Joint lactate and vasoactive-inotropic score trajectories and post-landmark mortality in sepsis: a retrospective cohort study in MIMIC-IV

Circulatory failure in sepsis is reflected jointly by blood lactate, a marker of tissue perfusion, and the vasoactive-inotropic score (VIS), an index of pharmacologic support, yet the two are usually modeled separately. Whether their joint 48-h courses define internally stable, clinically interpretable phenotypes with distinct post-landmark prognoses is uncertain. In this single-center retrospective cohort study of MIMIC-IV version 3.1, we included adults meeting Sepsis-3 criteria who were alive and still in the intensive care unit (ICU) at a fixed 48-h landmark. Four 12-h summaries of lactate and VIS were modeled as separate longitudinal processes connected through shared latent classes in a finite-mixture trajectory model. Class associations with post-landmark 28-day (primary) and 90-day mortality were estimated with proportional-hazards models, with posterior pseudo-class sampling as a misclassification sensitivity analysis. Incremental predictive value over a baseline clinical model (A) was quantified for lactate-only (B1), VIS-only (B2), joint-class (C), and dynamic-summary (D) models, with temporal validation in 2017–2019 admissions and a 2020–2022 drift stress test. Among 6,820 patients, three classes emerged: low lactate with no or minimal vasoactive support (C1, 51.5%); resolving lactate with declining low-dose support (C2, 30.0%); and slowly resolving hyperlactatemia with sustained high-dose support (C3, 18.5%). With full adjustment, 28-day mortality hazard ratios versus C1 were 1.39 (1.20–1.62) for C3 and 0.79 (0.69–0.91) for C2. However, when total SOFA was replaced by non-cardiovascular SOFA, the C2 estimate attenuated to 0.99 (0.86–1.13), while C3 remained at higher risk (1.83, 1.60–2.09). Standardized 28-day mortality was 18.9%, 15.5%, and 24.7% in C1, C2, and C3. In temporal validation, model C versus A improved the index of prediction accuracy from 0.125 to 0.143 and area under the curve from 0.762 to 0.772; model D performed better (AUC 0.782; IPA 0.151). All models degraded under the 2020–2022 stress test. Joint lactate-VIS trajectory phenotypes remained identifiable with high posterior certainty across within-center time periods, but independent external reproducibility was not assessed. The simpler dynamic-summary model slightly outperformed the latent-class model for individual prediction; the latter's principal contribution is therefore phenotyping rather than superior bedside risk prediction. The C2 association was adjustment-sensitive, and no class should be interpreted causally.