Development and validation of machine learning-based prediction model for outcome of cardiac arrest in intensive care units

Cardiac arrest (CA) poses a significant global health challenge and often results in poor prognosis. We developed an interpretable and applicable machine learning (ML) model for predicting in-hospital mortality of CA patients who survived more than 72 h. A total of 721 patients were extracted from the Medical Information Mart for Intensive Care IV database, divided into the training set (n = 576) and the internal validation set (n = 145). The external validation set containing 856 cases were collected from four tertiary hospitals in Zhejiang Province. The primary outcome was in-hospital mortality. Eleven ML algorithms were utilized to establish prediction models based on data from 72 h after return of spontaneous circulation (ROSC). The results indicate that the CatBoost model exhibited the best performance at 72 h. Eleven variables were ultimately selected as key features by recursive feature elimination (RFE) to construct a compact model. The final model achieved the highest AUC of 0.86 (0.80, 0.92) in the internal validation and 0.76 (0.73, 0.79) in the external validation. SHAP summary plots and force plots visually explained the predicted outcomes. In conclusion, 72-h CatBoost showed promising performance in predicting in-hospital mortality of CA patients who survived more than 72 h. The model still requires further optimization and improvement.