Landmark-based early perfusion phenotyping for prognostic stratification in septic shock: derivation in MIMIC-IV and fixed-parameter external transportability assessment in eICU.

BACKGROUND

Septic shock is clinically heterogeneous, and early perfusion failure may not be adequately captured by any single bedside variable. We aimed to derive a clinically interpretable early perfusion phenotyping framework within a landmark design and assess whether the high-risk phenotype contour could be transferred across critical care databases.

METHODS

In this retrospective observational study, MIMIC-IV was used for phenotype derivation and the eICU Collaborative Research Database for fixed-parameter external transportability assessment. Adults in the first intensive care unit stay were eligible if they had early vasopressor exposure and baseline lactate greater than 2 mmol/L. At an 8-h landmark, phenotypes were derived using mean arterial pressure deficit burden, body-weight-standardized urine output, and lactate clearance. The primary outcome was in-hospital mortality, selected for cross-database harmonization because it was consistently available in both databases, whereas fixed-time 28- or 90-day mortality was not reliably harmonized. Secondary renal outcomes were summarized descriptively.

RESULTS

The MIMIC landmark cohort included 4,374 patients, of whom 2,671 had complete data for phenotype assignment. Four early perfusion phenotypes were identified. Phenotype 1 was characterized by greater hypotensive burden, poorer lactate clearance, and lower urine output. In-hospital mortality was 47.0% in Phenotype 1 versus 17.7% in Phenotype 4. Phenotype 1 remained associated with higher mortality after primary adjustment for age, sex, and pre-landmark SOFA score (OR 3.67, 95% CI 2.74-4.93), but this estimate attenuated after expanded pre-landmark case-mix adjustment (OR 2.56, 95% CI 1.85-3.53). In eICU, the same high-risk contour remained detectable under frozen transfer among measured complete-case patients. However, the strict and relaxed transfer cohorts were small, monitoring-dense subsets, and class-prevalence drift with measurement-dependent complete-case formation precluded robust conclusions about transportability of the entire four-class model, particularly for Phenotypes 2-4.

CONCLUSIONS

A landmark-based early perfusion phenotype identified a high-risk perfusion-failure contour among septic shock patients with sufficient early monitoring density. In eICU, this contour was recoverable only within measured complete-case subsets, and the external data do not establish robust transportability of the entire four-class model. The framework should be interpreted as an observational, hypothesis-generating prognostic stratification approach rather than as an independently validated bedside decision tool.