The high-risk phenotype for gastrointestinal vulnerability in sepsis and 28-day mortality: an integrative study based...

Gastrointestinal dysfunction is common in sepsis but remains insufficiently represented in organ-specific risk stratification, partly because bedside gastrointestinal findings are often incompletely captured in structured electronic health record data. We defined gastrointestinal vulnerability phenotype (GIVP) high-risk as a prespecified rule-based electronic health record-operationalized phenotype of early gastrointestinal vulnerability and evaluated its prognostic relevance, incremental risk-stratification value, external prognostic directionality, and cross-level biological plausibility.

This retrospective integrative study used intensive care unit (ICU) admission events with sepsis from the medical information mart for intensive care IV (MIMIC-IV) to construct GIVP high-risk from three prespecified binary domains reflecting early hemodynamic support or hypoperfusion, absence of early enteral nutrition initiation, and early iatrogenic exposure burden. Multivariable logistic regression was used to evaluate the association between GIVP high-risk and fixed-window 28-day mortality. Supportive exposure analyses, sequential organ failure assessment (SOFA)-based stratified analyses, unit-of-analysis sensitivity analyses, Cox sensitivity analysis, incremental prediction analysis, decision curve analysis, and reduced external validation in the eICU collaborative research database (eICU-CRD) were performed. Cross-level biological plausibility was explored using publicly available peripheral blood single-cell transcriptomic data and animal intestinal tissue transcriptomic data.

The primary analysis cohort included 28,224 ICU admission events, of which 6,862 resulted in 28-day mortality. In the core model adjusted for age, sex, Charlson Comorbidity Index, infection source, the continuous SOFA score, and renal replacement therapy, GIVP high-risk retained an adjusted association with fixed-window 28-day mortality (OR = 1.216, 95% CI 1.143-1.295; P < 0.001). This association remained directionally consistent after additional adjustment for mechanical ventilation, in unit-of-analysis sensitivity analyses, and in Cox sensitivity analysis, although proportional hazards diagnostics suggested possible non-proportionality. Incremental prediction analysis showed only a very small discrimination improvement in the full cohort, which did not reach DeLong statistical significance, whereas the incremental signal was more concentrated in the high-SOFA subgroup. Decision curve analysis showed small net-benefit gains after adding GIVP high-risk, with maximum ΔNB values of 0.00291266 in the full cohort and 0.00474649 in the high-SOFA subgroup. In eICU-CRD, the reduced GIVP proxy was associated with hospital mortality after adjustment for age, sex, and Acute Physiology and Chronic Health Evaluation (APACHE) score (OR = 1.217, 95% CI 1.183-1.252; P < 0.001), supporting external prognostic directionality rather than exact replication of the full MIMIC-IV phenotype. Cross-level analyses showed directional concordance with an interferon (IFN)-high host-response pattern and intestinal tissue transcriptomic patterns characterized by enhanced inflammation, defense imbalance, and abnormal tissue remodeling.

GIVP high-risk is a rule-based electronic health record-operationalized phenotype that retained a prognostic association with 28-day mortality in sepsis after adjustment for clinical context and severity-related variables. Its incremental value was small and context-dependent, with the signal being more evident in higher-severity settings. Reduced external validation supported external prognostic directionality, whereas cross-level transcriptomic analyses supported biological plausibility; neither implied exact external replication nor causal inference.