Predicting Survived Events in Nontraumatic Out-of-Hospital Cardiac Arrest: A Comparison Study on Machine Learning and Regression Models.

Abstract

Prediction of early outcomes of nontraumatic out-of-hospital cardiac arrest (OHCA) by emergency physicians is inaccurate. Our aim was to develop and validate practical machine learning (ML)-based models to predict early outcomes of nontraumatic OHCA for use in the emergency department (ED). We compared their discrimination and calibration performances with the traditional logistic regression (LR) approach. Between October 1, 2017 and March 31, 2020, prehospital resuscitation was performed on 17,166 OHCA patients. There were 8157 patients 18 years or older with nontraumatic OHCA who received continued resuscitation in the ED included for analysis. Eleven demographic and resuscitation predictor variables were extracted to predict survived events, defined as any sustained return of spontaneous circulation until in-hospital transfer of care. Prediction models based on random forest (RF), multilayer perceptron (MLP), and LR were created with hyperparameter optimization. Model performances on internal and external validation were compared using discrimination and calibration statistics. The three models showed similar discrimination performances with c-statistics values of 0.712 (95% confidence interval [CI] 0.711-0.713) for LR, 0.714 (95% CI 0.712-0.717) for RF, and 0.712 (95% CI 0.710-0.713) for MLP models on external validation. For calibration, MLP model had a better performance (slope of calibration regression line=1.10, intercept=-0.09) than LR (slope=1.17, intercept=-0.11) and RF (slope=1.16, intercept= -0.10). Two practical ML-based and one regression-based clinical prediction models of nontraumatic OHCA for survived events were developed and validated. The ML-based models did not outperform LR in discrimination, but the MLP model showed a better calibration performance.