Deyu Kong, Y Long, Tingting Huang, Yonghu Zhang, Liu W, Cheng-Xi Liu, Liu S
Aims: Dysglycemia is common in critically ill patients and is associated with adverse outcomes, but the prognostic significance of early ICU glycemic patterns remains unclear. Methods: In this dual-database cohort study, 60,365 critically ill patients from MIMIC-IV were used for model development and 81,237 from eICU for external validation. Early blood glucose trajectories were clustered using dynamic time warping (DTW). Associations with mortality were assessed using survival and multivariable regression analyses. Least Absolute Shrinkage and Selection Operator (LASSO)-selected predictors were incorporated into machine-learning models, which were externally validated and interpreted using SHapley Additive exPlanations (SHAP). Results: Four glycemic trajectory phenotypes were identified in both cohorts. A rising stress hyperglycemia-like trajectory was consistently associated with the highest mortality risk. In MIMIC-IV, this phenotype was associated with higher in-hospital, 180-day, and 1-year mortality; in eICU, it was also associated with the highest in-hospital mortality. Machine-learning models integrating trajectory phenotype with other clinical variables outperformed Simplified Acute Physiology Score II (SAPS II) and Oxford Acute Severity of Illness Score (OASIS). Random Forest was selected as the representative model, and a web-based tool was developed for individualized risk estimation. Conclusions: A rising stress hyperglycemia-like phenotype identifies a high-risk subgroup of critically ill patients and may improve mortality risk stratification beyond conventional severity scores.