Multicenter Machine Learning-Based Prediction of Mortality in ICU Patients With Hypocalcemia.

Area of Science:

• Critical Care Medicine
• Data Science
• Biostatistics

Background:

• Hypocalcemia is common in ICUs and linked to higher mortality.
• Existing scoring systems inadequately capture complex patient physiology.
• There's a need for interpretable risk stratification models for hypocalcemic ICU patients.

Purpose of the Study:

• To develop and validate an interpretable machine-learning model for predicting in-hospital mortality in hypocalcemic ICU patients.
• To improve upon the accuracy of conventional severity scores.

Main Methods:

• A multicenter cohort of 13,979 adult ICU admissions with hypocalcemia was compiled from MIMIC-III, MIMIC-IV, and Chinese hospitals.
• LASSO regression filtered predictors for eight machine-learning algorithms.
• Model performance was assessed using AUC, F1-score, calibration plots, DCA, and CIC; SHAP was used for interpretability.

Main Results:

• The eXtreme Gradient Boosting (XGBoost) model achieved the highest discrimination (AUC = 0.914).
• The model demonstrated superior performance compared to logistic regression and conventional ICU scores.
• SHAP analysis identified key predictors, including noninvasive ventilation and hospital length of stay.

Conclusions:

• An interpretable, high-performance machine-learning model accurately predicts mortality in hypocalcemic ICU patients.
• The model surpasses established scoring systems in predictive accuracy.
• A SHAP-enabled web application provides real-time, patient-specific risk estimates to guide clinical decisions.