This study developed and externally validated a model to re-evaluate in-hospital mortality risk at a 72-hour intensive care landmark for patients with sepsis and pre-existing hypertension. The authors note that admission data may not capture evolving need for circulatory support and organ dysfunction; thus they combined baseline (static) information with physiological and treatment features accumulated during the first 72 hours in intensive care to inform subsequent risk stratification.
The analysis used two publicly available critical care databases. Model development used MIMIC-IV and external validation used the eICU Collaborative Research Database (eICU-CRD). Inclusion criteria required patients to have received vasoactive agents and to remain in the ICU for at least 72 hours. The development cohort included 1,019 patients with 338 deaths (33.2%), and the external cohort included 1,747 patients with 516 deaths (29.5%). No new patient-level data were generated.
The modeling approach combined static variables present at or near admission with dynamic features derived from six consecutive 12-hour windows spanning the first 72 hours. The dynamic features captured information that accumulated during early intensive care, reflecting changes in physiology and treatment intensity. The manuscript reports an exploratory comparison between models using static-only features and models using both static and dynamic features.
Imputation, feature selection, model fitting, and calibration were performed within training folds during five-fold cross-validation to limit information leakage. The authors selected an XGBoost classifier as the final model. After internal training and cross-validation, the final XGBoost model was applied to the external eICU-CRD cohort without refitting.
Model evaluation included discrimination, calibration, decision-curve analysis, and risk stratification. Reported performance metrics were:
These results indicate moderate discrimination retained on the external dataset and measurable calibration deviation (slope < 1).
When patients in the external cohort were divided into risk tertiles based on model predictions, observed mortality rose from 13.7% in the lowest tertile to 47.3% in the highest tertile. In an exploratory comparison, the static-plus-dynamic feature set produced a higher internal AUC than the static-only set by 0.071 (95% CI 0.042–0.109), supporting the added value of information accumulated during the first 72 hours.
The authors interpret their findings to mean that information accumulated during the initial 72 hours of intensive care can support subsequent mortality risk stratification in this selected population of patients with sepsis and pre-existing hypertension who received vasoactive agents. The model retained moderate discrimination on external validation, suggesting potential utility as an adjunct to clinical reassessment rather than a standalone decision tool.
Limitations explicitly noted in the abstract and discussion include the need for local calibration and evaluation of the clinical consequences of acting on model predictions before routine use. The study used retrospective, de-identified databases; the preprint status indicates the work has not been peer reviewed and therefore should not yet guide clinical practice.
The analysis was a secondary study of de-identified databases. The creation and sharing of MIMIC-IV were approved by the Institutional Review Board of Beth Israel Deaconess Medical Center with a waiver of individual consent; eICU-CRD contains de-identified data available through PhysioNet under data use agreements. The authors state they will make analysis code and reproduction instructions publicly available upon publication. Aggregate results are reported in the manuscript and supporting information. The authors declared no competing interests.
Note: This report summarizes the preprint's methods and results. The source article is a preprint and has not been peer reviewed; details beyond those reported in the manuscript were not provided.