Abstract 221: Machine Learning Algorithms For The Prognostication Of Return Of Spontaneous Circulation

Abstract

Introduction: Several factors are known to influence return of spontaneous circulation (ROSC) following out-of-hospital (OHCA) and in-hospital cardiac arrest (IHCA). Machine learning (ML) methods are capable of analyzing large datasets to elucidate the clinical and prognostic value of specific variables. In cardiac arrest, ML may help identify predictors of ROSC in OCHA and ICHA. Purpose: The present systematic review summarizes the literature on ML algorithms used to predict ROSC in OHCA and IHCA. Methods: PubMed, EMBASE, Web of Science, and Cochrane were searched to identify articles. Studies on human subjects with OHCA or IHCA which used ML methods to predict ROSC were included. Results: A total of 4,094 studies were identified through a literature search. Ten were included in the final analysis with a total sample size of 240,798 patients (240,798 (93%) male). Studies used logistic regression (n=7), random forest models (n=2), or deep learning (n=1). Sociodemographic variables beyond age (n=8) and sex (n=5) were not used. Variables used for predicting ROSC included age (n=8), type of rhythm (n=7), witnessed arrest (n=5), and time to emergency services arrival (n=5). Most models were built using data from OHCA patients (n=7), with only three focused on IHCA. Studies utilized a variety of parameters for reporting the prognostic value of their models, including predictive accuracy (69%-99%), area under the receiver operating characteristic curve (0.71-0.83), sensitivity (50.2%-76.0%), and specificity (61.0%-92.9%). Conclusion: Despite the prevalence of cardiac arrest and advances in AI, only ten studies examined ML in the prognostication of ROSC. Logistic regression was predominantly used in the included studies, and there is a paucity of data on the efficacy of deep learning models. Furthermore, heterogeneity in the predictive efficacy of studied ML models merits additional work and the need for large-scale trials comparing ML to conventional methods and clinician judgment. Critically, there is also a need for greater data inclusivity in model development, understanding that marginalized populations are less likely to receive CPR.