict, old-fashioned rulebook, you might miss the flour entirely because it wasn't written exactly as "flour." On the other hand, if you ask a super-smart, creative robot to guess what the ingredients are, it might make up a "flour" that doesn't exist or confuse sugar for salt. This is the exact problem doctors face when trying to check if they are following life-saving medical rules. The medical world is full of messy, unstructured notes where the same drug might have a dozen different names. Scientists have long debated whether to use rigid computer rules (which are safe but brittle) or smart AI (which is flexible but can make dangerous mistakes). The big question is: Can we build a system that is both flexible enough to understand messy human language and strict enough to guarantee safety?

This paper introduces a clever solution called an "Expert-Guided Neuro-Symbolic Pipeline" to solve this puzzle for treating sepsis, a dangerous body-wide infection. The researchers built a two-part team to act like a detective and a judge. First, they used a Large Language Model (LLM)—a type of AI that is great at understanding language—as a "semantic normalizer." Think of this AI as a translator that turns messy, confusing drug names and lab notes into a clean, standard list that everyone agrees on. However, because AI can sometimes "hallucinate" or make things up, this translator is strictly supervised. It doesn't make the final call; it just cleans up the data. Second, they used a "Fuzzy Inference System," which acts like a wise, experienced judge. Instead of giving a simple "pass" or "fail" grade, this judge uses a sliding scale of scores from 0 to 1 to decide how well the treatment followed the rules. This system was tested on 2,438 sepsis episodes from a massive hospital database called MIMIC-IV.

The team discovered that while doctors are getting better at some parts of the treatment, they are struggling significantly with the very first hour of care. The system found that the most critical breakdown was in the timing of antibiotics. On average, the compliance score for giving antibiotics within the first hour was only 0.24 out of 1.0, meaning only about 13% of patients actually received their antibiotics within that crucial first hour. Another major drop-off happened with lactate testing; 51% of the time, the system couldn't confirm that a necessary blood test was done. Interestingly, the study suggests that once patients survive the initial chaotic hour and reach the ICU, doctors do a much better job with later steps like giving fluids or blood pressure medication, with scores jumping up to 0.73 or even 0.97. However, the authors warn that this high score might be misleading because it only looks at patients who lived long enough to get to that stage, a phenomenon known as "survivorship bias."

The most striking finding connects these missed opportunities to real-world outcomes. The study suggests that when the first-hour rules are followed, patients stay in the ICU for a shorter time. Specifically, patients in the high-compliance group stayed for a median of 3.8 days, while those in the low-compliance group stayed for 5.1 days. The data also showed that if antibiotics are given within 30 to 60 minutes, the median stay is just 2.95 days, but if they are delayed by more than six hours, the stay jumps to 4.74 days. The researchers conclude that their hybrid approach—combining the language skills of AI with the strict logic of fuzzy rules—successfully highlights these gaps without the black-box confusion of pure AI or the brittleness of old rule-based systems. They suggest this method could be adapted for other medical protocols, like stroke or heart care, provided experts are there to set the safety boundaries. However, they note that the system currently relies heavily on human experts to define the rules and doesn't yet measure whether following these rules actually saves lives, only that it shortens hospital stays.

Technical Summary: An Expert-Guided Neuro-Symbolic Pipeline for Sepsis Compliance

Problem Statement
Verifying adherence to evidence-based clinical protocols, specifically the Surviving Sepsis Campaign (SSC) bundles, is a critical challenge in healthcare. While Large Electronic Health Records (EHR) like MIMIC-IV offer vast data for evaluation, the data is deeply unstructured and inconsistent. Traditional rule-based systems fail due to semantic rigidity (e.g., inability to handle trade names like "Zosyn" vs. "Vancomycin"), while pure neural networks produce uninterpretable "black box" scores that cannot guarantee adherence to safety protocols. There is a need for a system that is both robust to noisy clinical text and interpretable regarding safety rules.

Methodology
The authors propose an Expert-Guided Neuro-Symbolic Pipeline that integrates three distinct components to bridge the gap between unstructured data and logical reasoning:

Semantic Normalization (Neural Component):
- The pipeline utilizes MedGemma (a 4-bit quantized LLM) to resolve messy clinical text, mapping trade names, abbreviations, and synonyms to a fixed clinical vocabulary.
- To mitigate hallucination risks, the LLM operates in a zero-shot, structured prompting mode but is strictly constrained to semantic normalization. It does not make clinical decisions autonomously.
- A Regex-based system with domain-informed exclusion rules runs in parallel to verify classifications against documented drug information.
- Validation Strategy: The two systems are combined because their errors are complementary. Regex provides high precision on explicit matches but low recall on synonyms; MedGemma generalizes well but risks hallucination. Disagreements are adjudicated using embedding-based validation and domain expert consultation.
Fuzzy Inference System (Symbolic Component):
- Normalized events are fed into a Sugeno fuzzy inference system that encodes eight SSC bundle rules across three clinical phases:
  - Phase 1 (Hour-1): Blood culture sequencing, antibiotic administration, and lactate measurement.
  - Phase 2 (Resuscitation): Fluid administration and vasopressor initiation (conditional on hypotension/hyperlactatemia).
  - Phase 3 (Response): Mean Arterial Pressure (MAP) recovery and lactate clearance.
- Graded Scoring: Unlike binary judgments, the system uses half-Gaussian membership functions to generate compliance scores in the range $[0, 1]$ . This reflects clinical reality where a treatment at 62 minutes is not categorically different from one at 58 minutes.
- Expert Guidance: Domain experts define the decision boundaries, membership function parameters, and rule priorities (e.g., prioritizing initial lactate measurement over culture sequencing).
Data Cohort:
- The system was applied to 2,438 sepsis episodes derived from the MIMIC-IV v3.1 dataset (sampled from 17,926 patients).
- The pipeline processes medications, microbiology, labs, vitals, and static profiles with minute-level timestamp alignment.

Key Contributions

Architectural Innovation: The paper presents a hybrid architecture where the LLM acts solely as a semantic normalizer, while a deterministic fuzzy engine handles the safety-critical reasoning. This avoids the "black box" problem of pure neural approaches and the brittleness of pure symbolic systems.
Expert-Guided Fuzzy Logic: The introduction of graded compliance scores via Sugeno fuzzy logic allows for nuanced assessment of protocol adherence, moving beyond pass/fail binary metrics.
Validation Framework: A rigorous validation process combining regex, LLM, embedding-based verification, and human expert adjudication to ensure high-confidence classification of clinical entities.

Quantitative Results

Classification Agreement: The Regex and MedGemma systems agreed on 94.26% of drug strings and 60.62% of microbiology combinations. Cohen's Kappa ( $\kappa = 0.65$ ) indicated substantial agreement, confirming the classifiers are complementary.
Compliance Scores:
- The overall mean compliance score across the cohort was 36.7% (median 37.5%).
- Antibiotic timing (Rule 2) was the most critical breakdown, with a mean score of 0.24 and only 13% of episodes meeting the strict one-hour threshold.
- Blood cultures (Rule 1) and Lactate measurement (Rule 3) also fell below 50% (0.36 and 0.39, respectively).
- Conditional rules for later phases (e.g., MAP recovery, Rule 7) showed higher scores (0.97), though the authors note this likely reflects survivorship bias (patients surviving long enough to reach this stage).
Clinical Cascade: A significant 51% drop-off was observed at the elevated lactate threshold ( $\ge 2.0$ mmol/L), indicating either unmeasured or undocumented lactate levels.
Outcomes:
- ICU Length of Stay (LOS): High compliance groups had a median ICU stay of 3.8 days, compared to 5.1 days for low compliance groups.
- Antibiotic Timing: Patients receiving antibiotics within 30–60 minutes had a median LOS of 2.95 days, rising to 4.74 days for delays beyond six hours.

Significance and Claims
The paper claims that this neuro-symbolic approach successfully bridges the gap between the interpretability required for safety-critical clinical decision-making and the flexibility needed to handle unstructured EHR data. By constraining the LLM to normalization and delegating reasoning to a fuzzy system, the pipeline produces interpretable, graded compliance insights that align with subject matter expert perspectives.

The authors emphasize that the system identifies systemic failures in the "Hour-1" bundle as the primary driver of prolonged critical care. They explicitly state that the pipeline is domain-general and can be adapted to other protocols (e.g., stroke, cardiac care) provided the guidelines are well-defined and domain experts are available to set boundaries.

Limitations and Future Work
The authors modestly acknowledge several limitations:

Mortality: The study measured protocol adherence, not survival impact. High scores in later stages reflect survivorship bias rather than outcome modeling.
Timing Windows: The current onset-anchored window penalizes antibiotics administered before ICU admission, which may not reflect the actual start of sepsis care.
Expert Dependency: The fuzzy boundaries currently rely entirely on expert input; future work aims to integrate data-driven calibration to optimize membership parameters while preserving rule priorities.
Scalability: The reliance on subject matter experts may limit scalability in resource-constrained settings, though maturing generative models may eventually reduce this dependency.

Technical Summary: An Expert-Guided Neuro-Symbolic Pipeline for Sepsis Compliance

More like this