Predicting mortality in critically ill patients: a machine learning approach to electrolyte imbalances and clinical risk factors

The study aimed to enhance ICU mortality prediction by integrating electrolyte imbalances and clinical risk factors using machine learning, addressing gaps in traditional severity scores like SOFA and APACHE II. Dominant Predictors of Mortality: SOFA score ≥ 8 (aOR = 4.0), mechanical ventilation (aOR = 3.1), hyperlactatemia (> 4 mmol/L; aOR = 2.6), vasopressor use (aOR = 2.7), and GCS ≤ 8 (aOR = 2.4). High-Risk Phenotypes: “Triad of Death” (SOFA ≥ 8 + mechanical ventilation + vasopressors), metabolic crisis (lactate > 4 mmol/L + pH < 7.2), and frail elderly (age ≥ 65 with comorbidities). The XGBoost model achieved superior discrimination (AUC-ROC = 0.86) compared to logistic regression (AUC = 0.84), SOFA (AUC = 0.72), and APACHE II (AUC = 0.75). A nomogram was developed to stratify mortality risk into four actionable tiers (e.g., scores > 150 predicted > 60% mortality). The model performed consistently across subgroups (sepsis, age strata, AKI stages), with AUCs > 0.80. Sepsis patients: Lactate and vasopressor use were stronger predictors (38% contribution to SHAP values) Early correction of modifiable factors (e.g., sepsis, acidosis) may improve outcomes. The nomogram offers a practical bedside tool for risk stratification and clinical decision-making. Critically ill patients in intensive care units (ICUs) exhibit high mortality due to complex interactions among metabolic instability, organ dysfunction, and therapeutic interventions. While conventional severity scores like SOFA and APACHE II are widely used, they insufficiently integrate dynamic metabolic markers. We performed a retrospective cohort study of 1,004 adult ICU patients. Demographic, clinical, and laboratory data were analyzed using multivariate logistic regression and machine learning algorithms (Random Forest, XGBoost). Model performance was assessed using AUC-ROC, calibration slope, and sensitivity at fixed specificity threshold. The ICU mortality was 30.7%. Independent predictors of mortality included SOFA score ≥ 8 (aOR = 4.0), mechanical ventilation (aOR = 3.1), hyperlactatemia ≥ 4 mmol/L (aOR = 2.6), vasopressor use (aOR = 2.7), and GCS ≤ 8 (aOR = 2.4). The XGBoost model achieved superior discrimination (AUC = 0.86; 95% CI: 0.83–0.89) compared to logistic regression (AUC = 0.84), SOFA (AUC = 0.72), and APACHE II (AUC = 0.75). A validated nomogram incorporating eight key predictors stratified mortality risk into four actionable tiers. Subgroup analysis confirmed robust performance across age strata, sepsis status, and AKI stages. Integrating metabolic markers with clinical severity scores enhances ICU mortality prediction. Our nomogram offers a practical, interpretable tool for bedside risk stratification and clinical decision support.