Deep Learning for Pulse Detection in Out-of-Hospital Cardiac Arrest Using the ECG

Abstract

Pulse detection during out-of-hospital cardiac arrest is necessary to identify cardiac arrest and detect return of spontaneous circulation. Currently, carotid pulse checking and checking for signs of life are the most widespread procedures for pulse detection, but both have been proven inaccurate and time consuming. Automatic methods that could be integrated in Automated External Defibrillators (AEDs) are needed. In this study we propose a deep neural network classifier to detect pulse using exclusively the ECG. We extracted 3914 segments of 4s from 279 patients, all of them with an organized rhythm. They were annotated as pulsed rhythm or pulseless rhythm based on clinical information. A total of 2372 pulsed rhythms and 1542 pulseless rhythms were included in the study. To train and test the model 3038 (223 patients) and 876 segments (56 patients) were used, respectively. The model was evaluated in terms of Sensitivity (Se) and Specificity (Sp) for pulse detection. The network showed a Se/Sp of 89.4%/97.2% during training process and 91.7%/92.5% for the test set.