Abstract 178: Development And Validation Of The Utstein-based Machine Learning Algorithms To Predict Neurological Outcomes After Out-of-hospital Cardiac Arrest: A Retrospective Analysis Of The Pan-asian Registry

Abstract

Introduction: The proportion of favorable neurologic outcome from the survival after Out-of-Hospital Cardiac Arrest (OHCA) remains poor, esp. in the Asian communities. By using the machine learning (ML) approach, we aimed to develop and validate the the Utstein-based ML algorithms to predict neurological outcomes after OHCA in this population. Methods: We conducted a retrospective analysis of data collected from the registry of Pan Asian Resuscitation Outcomes Study (PAROS) Network, an international and multi-center cohort study of OHCA across 13 countries in this region. All adult (≥ 18 years) EMS-treated OHCA patients between Jan 2009 to May 2018 in the registry were included, with variables followed the Utstein recommendations and conform to a unified taxonomy established by the network. The primary outcome was defined as good neurological outcome with cerebral performance category 1 or 2. Random splitting was used to divide the dataset into the training/validation and testing cohorts at around 2-to-1 ratio. Four supervised ML models, including Random Forest (RF), Gradient Boosting (GB), Extra Trees (ET), and CatBoost (CB) classifiers were employed to construct the prediction models, and performances were evaluated and compared with traditional logistic regression (LR) by the area under the receiver operating characteristic curve (AUC) in the testing cohort. Results: We included 194,300 records for analysis. Of them, 6,342 (3.3%) achieved good neurologic outcome. Among the constructed ML models, RF obtained the best AUC performance (0.924, 95% CI: 0.917-0.932), followed by ET (0.922, 95% CI: 0.915-0.930), CB (0.922, 95% CI: 0.915-0.930), and GB (0.910, 95% CI: 0.901-0.920). Although there was no difference in performance between each ML models, RF performed significantly better than using the traditional LR. The top important features were first arrest rhythm, age, witnessed by EMS, event location, and duration from call to EMS arrival. Conclusions: The ML approach showed excellent discriminatory performance to predict good neurologic outcome for EMS-treated OHCA patients in the Pan-Asian communities. It has the potential to save more life or provide termination-of-resuscitation if successfully implemented in the EMS system.