Early Glucose States and 72-Hour Care-Metabolic Clusters in Critically Ill Patients

This is a retrospective observational cohort study using electronic health record data from intensive care units at Zhongshan Hospital, Fudan University. Source clinical data were recorded from 2021 through 2025. The retrospective research study was initiated after institutional ethics approval in July 2026.

Two nested observation windows are evaluated. Early glucose states are derived from routinely recorded glucose data during the first 24 hours after ICU admission. The early-state features include mean glucose, coefficient of variation, and mean absolute change between adjacent glucose measurements. ICU stays used for early-state derivation require at least five valid glucose measurements during hours 0-24.

A second classification summarizes integrated care-metabolic information during the first 72 hours after ICU admission. The 72-hour cluster feature set contains 26 variables comprising glucose characteristics, monitoring intensity, documented insulin exposure, nutrition documentation and volume, corticosteroid exposure, mechanical ventilation, vasopressor use, and renal replacement therapy. Demographic variables, severity scores, microbiology variables, and outcomes are not used to derive the clusters.

The 0-24-hour glucose states and 0-72-hour integrated clusters are compared within ICU stays carrying both labels. Because the observation windows overlap, this comparison is interpreted as a nested-window cross-sectional mapping rather than a temporal transition between states.

Outcome follow-up for the cluster analysis begins after completion of the 72-hour landmark. The primary clinical outcome is subsequent ICU mortality. Associations between cluster membership and subsequent ICU mortality are estimated using logistic regression with adjustment for age, sex, and first-day SOFA score. Missing covariate data are examined using multiple-imputation sensitivity analyses, while missing ICU outcomes are examined using prespecified sensitivity bounds rather than outcome imputation.

Robustness analyses assess patient-disjoint reproducibility, bootstrap stability, dependence on explicit documentation-count variables, glucose-measurement interval, and chronic glycemic context. Cross-validated analyses assess whether the derived labels add predictive information beyond their continuous source variables.

The fixed institutional early-glucose-state transformation and centroids are projected without refitting to eligible MIMIC-IV ICU stays. This external analysis evaluates the direction of the early-state mortality gradient only and is not considered external validation of the four 72-hour clusters.

All glucose management, insulin use, nutritional support, organ support, microbiological testing, and antimicrobial treatment reflect routine clinical care. No exposure, glucose target, insulin regimen, nutritional intervention, or organ-support strategy is assigned by the investigators.