A Machine Learning Framework for Pulse Detection during Out-of-Hospital Cardiac Arrest

The availability of an automatic pulse detection during out-of-hospital cardiac arrest (OHCA) would allow the rapid identification of cardiac arrest and the prompt detection of return of spontaneous circulation. The aim of this study was to develop a reliable pulse detection algorithm using the electrocardiogram (ECG) and thoracic impedance (TI), the signals available in most defibrilators. The dataset used in the study consisted of 1140 ECG and TI segments from 187 OHCA patients, whereof 792 were labelled as pulse-generating rhythm (PR) and 348 as pulseless electrical activity (PEA) by a pool of experts in OHCA. First, an adaptive filtering scheme was used to extract the impedance circulation component and its first derivative from the TI. Then, the wavelet decomposition of the ECG was carried out to obtain the different subband components and the denoised ECG. Pulse/no-pulse (PR/PEA) discrimination features were extracted from those signals and fed into a support vector machine (SVM) classifier that made the pulse/no-pulse decision. A quasi-stratified and patient wise nested cross validation procedure was used to select the best feature subset and to tune the SVM hyperparameters. This procedure was repeated 50 times to estimate the statistical distributions of the performance metrics of the method. The optimal solution consisted in a five feature classifier that yielded a mean (standard deviation) sensitivity, specificity, balanced accuracy and total accuracy of 92.4% (0.7), 93.0% (0.8), 92.7% (0.5) and 92.6% (0.5), respectively. When compared to available methods, our solution presented an improvement in balanced accuracy of at least 2.5 points. A reliable pulse detection algorithm for OHCA using the signals available in defibrillators was acomplished.