Automatic Cardiopulmonary Resuscitation Quality Estimation from Photopleythysmography Signal using Convolution Neural Network

Abstract

Cardiopulmonary resuscitation (CPR) is a first aid procedure to preserve the function of the body with chest compression and artificial respiration. The CPR guideline recommends compressing the chest to a depth of 5cm between 100 – 120 times per minute. However, it is not appropriate to perform CPR in accordance with the guideline to all patients because of their different physical characteristics. In the United States, about 600,000 patients with cardiac arrest occur each year and the survival rate of cardiac arrest patients outside the hospital is about 10%, which is considerably lower than survival rate of patients in hospital. According to the American Heart Association(AHA), the quality and coping speed of CPR have a significant impact on the survival rate of patients with cardiac arrest. A study to monitor the patients condition during CPR and detecting the return of spontaneous circulation (ROSC) are progressing actively for monitoring the quality of CPR. End-tidal carbon dioxide (ETCO2), trans-thoracic impedance (TTI), arterial blood pressure, and other bio-signals have been used in the research, which suggest many possibilities of monitoring CPR. These signals, however are invasive or time consuming. Among the various bio-signals, the photoplethysmography (PPG) used in this study is noninvasive and has the advantage of reflecting information in real time. The goal of this study is to estimate the quality of CPR using convolution neural network of PPG signal. The PPG and ETCO2 used in this study were obtained from preclinical experiments using 15 pigs. Cardiac arrest was induced in each pig, and both data were acquired for 290seconds during CPR in cardiopulmonary induction pigs. The PPG data was divided into 5 seconds and stored again. Each of the stored data was converted into images by applying spectrogram transformation and wavelet transformation. The data stored as an image contains the ETCO2 value corresponding to the same time interval. The quality of CPR was divided into two classes. The criteria of the two classes was based on ETCO2 values which were obtained during CPR of survived pigs and non-survived pigs. All image data were divided into two classes based on the criteria. The obtained data in this study were trained using the VGG based neural network. The training set of data were 90% of total data and test set of data were 10% of total data. The classification accuracy is 84.09% when the image is applied with the spectrogram transformation of the PPG signal and classification accuracy is 88.37% when the wavelet transformation is used. The result of this study that training of the transformed image from the PPG signal using the convolution neural network is effective in CPR quality estimation.