Early Risk Stratification for 30-Day Mortality After In-Hospital Cardiac Arrest: SHAP Interpretable CatBoost Model with m-NUTRIC and Micronutrient Biomarkers

TL;DR

This study developed a machine learning model to predict 30-day mortality after in-hospital cardiac arrest using clinical scores and micronutrient biomarkers.

Contribution

The novel contribution is an interpretable CatBoost model using SHAP values and micronutrient biomarkers for early risk stratification.

Findings

CatBoost achieved a ROC-AUC of 0.850 in training and 0.827 in testing, outperforming logistic regression.

SHAP analysis showed m-NUTRIC and age as dominant predictors, with micronutrients as complementary factors.

The model demonstrated net benefit across threshold probabilities of 0.20–0.70 in decision curve analysis.

Abstract

Background/Objectives: Predicting 30-day mortality after in-hospital cardiac arrest (IHCA) remains challenging. We developed an interpretable CatBoost model that incorporates the m-NUTRIC score, age, and selected micronutrient biomarkers (i.e., magnesium, zinc, vitamin D, and vitamin B12). We compared its performance with that of logistic regression and quantified variable contributions using SHAP. Methods: Variables were extracted from the electronic medical records of 880 patients with IHCA admitted to a medical intensive care unit. The CatBoost and logistic regression models were trained on a stratified 80/20 split. The decision threshold was optimized using the Youden index (0.482). Discrimination (ROC-AUC with bootstrap confidence intervals), classification metrics, precision–recall analysis, calibration, and decision curve analysis were reported. Results: CatBoost achieved a ROC-AUC of 0.850 (95% confidence interval [CI]: 0.822–0.879) in the training set and 0.827 (95% CI: 0.760–0.887) in the internal test set, outperforming logistic regression (0.797; 95% CI: 0.720–0.861). The test set accuracy, precision, recall, F1-score, specificity, and average precision were 0.761, 0.847, 0.790, 0.817, 0.702, and 0.909, respectively. The Brier score was 0.186. Decision curve analysis showed net benefit across threshold probabilities of 0.20–0.70. The SHAP analysis identified m-NUTRIC and age as the dominant predictors, whereas micronutrients served as complementary contextual factors. Conclusions: The CatBoost model consistently outperformed the logistic regression and warrants prospective multicenter validation.

Figures