Burst suppression is a unique electroencephalogram (EEG) pattern commonly seen in cases of severely reduced brain activity such as overdose of general anesthesia. It is important to detect burst suppression reliably during the administration of anesthetic or sedative agents, especially for cerebralprotective treatments in various neurosurgical diseases. This study investigates recurrent plot (RP) analysis for the detection of the burst suppression pattern (BSP) in EEG. The RP analysis is applied to EEG data containing BSPs collected from 14 patients. Firstly we obtain the best selection of parameters for RP analysis. Then, the recurrence rate (RR), determinism (DET), and entropy (ENTR) are calculated. Then RR was selected as the best BSP index oneway analysis of variance (ANOVA) and multiple comparison tests. Finally, the performance of RR analysis is compared with spectral analysis, bispectral analysis, approximate entropy, and the nonlinear energy operator (NLEO). ANOVA and multiple comparison tests showed that the RR could detect BSP and that it was superior to other measures with the highest sensitivity of suppression detection (96.49%,
The electroencephalographic burst suppression pattern (BSP) consists of high amplitude bursts interrupted by low amplitude suppressions. It can be observed in different clinical conditions (head trauma, stroke, coma, anoxia, and hypothermia) [
Many researchers have investigated methods for BSP detection. Early methods were based on the spectral analysis, such as the spectral edge frequency and the median frequency [
Recurrence quantification analysis (RQA) [
The paper is organized as follows. In Section
The data used in this study were obtained from a previously reported study on dreaming during general anesthesia [
Artifacts in scalp EEG recordings mainly come from eye movement, muscle activity, and power frequency noise. To reduce these artifacts, the following steps are carried out. First, statistical mean and standard deviation methods were used to remove the outlier points. Then, a stationary wavelet transform [
The process of BSP detection is depicted with a block diagram in Figure
The block diagrams of EEG signal processing.
(a) A composite EEG signal from a patient. It consists of suppression (1000point), burst (1000point), and normal (1000point) records artificially joined together; (b) different RP patterns during suppression, burst, and normal states, respectively. The blue box represents the suppression, the green box the burst, and the red box the normal state.
The ratio of recurrence points on the diagonal structures to all recurrence points is called determinism (DET). The DET is a determinism (or predictability) measure of a system, calculated by
The ENTR is considered as a complexity measure of a deterministic structure in a dynamical system. The ENTR refers to the Shannon entropy of the frequency distribution of the diagonal line lengths. The more complex the deterministic structure, the larger the ENTR value. ENTR is calculated as
To test the performance of the three RQA indexes RR, DET, and ENTR to detect BSP in the EEG series, the oneway ANOVA and multiple comparison tests were performed on averaged RR values. We also compared RR and NLEO for detecting and classifying BSP with confusion matrixes [
Prior to calculating a recurrent plot index from EEG data, the phase space reconstruction should first be determined. In consideration of the nonstationary characteristic of the EEG signals exhibiting rather sudden changes of state, the notion of a “correct” embedding or delay is inappropriate—as demonstrated by Grassberger and Schreiber [
The embedding dimension and delay time of the EEG signals during the burst suppression state. (a) The false nearest neighbors versus the dimension with scales from 0 to 40. (b) The local plot of (a) with the dimension scales from 1 to 10. (c) The mutual information versus the delay time with scales from 0 to 40. (d) The local plot of (c) with the delay time scales from 1 to 10.
The first local minimum of the mutual information measure was used to determine the time delay parameter [
Another crucial parameter of RP is the radius
(a) An EEG signal consists of suppression and burst; (b) the different RP under four different radiuses for the signals in (a). (A)
The comparison of the three RQA measures, and, hence, selection of the BSP index, is another important issue. Figure
The boxplot of three different indexes at the burst suppression normal states. (a) The RR index, (b) The DET index. (c) The ENTR index.
As can be seen in Figure
To evaluate the performance of the three RQA indexes, we applied the oneway repeated measure ANOVA and multiple comparisons. As shown in Table
Oneway ANOVA and multiple comparison test of three RQA indexes.
Oneway ANOVA  Multiple comparison test  







RR 

<0.001 



DET  246.10  <0.001 



ENTR  920.06  <0.001 



Multiple comparison tests showed that all three indices could distinguish between the burst and suppression states and the burst and normal states (difference of mean >0). However only the RR measure could distinguish between the suppression and normal states (the other two indexes’ difference contains 0 and thus are not significant). So the RR measure was chosen to be the index of the BSP identification.
In order to detect BSP quantitatively and automatically, the optimal value of threshold
Four methods (spectral analysis, bispectral analysis, approximate entropy, and NLEO) have been employed for the detection of BSPs. In the following section, the RR is compared with the above mentioned four methods.
First, we discuss the spectral analysis based methods. Figure
Comparison between the RR method and the spectral analysis based methods for the BSPs detection. (a) A burst suppression interval EEG signal of 80s, (b) frequency spectrum, (c) spectral edge frequency 95 parameter, (d) median electroencephalogram frequency parameter, (e) the RR index, and (f) the BS index of RR.
The bispectral analysis is another method used for BSP detection. In Figure
The bispectrum of burst, suppression, and normal states, respectively.
Approximate entropy has also been proposed to detect the BSP [
The boxplot of two different indexes at the burst, suppression, and normal states. (a) Approximate entropy. (b) The RR index.
Nonlinear energy operator (NLEO) is one of the most popular methods in burst suppression pattern detection. The NLEO and RR methods were used to analyze the EEG signals of all subjects and the statistical results are shown in Table
The classification of manual, NLEO, and RR methods.
Manual  NLEO  RR  

Burst  427  459  426 
True burst  —  408  412 
False burst  —  51  14 


Suppression  427  395  428 
True suppression  —  376  413 
False suppression  —  19  15 
The confusion matrix between the NLEO and the RR.
NLEO  RR  

Burst (%)  Suppression (%)  Burst (%)  Suppression (%)  
Manual  
Burst (%)  95.50  4.50 

3.51 
Suppression (%)  11.94  88.06  3.28 

To assess the changes of RR over time, the RR of the longterm EEG records for 14 subjects was obtained. The RR was calculated on 10 s EEG moving window, and parameters for RR were
Figure
(a) The longterm EEG recordings with burst suppression patterns. (b) The observation of RR over the entire EEG recordings. (c) Suppression is represented with 0 and burst with 1 to obviously distinguish the two states. (d) The BSR is calculated.
Previously, several methods have been proposed to detect the burst suppression pattern of EEG signals [
The advantage of RR is that it does not have constraints and assumptions, because it only counts similar events in an embedded space [
The RR index is not very sensitive to choice of threshold, because it is based on the distance of different dots and is independent of the signal amplitude. Amplitude differences between individual recordings were eliminated through the normalization of the RR index. In contrast, the NLEO method is very sensitive to the choice of an appropriate threshold. Thus we would conclude that the RR method is more robust than other methods and is suitable for further development of a BSP detector.
The authors declare that there is no conflict of interests regarding the publication of this paper.
This research is supported by National Natural Science Fund for Distinguished Young Scholars of China (no. 61025019) and National Natural Science Foundations of China (nos. 61304247, 61105027, and 61203210).