EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Explainable AI Framework for Evaluating Sepsis Management in Critical Care: A MIMIC-IV Implementation Study

Kevin Ding

Undergraduate Research in Natural and Clinical Science and Technology Journal · 17.08.2026 · 1-10

DOI: 10.26685/urncst.1004 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Undergraduate Research in Natural and Clinical Science and Technology Journal, online erschienen am 17.08.2026. Verfasst von Kevin Ding. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Sepsis requires fast, organized care, but it remains difficult to measure in real hospitals whether recommended steps occur on time. Reported compliance with the Surviving Sepsis Campaign (SSC) 1-hour bundle varies widely because guideline elements are hard to translate into electronic health records (EHRs), time zero is inconsistently defined, and circular logic can occur. Although multiple studies have evaluated SSC 1-hour bundle adherence using EHR data, existing implementations rely on heterogeneous or opaque logic, limiting their reproducibility and interpretability. There remains a need for a transparent, rule-based computational framework that expresses clinical guideline elements as formal temporal logic over event data. The purpose of this study is to develop and apply such a framework—an explainable AI system that operationalizes the SSC 1-hour bundle as computable temporal rules over the Medical Information Mart for Intensive Care IV (MIMIC-IV) dataset—and to evaluate how definitional choices affect measured adherence. The framework encodes each bundle element as rules with explicit time windows, introduces a second time-zero anchor that excludes antibiotics to avoid circularity, and applies a strict lactate rule requiring repeat testing when lactate itself defines time zero. We used an International Classification of Diseases (ICD)-based sepsis cohort (n = 14,399) and found modest step-level compliance: lactate 35.3%, culture-before-antibiotic 20.8%, antibiotics 74.3%, fluids when indicated 31.8%, and vasopressors when indicated 57.3%. Overall “all-steps” full compliance was 2.0%, falling slightly with the strict lactate rule. This study demonstrates that a formal explainable AI framework can reproducibly represent and audit clinical guideline logic, revealing that measured performance depends strongly on definitional choices. The framework provides a transparent, computationally rigorous foundation for future decision support and quality improvement in sepsis care.

Zitieren

APA

Ding, K. (2026). Explainable AI Framework for Evaluating Sepsis Management in Critical Care: A MIMIC-IV Implementation Study. Undergraduate Research in Natural and Clinical Science and Technology Journal, 1-10. https://doi.org/10.26685/urncst.1004

Vancouver

Ding K. Explainable AI Framework for Evaluating Sepsis Management in Critical Care: A MIMIC-IV Implementation Study. Undergraduate Research in Natural and Clinical Science and Technology Journal. 2026:1-10. doi:10.26685/urncst.1004

BibTeX
@article{Ding2026,
  author = {Ding, Kevin},
  title = {Explainable AI Framework for Evaluating Sepsis Management in Critical Care: A MIMIC-IV Implementation Study},
  journal = {Undergraduate Research in Natural and Clinical Science and Technology Journal},
  year = {2026},
  pages = {1--10},
  doi = {10.26685/urncst.1004}
}
Kostenlos starten →