Early mortality prediction of prognosis in cardiac arrest patients using machine learning: Development, external validation, and explainability with SHAP.

BACKGROUND: Cardiac arrest (CA) is a major global health challenge, accounting for a significant proportion of deaths and healthcare resource utilization. However, due to the lack of interpretability, most predictive models for CA mortality have not been applied in clinical practice. We aimed to construct an interpretable model for predicting in-hospital mortality for CA patients in the intensive care units (ICU). METHODS: Data from the MIMIC-IV database were split, with 70% allocated to the training set and 30% to the internal validation set. The eICU database served as an external validation set. Using the variables selected by least absolute shrinkage and selection operator (LASSO), we constructed and evaluated seven machine-learning (ML) models. The optimal model was rendered explainable through the Shapley additive explanations (SHAP) approach. And, a web-based calculator was developed based on the optimal performance model for easy use. RESULTS: A total of 1749 patients with CA were finally enrolled from MIMIC-IV in this study. By LASSO regression, 21 variables were selected to construct the ML models. Among seven constructed models, the eXtreme Gradient Boosting (XGBoost) model emerged as the best-performing model in internal validation set (AUC = 0.8486) and external validation set (AUC=0.8471). In addition，XGBoost model showed higher net benefit and wider threshold probability in DCA. Furthermore, SHAP force analysis visualized the individualized mortality prediction performed by the model. CONCLUSIONS: We developed an XGBoost-based prediction model with good performance for early mortality prediction in CA patients. Model interpretability was enhanced using SHAP analysis. An online calculator based on the model is also provided.