Predicting dynamic individual out-of-hospital cardiac arrest risks using explainable machine learning: a multicenter study in China.

Abstract

Out-of-hospital cardiac arrest (OHCA) poses a significant public health challenge, with limited tools available for dynamic risk estimation under changing environmental conditions. This study aimed to develop and validate an OHCA risk prediction model that integrates individual and environmental factors. Using a multicenter nested case-control design, we analyzed data from 26,145 OHCA cases and 162,160 controls in Chongqing, China, collected between January 2021 and July 2024. By linking clinical data with daily meteorological variables, we evaluated several modeling approaches, including logistic regression, Lasso, ridge, and Extreme gradient boosting (XGBoost). The XGBoost model incorporating both meteorological and individual characteristics outperformed models based solely on individual factors, achieving an the receiver operating characteristic curve (AUC) of 0.810 compared to 0.789 (P < 0.001). XGBoost achieved a 5.4 fold higher positive predictive value than logistic regression at 99% specificity (2.45% vs. 0.45%), with 0.96 sensitivity at the Youden-optimal threshold. These results indicate that the model maintains high sensitivity while significantly improving predictive value at high-specificity thresholds. Our findings underscore the importance of integrating real-time meteorological data into OHCA risk prediction, which may help inform situational awareness and contribute to emergency medical services (EMS) preparedness planning.