SHAP-audited machine-learning model for short-term mortality risk stratification in ICU patients with documented hypokalemia: development and temporal external validation

Hypokalemia is a frequent electrolyte abnormality in the ICU and is associated with short-term mortality. However, risk prediction in this population remains challenging, and the prognostic information provided by routinely collected physiologic and laboratory variables has not been fully characterized in externally evaluated machine-learning models. Adult ICU admissions with documented hypokalemia (serum K⁺ < 3.5 mmol/L) and ICU stay > 24 h were analyzed as a retrospective admission-window risk-stratification study among patients retrospectively identified as hypokalemic, rather than as an admission-time deployment model for all ICU admissions. MIMIC-III was used for model development, MIMIC-IV for temporal validation, and two Chinese ICU cohorts for exploratory geographic evaluation. Because baseline-characterization, locked endpoint-specific evaluation, subgroup, and sensitivity analyses use different denominators, all analysis subsets and event counts are mapped explicitly in Supplementary Table S7. Twenty routinely available variables were selected using leakage-conscious preprocessing and LASSO. Model performance was evaluated with discrimination, calibration, decision-curve analysis, subgroup analyses, and sensitivity analyses; SHAP was used as a post-hoc model-audit method rather than as evidence of causality, actionability, or bedside interpretability. In the locked endpoint-specific evaluation, the 20-variable stacking model achieved an AUROC of 0.783 (95% CI, 0.708–0.857; n = 163; events = 50) in the MIMIC-III internal test subset and 0.838 (95% CI, 0.808–0.868; n = 1,413; events = 130) in MIMIC-IV temporal validation. The simplified 8-variable model achieved corresponding AUROCs of 0.717 (95% CI, 0.627–0.807) and 0.835 (95% CI, 0.805–0.866). Exploratory geographic evaluation showed heterogeneous performance in Third Xiangya (AUROC 0.830; 95% CI, 0.750–0.910) and Nanhua (AUROC 0.717; 95% CI, 0.616–0.819), supporting cautious interpretation and likely center-specific heterogeneity. We developed a SHAP-audited model for retrospective short-term mortality risk stratification in ICU patients with documented hypokalemia. The primary analysis does not eliminate temporal-selection bias; therefore, the findings support methodological and prospective evaluation but do not establish clinical deployment readiness, superiority over established severity scores, or first-hypokalemia-triggered use.