Abstract 14468: Machine-Learning Identifies Out-of-Hospital Cardiac Arrest Patients With Higher Probability of Survival While Undergoing Extracorporeal Cardiopulmonary Resuscitation

Abstract

Introduction: Neurologically favorable survival following OHCA depends on many factors including patient age, presenting rhythm, bystander CPR, CPR duration, and return of spontaneous circulation (ROSC). This study aims to utilize machine learning to construct predictive models of neurologically favorable survival in OHCA patients receiving standard ACLS and extracorporeal cardiopulmonary resuscitation (ECPR). Methods: 3011 patients from the Amiodarone, Lidocaine, Placebo Study (ALPS) were analyzed using supervised and unsupervised machine learning approaches based on several variables available at the time of hospital presentation after OHCA. Machine learning platforms were developed to create real world predictive models of resuscitation. The model was then refined for 180 patients from the UMN-ECPR study with refractory VF/VT cardiac arrest who underwent ECPR. Resuscitative variables were selected and ranked on the likelihood ratio of predicting survival. Results: Machine learning analysis of the ALPS cohort yielded a model of survival with an area under the receiver-operating characteristic curve (AUC) of 0.92. ROSC, CPR duration and age were the highest determinants of neurologically favorable survival. The model was tested on the UMN-ECPR cohort, showing an AUC of 0.74 with a sensitivity and negative predictive value for survival of 12% and 60%, respectively. Machine learning was again used to render a model tailored for the UMN-ECPR group. The ECPR predictive algorithm had an AUC of 0.83, sensitivity and negative predictive value for survival of 70% and 80%, respectively, in which presenting rhythm to the cath lab, ROSC and CPR duration were the main determinants of neurologically favorable survival. Conclusion: Simple and effective predictive algorithms can predict the outcomes of cardiac arrest using only cardiac arrest characteristics. Use of ECPR changes the predictive determinants substantially. While this information is helpful in the appropriate context, the specificity required to withdraw care or halt resuscitation will require further data points and refinement.