Machine learning algorithms for predicting days of high incidence with out-of-hospital cardiac arrest

Abstract

Abstract Background: Predicting out-of-hospital cardiac arrest (OHCA) events might contribute to the improvement of OHCA patients’ outcomes. We hypothesized that machine learning algorithms using meteorological and chronological information would predict high OHCA incidence.Methods: We used the large Japanese population-based repository database of OHCA and weather information. The data of Tokyo (2005-2012) were used as the training (derivation) cohort and the data of the top six most populated prefectures of Japan (2013-2015) as the testing (validation) cohorts. Eight machine learning, including eXtreme Gradient Boosting (XGBoost), were used. The primary outcome was high-incidence days, defined as the daily events exceeding 75% tile of our dataset in Tokyo between 2005-2015. In addition, we used the Shapley Additive exPlanations (SHAP) values to evaluate the contribution of each feature to the model. Secondly, we compared the daily OHCA incidence between the elderly and non-elderly patients to determine the impact of meteorological and chronological information. Results: The training cohort included 96,597 OHCA patients. In the primary analysis of the training cohort, eight machine learning models achieved an area under the receiver operating curve (AUROC) above 0.89. Among these, XGBoost had the highest AUROC of 0.906 (95% confidence interval [CI] 0.868–0.944). In the test cohorts, the XGBoost prediction algorithms had the similarily high AUROC values (Tokyo 0.923, Kanagawa 0.882, Osaka 0.888, Aichi 0.889, Saitama 0.879, Chiba 0.862). The SHapley Additive exPlanations values indicated that the “mean temperature on the previous day” had the highest impact on the model. In the secondary analysis, the lower mean temperature of the previous day was associated with the higher daily incidence in the elderly population. OHCA incidence was highest on Sundays and Mondays in the elderly group, whereas on Mondays in the non-elderly group.Conclusions: Algorithms using machine learning with meteorological and chronological information could accurately predict OHCA events.