Prediction of in-hospital mortality at a 72-hour intensive care landmark in patients with sepsis and pre-existing hypertension: development and external validation

Abstract Background: In patients with sepsis and pre-existing hypertension, admission findings may not adequately describe the subsequent course of circulatory support and organ dysfunction. We developed and externally validated a model to reassess in-hospital mortality risk after the first 72 hours of intensive care. Methods: This retrospective study used MIMIC-IV for development and eICU-CRD for external validation. Patients had received vasoactive agents and remained in the ICU for at least 72 hours. Static variables were combined with features from six consecutive 12-hour windows. Imputation, feature selection, model fitting, and calibration were performed within training folds during five-fold cross-validation. The final XGBoost model was applied externally without refitting. Evaluation included discrimination, calibration, decision curves, and risk stratification, with an exploratory comparison of static-only and static-plus-dynamic feature sets. Results: The development and external cohorts included 1019 and 1747 patients, with 338 (33.2%) and 516 (29.5%) deaths, respectively. Internal area under the receiver operating characteristic curve (AUC) was 0.752 (95% CI 0.720-0.784). External AUC was 0.713 (95% CI 0.687-0.739), average precision was 0.536, the Brier score was 0.185, and the calibration slope was 0.744. Mortality increased across external risk tertiles from 13.7% to 47.3%. In an exploratory feature-set comparison, the static-plus-dynamic feature set showed a 0.071 higher internal AUC than the static feature set (95% CI 0.042-0.109). Conclusions: Information accumulated during the first 72 hours supported subsequent mortality risk stratification in this selected population. The model retained moderate discrimination externally, supporting further evaluation as an adjunct to clinical reassessment. Local calibration and the clinical consequences of acting on model predictions require evaluation before routine use.