Poster 027: Glycemic Variability in first 24 hours Predicts 6-Month Mortality in Critically Ill Patients: A Multisite Study Using Mixed-Effects Machine Learning

Background: Glycemic variability is a hallmark of the acute physiological response to shock and critical illness. The ability to leverage the first 24-hour glycemic dynamics to predict long-term mortality across heterogeneous ICU populations remains challenging. We aimed to develop and validate machine-learning models for 6-month survival using standardized EHR patient data.

Methods: We conducted a retrospective, multisite observational study of 6,505 unique ICU patients across six sites using an EHR data warehouse mapped to the OMOP Common Data Model (CDM). We extracted all glucose measurements, vitals, labs, and comorbidities from the first 24 hours of ICU admission. Predictive features included minimum and maximum glucose readings, coefficient of variation (CV), Time in Range (70–180 mg/dL), and advanced risk indices (HBGI/LBGI). We utilized Mixed-Effects Random Forest (MERF) and BME-XGBoost to account for patient and site-level clustering. Model performance was evaluated using Concordance Index, Time-dependent AUC, integrated Brier Score and Calibration (slope/intercept).

Outcome: The models achieved strong discrimination showing concordance of 0.76 and demonstrated resilience in LOSOCV, suggesting capture of portable risk signals rather than site-specific practices. SHAP (SHapley Additive exPlanations) analysis revealed that glucose trajectories on days 2–4 were the most influential predictors of long-term death. In non-diabetic patients, hypoglycemia and high variability were strongly associated with metabolic acidosis (low bicarbonate) and acute kidney injury. Notably, insulin administration significantly attenuated the negative impact of hyperglycemia. SHAP values showed reduced mortality risk when insulin was given early. This suggests a systemic protective effect on renal and metabolic homeostasis independent of simple glucose lowering.

Conclusion: Early glycemic dynamics carry important data that prognosticates 6-month-survival for critically-ill patients. Mixed-effects machine-learning frameworks provide accurate predictions within 24 hours of admission. These findings support the transition from standardized glycemic targets to model-guided decisions that identifies high-risk patients early, enabling interventions to mitigate the catabolic, inflammatory consequences of shock.