Dynamic variability of serum sodium, potassium, and calcium and mortality after acute myocardial infarction: insights from traditional and machine learning approaches

Electrolyte variability may capture dynamic homeostatic instability beyond single admission values, yet the comparative prognostic importance of sodium, potassium, and calcium variability in critically ill patients with acute myocardial infarction (AMI) remains unclear. Their relative predictive importance has not been systematically evaluated using machine learning (ML) approaches. We conducted a retrospective cohort study using MIMIC-IV (v3.1). Patients were divided into quartiles based on electrolyte coefficient of variation (CV). Associations between CV quartiles and ICU mortality (28- and 90-day) were analyzed using Cox proportional hazards models. Kaplan–Meier curves with log-rank tests compared survival across quartiles, and restricted cubic splines (RCS) examined non-linear relationships. Subgroup and interaction analyses were prespecified for age (< 65 vs. ≥65), sex, and BMI (< 30 vs. ≥30 kg/m²). Sensitivity analyses excluded early deaths (≤ 7 days), patients with CKD, and recalculated CV using the first 48 h of ICU data. Additionally, machine learning models were developed to predict 28-day mortality. The best-performing model was interpreted using Shapley Additive Explanations to identify important features. In total, 3,632 ICU patients were included, median age was 70.0 years (IQR 61.0–78.0), and 65.0% were male. In Model II, higher electrolyte variability was independently associated with 28-day ICU mortality for sodium (HR = 1.11), potassium (HR = 1.02), and calcium (HR = 1.06). Using quartiles (Q1 as reference), the highest variability group (Q4) had increased risk for sodium (HR = 1.38), potassium (HR = 1.26), and calcium (HR = 2.34). For 90-day ICU mortality, Model II showed consistent associations for continuous CV: sodium (HR = 1.09), potassium (HR = 1.02), and calcium (HR = 1.06). Compared with Q1, Q4 remained significant for sodium (HR = 1.25), potassium (HR = 1.32), and calcium (HR = 1.80). RCS indicated non-linear positive relationships (P for non-linearity < 0.001), and calcium CV had the best discrimination. Sensitivity analyses confirmed robustness. ML analyses showed that calcium CV ranked second in importance, comparable to age, while sodium CV and potassium CV were also among the top six predictors of 28-day mortality. Greater sodium CV, potassium CV, and particularly calcium CV were independently associated with increased short- and medium-term ICU mortality in patients with AMI. Complementary machine learning analyses further underscore their prognostic importance for short-term ICU mortality risk stratification, supporting a clinical focus on maintaining electrolyte stability rather than solely correcting absolute levels.