Abstract 220: The Prediction Of Pulseless Electrical Activity Evolution During In-hospital Cardiac Arrest Using Machine Learning

Abstract

Background: Pulseless electrical activity (PEA) is the most common rhythm during in-hospital cardiac arrest (IHCA) with a prevalence around 50%. Knowing the prognosis of PEA evolution towards return of spontaneous circulation (ROSC) could help optimizing both resuscitation maneuvers and pharmacological therapy. The aim of this study was to develop an automatic method to predict the evolution of PEA during resuscitation based on the ECG-waveform. Materials and Methods: The dataset consists of 164 IHCA cases recorded by St. Olav University Hospital (Norway), Hospital of the University of Pennsylvania (USA) and Penn Presbyterian Medical Center (USA). ROSC was verified in 108 cases of the patients by physicians and bioengineers based on episode waveforms and clinical data. PEA segments of 5 sec were extracted from the last 10 min before ROSC or the end of resuscitation therapy. Three machine learning models were designed for segment binary classification based on an SVM (Gaussian) model using: 1) ECG-waveform features (9); 2) QRS-features (8); and 3) both ECG-waveform and QRS-features (17). Ten-fold cross validation was applied to train and test the models, and the performance was given in terms of area under the curve (AUC), sensitivity (Se) to correctly detect cases evolving to ROSC, specificity (Sp) and balanced accuracy (BAC). Results: A total of 780 segments were extracted (472 with ROSC). The median (IQR) for the models with the best feature combination are shown in the Table. The most important ECG-waveform features are associated to the ECG spectral distribution. QRS features that showed relevant information about the evolution of PEA were heart rate median and standard deviation, and QRS width, slope and amplitude. Conclusions: ROSC/no ROSC prediction of PEA segments is feasible using ECG signal information. The combination of ECG-waveform and QRS features enhances performance of predictive model.