Prediction of cerebral perfusion pressure during CPR using electroencephalogram in a swine model of ventricular fibrillation.

BACKGROUND:Measuring the quality of cardiopulmonary resuscitation (CPR) is important for improving outcomes in cardiac arrest. Cerebral perfusion pressure (CePP) could represent cerebral circulation during CPR, but it is difficult to measure non-invasively. In this study, we developed the electroencephalogram (EEG) based brain index (EBRI) derived from EEG signals by machine learning techniques, which could estimate CePP accurately in a porcine cardiac arrest model. METHODS:We conducted a randomised crossover study using nine female pigs. After 1 min of untreated ventricular fibrillation, we performed CPR with 12 different 2-min tilting angle sessions, including two different head-up tilt (HUT) angles (30°, 15°) twice, horizontal angle (0°) four times and two different head-down tilt (HDT) angles (-15°, -30°) twice with the random order. We collected EEG signals using a single channel EEG electrode in real-time during CPR. We derived the EBRI models to predict the CePP classified by the 5 or 10 groups using three different machine learning algorithms, including the support vector machine (SVM), k-nearest neighbour (KNN) and random forest classification (RFC) method. We assessed the accuracy, sensitivity and specificity of each model. RESULTS:The accuracy of the EBRI model using an SVM algorithm in the 5-group CePP classification was 0.935 with a standard deviation (SD) from 0.923 to 0.946. The accuracy in the 10-group classification was 0.904 (SD: 0.896, 0.913). The accuracy of the EBRI using the KNN method in the 5-group classification was 0.927 (SD: 0.920, 0933) and in the 10-group was 0.894 (SD: 0.880, 0.907). The accuracy of the RFC algorithm was 0.947 (SD: 0.931, 0.963) in the 5-group classification and 0.920 (SD: 0.911, 0.929) in the 10-group classification. CONCLUSION:We developed the EBRI model using non-invasive acquisition of EEG signals to predict CePP during CPR. The accuracy the EBRI model was 0.935, 0.927 and 0.947 for each machine learning algorithm, and the EBRI could be used as a surrogate indicator for measuring cerebral perfusion during CPR.