How Compliant is Sepsis Treatment? An Expert-Guided Neuro-symbolic Pipeline for Generating Clinical Compliance Insights
06:00 · August 17, 2026 · arXiv cs.AI RSS
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).
Summary
A neuro-symbolic pipeline addresses the challenge of verifying adherence to evidence-based sepsis protocols in large, unstructured electronic health records. Purely symbolic systems struggle with semantic variations in drug names and clinical notes, while standalone neural models produce opaque outputs unsuitable for safety-critical decisions. The presented approach confines a large language model, MedGemma, to semantic normalization of messy clinical strings, mapping them to a fixed vocabulary of antibiotics, vasopressors, and laboratory values. A separate Sugeno fuzzy inference system then applies eight Surviving Sepsis Campaign bundle rules to generate graded compliance scores in the interval [0,1], preserving interpretability and allowing partial credit for near-misses.
The fuzzy layer encodes rules across three clinical phases: immediate actions such as blood-culture collection, antibiotic administration, and lactate measurement within the first hour; hemodynamic resuscitation involving fluid and vasopressor use; and response assessment covering mean arterial pressure recovery and lactate clearance. Domain experts define membership functions and decision boundaries, while the language model operates under strict zero-shot prompting and is cross-validated against a regex baseline. On 1,691 unique drug strings the two classifiers agreed in 94.26 percent of cases, yielding a Cohen’s kappa of 0.65 and confirming complementary error patterns rather than redundancy.
Applied to 2,438 sepsis episodes sampled from MIMIC-IV v3.1, the pipeline reveals pronounced gaps in early intervention. Antibiotic timing produced the lowest mean compliance score of 0.24, with only 13 percent of episodes meeting the one-hour target. Hour-1 bundle performance averaged 36.7 percent, and lactate re-measurement showed a 51 percent drop-off after an initially elevated result. Episodes with higher overall compliance correlated with shorter ICU stays (3.8 versus 5.1 days), supplying descriptive evidence that the graded scores align with observable patient trajectories.
Why it matters
The paper's focus on transparent, neuro-symbolic AI directly aligns with the Dutch and EU emphasis on trustworthy and explainable AI in safety-critical domains like healthcare. Dutch AI researchers and medical centers can leverage this hybrid methodology to develop compliant clinical decision-support systems that adhere to strict EU regulations.
