Persistent stress hyperglycemia trajectories predict mortality in sepsis: a machine learning-enhanced cohort study.

backgroundStress-induced hyperglycemia is a common metabolic derangement in sepsis linked to adverse outcomes. However, the prognostic significance of the dynamic trajectories of the stress hyperglycemia ratio (SHR) remains poorly understood.

methodsWe conducted a retrospective cohort study using data from the MIMIC-IV database. The SHR was calculated at six time points within the first 24 h of ICU admission. Group-based trajectory modeling (GBTM) was used to identify distinct SHR trajectory phenotypes. The associations between these trajectories and in-hospital, 28-day, and 90-day mortality were assessed using multivariable logistic and Cox regression models. The prognostic value of SHR trajectories was further validated using multiple machine learning models.

resultsAmong 1,834 patients with sepsis, four distinct SHR trajectories were identified: ‘Low-Stable’, ‘Moderate-Stable’, ‘High-Improving’, and ‘High-Persistent’. The ‘High-Persistent’ trajectory group displayed the highest inflammatory burden, greatest organ dysfunction, and worst mortality rates. After multivariable adjustment, the ‘High-Persistent’ trajectory was independently associated with increased 28-day mortality (HR 2.02, 95% CI 1.42–2.86), in-hospital mortality (OR 1.16, 95% CI 1.07–1.25), and 90-day mortality (HR 1.90, 95% CI 1.37–2.63; all P < 0.001). Feature selection algorithms consistently identified the SHR trajectory as a key predictor of mortality. The Gradient Boosting Machine (GBM) model, which incorporated the SHR trajectory, achieved excellent discrimination for 28-day mortality with an AUC of 0.863.

conclusionsDynamic SHR trajectories identify distinct prognostic phenotypes in patients with sepsis. A persistently elevated SHR is a powerful, independent predictor of mortality, signifying a state of sustained and maladaptive metabolic stress. Integrating SHR trajectory analysis, particularly within machine learning frameworks, holds significant promise for advancing early risk stratification and personalizing glycemic management in critical care. CLINICAL TRIAL NUMBER: Not applicable.