Neuro-symbolic Pipeline Assesses Sepsis Treatment Compliance in MIMIC-IV

A novel neuro-symbolic framework aimed at evaluating adherence to sepsis treatment protocols has been introduced, as outlined in a paper on arXiv. This system integrates a large language model (LLM) focused on semantic normalization with a Sugeno fuzzy inference mechanism. The LLM translates disorganized drug and microbiology data into a standardized clinical vocabulary, while the fuzzy inference system incorporates eight rules from the Surviving Sepsis Campaign, using graded scores ranging from 0 to 1 instead of binary assessments. Analyzing 2,438 sepsis cases from MIMIC-IV v3.1, the framework revealed that antibiotic timing was the most significant issue, with a mean score of 0.24, as only 13% received antibiotics within one hour. Additionally, there was a mean score of 36.7% for hour-1 performance, and a 51% decrease in elevated lactate readings was noted. Compliance groups showed notable differences in ICU length of stay: 3.8 days for compliant patients compared to 5.1 days for those non-compliant. The research underscores the promise of neuro-symbolic techniques in high-stakes clinical environments, where neither purely neural nor purely symbolic methods suffice.