BEYOND SOFA AND APACHE: A SMARTER PATH TO ICU RISK STRATIFICATION

ICU mortality remains high due to complex interactions between organ dysfunction, metabolic derangements, and treatment interventions. Traditional severity scores (SOFA, APACHE II) lack dynamic integration of metabolic markers such as electrolyte disturbances, which may provide early prognostic signals.

We retrospectively analyzed 1,004 adult ICU patients admitted between January 2019–December 2022. Demographic, clinical, and laboratory data (including electrolytes, lactate, pH) were collected at admission. Independent predictors of ICU mortality were identified using multivariate logistic regression. Predictive performance was compared between logistic regression, SOFA, APACHE II, and machine learning models (Random Forest, XGBoost). Discrimination (AUC-ROC), calibration, and subgroup performance (sepsis status, age strata, AKI stage) were assessed. A nomogram was derived from multivariate predictors.

ICU mortality was 30.7%. Independent predictors included SOFA ≥8 (aOR=4.0), mechanical ventilation (aOR=3.1), vasopressor use (aOR=2.7), lactate >4 mmol/L (aOR=2.6), GCS ≤8 (aOR=2.4), AKI stage 2–3 (aOR=2.2), pH <7.2 (aOR=1.9), age ≥65 (aOR=1.8), and sepsis (aOR=2.5) (Figure 1). XGBoost achieved the highest discrimination (AUC=0.86) vs. logistic regression (AUC=0.84), SOFA (AUC=0.72), and APACHE II (AUC=0.75), with consistent performance across subgroups (AUC >0.80). A validated 8-variable nomogram stratified patients into four mortality risk tiers, with scores >150 predicting >60% mortality. In sepsis, lactate and vasopressor use contributed 38% to model predictions.

Integrating metabolic markers with established severity scores significantly improves ICU mortality prediction. The developed nomogram is practical, interpretable, and applicable at the bedside, offering a tool for early identification of high-risk patients and guiding targeted interventions.