Multi-Center Machine Learning-Based Prediction of Mortality in ICU Patients with Hypocalcemia.

Background: Hypocalcemia occurs frequently in intensive-care units (ICUs) and is independently associated with excess mortality. Conventional severity scores-such as APACHE II and SOFA-assign fixed weights to a limited set of variables and therefore fail to capture the nonlinear, high-dimensional physiology characteristic of hypocalcemic patients. Although machine-learning (ML) approaches can enhance risk stratification, no interpretable model tailored to this cohort has been available.

Methods: We harmonised de-identified data from MIMIC-III, MIMIC-IV and two Grade III Level A hospitals in China, generating a multicentre cohort of 13,979 adult ICU admissions with total serum calcium < 2.12 mmol L-1. MIMIC-IV records were randomly divided into a training set (n = 7,749) and an internal-validation set (n = 1,550). External validation employed MIMIC-III (n = 4,771) and the Chinese multicentre dataset (n = 209). Predictors were filtered with least-absolute-shrinkage-and-selection operator (LASSO) regression and applied to eight ML algorithms: logistic regression, k-nearest neighbors (KNN), support-vector machine, decision tree, random forest, artificial neural network, eXtreme Gradient Boosting (XGBoost) and LightGBM. Model discrimination, calibration and clinical utility were quantified using the area under the receiver-operating-characteristic curve (AUC), F1-score, sensitivity, specificity, calibration plots, decision-curve analysis (DCA) and clinical-impact curves (CIC). SHapley Additive exPlanations (SHAP) were used for interpretability, and the final model was deployed as a public web application.

Results: LASSO retained 20 predictive variables; is_noninvasive_ventilator and hospital length of stay were the most influential in SHAP analysis. XGBoost provided the highest discrimination (AUC = 0.914; F1 = 0.844), surpassing logistic regression (AUC = 0.896; F1 = 0.829), LightGBM (AUC = 0.909; F1 = 0.816) and conventional ICU scores. Calibration curves, DCA and CIC confirmed consistent performance and superior net benefit across internal and external validation cohorts.

Conclusions: We present and externally validate an interpretable, high-performance ML model that predicts in-hospital mortality in hypocalcemic ICU patients more accurately than established scoring systems. The SHAP-enabled web interface provides real-time, patient-specific risk estimates, facilitating data-driven clinical decisions within the early, critical window of ICU care.