EEG-Based Prediction of the Recovery of Carotid Blood Flow during Cardiopulmonary Resuscitation in a Swine Model

Figures

• 

  Figure 1. Flow chart of the study design [17]: After the initial experimental setup was completed (Step 1), VF was induced and an animal was left untreated for 1 min (Step 2). Four consecutive 2−minute BLS sessions were performed by the bystanders (Step 3). The EMT was supposed to arrive at the site and ECG was checked (Step Q1). If ECG was shockable, a biphasic defibrillation shock of 200 J was applied by the EMT (Step 4-1). Otherwise, an additional BLS session was performed (Step 4-2) and ECG was checked again (Step Q1). If ROSC occurred after the defibrillation attempt (Step Q2), the monitoring session was initiated (Step 5) and sustained ROSC was confirmed if ROSC lasted for 20 min (Step Q3 → Step 6). However, if ROSC did not occurred (Step Q2) or VF occurred again during the monitoring session (Step Q3), ACLS sessions for 2 min with an injection of epinephrine were performed up to 10 times (Step Q4, 7). If ROSC occurred (Step Q2), the monitoring session was initiated (Step 5). Otherwise, an additional ACLS session was performed (Step Q4, 7). Non-ROSC was confirmed if sustained ROSC was not achieved after the 10th ACLS session (Step Q4 → Step 6).

• 

  Figure 2. Total developmental stages of the prediction models; from animal experiments to performance evaluation: the frontal EEG and CBF data were acquired during the process of CPR. CBF recovery rates were calculated and categorized into two groups with a division criterion of 30%. Time- and frequency-domain EEG parameters and entropy endices were determined. The top 10 significant EEG parameters were selected through the NCA and Student’s t-Test. An EEG-based binary classifier for CBF recovery was established based on LOOCV and SMOTE methods. Finally, the performances of prediction models were evaluated with the confusion matrix and ROC curve.

• 

  Figure 3. EEG changes in animal 1: (a) irregular baseline EEG activity before VF, (b) isoelectric EEG pattern during untreated VF, red dotted lines denote the limits of the isoelectric state (±5 μV), (c) severely contaminated signal during CPR, (d) recovered EEG activity after the 1st ACLS, (e) frequency response of the baseline EEG before VF, (f) frequency response of the isoelectric EEG during untreated VF, (g) frequency response of CPR artifacts, red marks denote the frequency components of CPR artifacts, (h) frequency response of the EEG after the 1st ACLS, a black dotted line in (f−h) is the frequency responses of the baseline EEG (e).

• 

  Figure 4. An oversampling example by SMOTE application: (a) Before applying SMOTE, (b) After applying SMOTE.

• 

  Figure 5. Confusion matrices of classifier models: (a) LR, (b) SVM, (c) KNN, (d) RF, (e) MLP.

• 

  Figure 6. The ROC curves and AUC values of classifier models: (a) LR, (b) SVM, (c) KNN, (d) RF, (e) MLP.