Abstract
Fluctuations in blood glucose during acute neurocritical illness are associated with poor outcomes, but the role of stress hyperglycemia ratio (SHR) in predicting outcomes across diverse neurocritical conditions remains unclear. This study evaluated SHR as a prognostic indicator for short and long-term mortality and its utility in machine learning-enhanced predictive models. Using data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) version 3.0 database, 2376 patients with acute neurocritical illness were grouped by admission SHR quartiles (Q1-Q4). Kaplan-Meier survival analysis, Cox proportional hazards models, and restricted cubic splines assessed the associations between the SHR index and the mortality at 30, 90, 180, and 360 days. Machine learning algorithms optimized feature selection for 30-day mortality prediction, with SHAP (SHapley Additive exPlanations) values to visualize model contributions. A total of 2376 participants were included in the study. Higher SHR quartiles were associated with increased all-cause mortality across different time points compared to Q2 as shown by Cox proportional hazards models. Risk escalated when SHR ≥ 0.86, with stronger associations in non-diabetic patients. Higher SHR is strongly associated with increased all-cause mortality across different time horizons and may serve as a valuable prognostic tool for neurocritically ill patients, especially in non-diabetics. Combining SHR with Glasgow Coma Score (GCS) improved outcome prediction compared to GCS alone.