A Machine Learning-Based Decision Support System for the Prognostication of Neurological Outcomes in Successfully Resuscitated Out-of-Hospital Cardiac Arrest Patients.

This study uses machine learning and multicenter registry data for analyzing the determinants of a favorable neurological outcome in patients with out-of-hospital cardiac arrest (OHCA) and developing decision support systems for various subgroups. The data came from the Korean Cardiac Arrest Research Consortium registry, with 2679 patients who underwent OHCA aged 18 or above with the return of spontaneous circulation (ROSC). The dependent variable was a favorable neurological outcome (Cerebral Performance Category score 1-2), and 68 independent variables were included, e.g., first monitored rhythm, in-hospital cardiopulmonary resuscitation (CPR) duration and post-ROSC pH. A random forest was used for identifying the major determinants of the favorable neurological outcome and developing decision support systems for the various subgroups stratified by the major variables. Based on the random forest variable importance, the major determinants of the OHCA patient outcomes were the in-hospital CPR duration (0.0824), in-hospital electrocardiogram on emergency room arrival (0.0692), post-ROSC pH (0.0579), prehospital ROSC before emergency room arrival (0.0565), coronary angiography (0.0527), age (0.0415), first monitored rhythm (EMS) (0.0402), first monitored rhythm (community) (0.0401), early coronary angiography within 24 h (0.0304) and time from scene arrival to CPR stop (0.0301). It was also found that the patients could be divided into six subgroups in terms of their prehospital ROSC and first monitored rhythm (EMS), and that a decision tree could be developed as a decision support system for each subgroup to find the effective cut-off points regarding the in-hospital CPR duration, post-ROSC pH, age and hemoglobin. We identified the major determinants of favorable neurological outcomes in successfully resuscitated patients who underwent OHCA using machine learning. This study demonstrates the strengths of a random forest as an effective decision support system for each stratified subgroup (prehospital ROSC and first monitored rhythm by EMS) to find its own optimal cut-off points for the major in-hospital variables (in-hospital CPR duration, post-ROSC pH, age and hemoglobin).