Quantitative Analysis of EEG Slow Wave Activity Based on MinPeakProminence Method

Electroencephalography (EEG) is a technique where the electrical activities of the brain are recorded and it has an important place in the diagnosis of neurological diseases. With the developments in signal processing studies in recent years, it is possible to reveal the hidden and important signals in EEG. For example, changes in amplitude and frequency in background activities may indicate a lesion. In addition, epileptiform activities determined in the EEG signal also contribute greatly to the diagnosis of epilepsy [1].

Epilepsy is a chronic neurological disease that occurs as a result of abnormal and excessive electrical discharge in cortical neurons and the essential syndrome which is characterized by recurrent epileptic seizures [2].

In the most common definition made by the International League Against Epilepsy (ILAE), epilepsy is a transient functional disorder of the central nervous system caused by excessive and synchronous abnormal neuronal activity in the cortex [3-7]. This neurological disease, which occurs as a consequence of recurrent seizures, negatively influences around 1% of the world population [8]. Even with the use of various medications and surgical treatment, seizures cannot be controlled in more than 25% of epilepsy patients. Due to the incapability to detect the occurrence of seizures in advance, sudden and unexpected deaths occur in some patients with this disease or the risk of morbidity increases [9].

The World Health Organization (WHO) reports that about 50 million people worldwide have epileptic seizures, according to epidemiological data. It was revealed in a recent study that epilepsy is one of the severest disorders among neurological diseases [10].

It is known that epilepsy is the most common in childhood and adolescence, and the second most common after cerebrovascular diseases in adults. In spite of its widespread prevalence, various difficulties are encountered in establishing the definitive diagnosis of epilepsy, especially in distinguishing it from psychogenic seizures. To eliminate these difficulties and to obtain more precise and clear results, Video-EEG imaging (VEEG) method is used. However, this imaging technique, a gold standard in the differentiation of non-epileptic seizures and epilepsy, has a limited use because of the difficulties faced in its application and its high cost. Due to these reasons, this technique is often preferred to be used in situations where a decision cannot be made with anamnesis, clinical examination, and EEG [11].

Before starting EEG, an EEG procedure suitable for the patient should be selected, taking into account their clinical information. The recorded signals should be carefully examined and interpreted because spikes or other recorded abnormalities that have been observed contain useful information in diagnosis and treatment [12]. Despite the advances in technology, the EEG recordings are mostly interpreted manually. Manual review/examination of EEG recordings of 20-30 minutes or even hours takes quite a long time [13, 14]. To make a faster and more accurate diagnosis, various signal processing algorithms and computer programs have been developed in recent years [15]. In studies regarding automated EEG analysis, seizures, spikes, and sharp waves, mostly indicative of epileptic activity, were examined [13, 14]. Focal slow waves in the delta and theta frequency range are often evaluated as psychopathological. Therefore, such waves with high amplitude and low frequency activity are very important in the electroencephalogram. Wienbruch et al. aimed to detect focal slow wave activities. In their study, they compared the records of the patients diagnosed with schizophrenia and affective disorder with normal records [16]. In another study, a method was developed for detecting Slow Wave Sleep (SWS) on an automatic single channel in sleep EEG. With this study, the uncertainty caused by individual differences was minimized, and the effects of the SWS ratio on the effectiveness of sleep staging were investigated [17]. On the other hand, there has not been sufficient number of studies on the analysis of slow-waves since they vary more greatly than the morphology of sharp transients [18]. However, Slow Waves (SWs) are as important as spikes in both clinical diagnosis and automatic detection [19]. For this reason, in this study, a new MinPeakProminence-based (MPP) method has been developed for the detection of Slow Wave (SW), which is an indicative of EEG anomaly. The aim of this method is to determine the relationship of SW activities with epilepsy and other neurological diseases and to detect which regions of the brain are frequently seen (focal, lateralized, and generalized). Consequently, in this study, a new MPP-based method has been generated for the detection of SWs, which is the sign of EEG anomaly. With this method, it is intended to find out the relevance of SW activities with epilepsy and other neurological diseases and to locate which regions of the brain are commonly seen (focal, lateralized, generalized).

The other parts of this paper are organized as follows. In Section II, the methods that form the basis of the whole study and the datasets used are mentioned. The experimental results obtained with the performed method are discussed in Section III. Finally, Section IV presents a brief summary of the study and the scope of future research.

General background slowing (theta and delta frequency ranges) is a normal finding on EEG when it signifies developmental slowing or the development of lethargy and sleep activity in children, adolescents, and some young adults. The background EEG is generally normal in patients with epilepsy [50]. On the other hand, general or focal background slowing can be seen in the EEG records of epilepsy patients. When it is not diagnosed accurately, patients with psychogenic non-epileptic seizures (PNES) encounter serious and multifaceted problems such as delay in starting the correct treatment, the long-term financial burden of antiepileptic drugs, and possible side effects.

An epileptic seizure manifests itself with sudden changes in sensory-motor functions, behaviour, memory, and consciousness. In general, the way the seizure manifests itself behaviourally varies depending on the localization and area of the affected brain area [51]. Epilepsy classification is most frequently made according to this clinical situation. According to this classification, epileptic seizures are examined in two parts: partial seizures affecting only a certain part of the brain and not always accompanied by loss of consciousness, and generalized seizures affecting the whole brain, always with loss of consciousness [52]. Etiologically, epilepsies are examined in two groups as primary or idiopathic and secondary epilepsies [53].

The main question that needs to be answered in patients coming to the clinic with loss of consciousness is whether the situation is an epileptic seizure. Since the seizure has already stopped when the patient is taken to the hospital, the most critical information in distinguishing this can be gathered from the patient or the eyewitnesses present during the loss of consciousness. In cases where anamnesis cannot be taken adequately, it becomes difficult to make a differential diagnosis to the patient. Besides, having a single seizure does not mean that a person has epilepsy. Epilepsy, as mentioned earlier, is a disease that occurs without any triggering cause and is defined by recurrent (two or more) seizures [54]. For this reason, it is critically essential to consider all epileptic activities that can be observed in EEG recordings. The main focus of recent studies has been on transient activities, which are signs of epileptic activity. However, the detection of slow waves is at least as critical as the detection of transient activities [19]. That being the case, a new MPP-based method has been developed to determine the relevance of SWs, which are the signs of EEG anomalies, with epilepsy and other neurological diseases, and to determine which regions of the brain are frequently detected (focal, lateralized, generalized).

For a better understanding of the working principle of the method, all examples are explained using a single data (suh-01) in the SUH dataset. The image of a 15-second EEG recording obtained from subject suh-01 data is shown in Figure 9. Even though the ground activities generally include alpha waves, SWs are encountered in some channels, as can be seen in the Figure 9. Photic Stimulation is one of the activation methods applied during routine EEG recording. Eyelid myoclonus, generalized myoclonic jerks, absence seizures, generalized tonic clonic seizures, focal seizures, and more rarely tonic versive seizures and focal asymmetric myoclonic seizures can be triggered by photic stimulation [55]. In addition, the emergence of generalized discharges consisting of spike-slow wave, sharp wave-slow wave and multiple spike-slow wave elements in epilepsies generalized with photic stimulation becomes easier. In this technique, a bright light stimulus flashing at different frequencies for 5-10 seconds is applied to a patient from a distance of 30 cm with eyes open and closed. The blue line in Figure 9 and the blue line shown at the bottom (photic stimulation) was applied for 10 seconds. The fact that this blue line continues straight means that photic stimulation is not applied.

Detecting background activities that indicate the anomaly is quite significant for diagnosis and treatment. Therefore, the signal that Discrete Wavelet Transform (DWT) or any pass filter (low pass, high pass, etc.) is not applied has been examined with threshold and MPP. For the signal analysis here, various studies have been conducted on the peak values of the original signal and the peaks obtained after applying the threshold and MPP. That's why the focus has been on the signals belonging to all channels in the second of the subject suh-01 dataset. First of all, 1st and 2nd channels of these signals in the 2nd second were examined. Various trials were carried out by giving different values for Threshold and MPP. As it is seen in Figure 10, the peaks and peak numbers obtained with both techniques are indicated. All peak numbers (original signal, threshold, and MPP) obtained as a result of the trials are taken into a matrix. Afterward, the peak numbers obtained by threshold and MPP were subtracted from the original peak numbers and the differences between them were analysed. Analysing the obtained differences, it was determined that the channels with the least subtraction result had SW as given in Figure 10. The differences between all the peak numbers in each dataset were obtained through the program written in the Matlab program.

The numerically obtained differences between the peak numbers of the original signal obtained before any processing and the peaks obtained as a result of the threshold and MPP are shown in Table 8. The difference 1, difference 2 and difference 3 indicate the values obtained by the MPP method, while the difference 4, difference 5 and difference 5 indicate the values obtained by the Threshold method, respectively. In Table 8, the values where the difference between the peak numbers of the original signal and the peak numbers obtained with the threshold and MPP techniques is 3 and below 3 are marked with a green background colour. These values marked signify possible SW. Examining the table, when MPP 2 and 5 values were given, 8 and 4 SWs were found, respectively. On the other side, 3 and 1 SWs were found for threshold 2 and 5, respectively, but no SW was found for the value 7. Signals that may be SWs were almost never detected, especially in trials with the threshold. As it is shown in the table, when the value 7 for MPP is given, SWs marked manually by a neurologist were found to be high. That's why the value 7 was taken as the basis for the MPP based method implemented, and this value was used on all the next dataset. SUH and Bonn datasets were used to test the reliability and accuracy of the MPP method.

To determine which channels and on which pages the SWs are seen frequently, the data were processed page by page. Besides, it was determined in which channels it was frequently seen by taking the average SWs value for each channel. In this way, it was determined in which areas of the brain the SWs are seen or whether they are seen in all (focal, lateralized, and generalized).

Statistical measures of the performance, commonly used in scientific studies, were used to evaluate the success of the method. As it is known, the confusion matrix is one of the easiest and well-known methods used to interpret the success of classifiers [56]. This confusion matrix provides us with detailed information regarding correct and incorrect predictions in classification. In this method that generally evaluates the performance of data in the form of matrixes, there are four possible results: True Positive (TP), True Negative (TN), False Positive (FP), and False Negative (FN). Explanations of these possible results are given in Table 9 [57, 58].

9.png

Figure 9. 18-channel EEG recording (only 15 sec) of the subject suh-01. Generally, the ground activity is well developed in the posterior areas of the hemispheres and includes 8-13 Hz alpha waves. However, SW is observed in some of the channels marked

10a.png

10b.png

Figure 10. Peak detection trials of the signals in the 2nd second of channels 1 and 2 of the subject suh-01. a) 1. Channel on 2nd second, b) 2. Channel on 2nd second

The performance of a classifier is assessed by statistical parameters such as Sensitivity (SEN), specificity (SPE), and classification accuracy (ACC) (Eqns. (1), (2), and (3)).

SEN=TP/(TP+FN)          (1)

SPE=TN/(TN+FP)              (2)

ACC= (TP+TN)/(TP+TN+FP+FN)               (3)

Examining the classifier success (ACC) alone is not always adequate to evaluate the success of a system. Especially in imbalanced datasets, it will be beneficial to examine the parameters of precision (Eq. (4)) and F-score (Eq. (5)) together with these parameters. The Positive Predicted Value (PPV) or Precision value provides information regarding how many of those identified as patients are actually sick, and the F-score (also F-measure or F1-score) is used to calculate the truth of a test and it balances the use of precision. The F-score value can provide the measurement of a test's performance in a more realistic way through both precision and recall.

PPV=TP/(TP+FP)         (4)

F-score=2TP/(2TP+FP+FN)           (5)

Table 8. The peak numbers obtained in the 2nd second of all channels of the subject suh-01 (Available/No Available (A/N))

Table 9. Possible results on confusion matrix

The average of ACC values obtained with the confusion matrix found for the SUH dataset is given in Table 10. Comparing the average ACC values obtained by the method performed with clinical findings, it was found that there was a high rate of SWs in places where the average value was 1 and greater than 1 (channel numbers with SWs are marked in the relevant table for each data). For this reason, those with an average ACC value less than 1 were not taken into account.

For a better analysis, the average ACC values given in Table 11 were plotted with glyphplot (Figure 11). In this way, it is better to analyse which channels have more SWs by examining these values visually for each channel. The graph shows that the channels with the most SWs are the 8th (P3-O1) and 18th (P4-O2) channels. However, in general, a considerable number of SWs were detected in the 3rd (T3-T5), 4th (T5-O1), 10th (CZ-PZ), 13th (T4-T6), and 14th (T6-O2) channels. When the electrode connection points of the channels where SWs are most frequently seen were examined, it was found that they were in both hemispheres (both right and left), around the occipital area, that is, in the posterior part of the skull (Figure 12). This indicates that such SWs are observed in the posterior areas of the brain, particularly in the temporal, parietal, and occipital regions.

Table 10. Confusion matrix for the SUH dataset

Table 11. Average of SWs numbers found for all channels in the SUH dataset

11.png

Figure 11. The indication of the peak averages monitored by channels through glyhplot for the SUH dataset

The number of SWs seen in the channels is given in Table 12. Examining the frequency values 6 and higher than 6 in the table, it was determined that the channels with the most SWs are T3-T5, T5-O1, P3-O1, CZ-PZ, T4-T6, T6-O2, and P4-O2.

12.jpg

Figure 12. The indication of the areas where SWs are most frequently seen in the electrode connection (channels marked in blue are the channels with the most SWs), according to the results obtained (SUH dataset)

Also, all SWs found for the SUH dataset were examined with a box plot on a channel basis. A box plot (also known as box and whisker plot) is a type of chart used to visually indicate the distribution of quantitative data, the distributions and skewness between variables or between levels of a categorical variable. The summary of a dataset consists of five parts: minimum score, first (lower) quarter, median, third (upper) quarter, and maximum score (Figure 13).

Table 12. Frequency of SWs by channels for the SUH dataset (numerically)

13.png

Figure 13. Boxplot [59]

In the box plot, the analysis is made according to the location of the median on the box. If the median is in the middle of the box and its whiskers are equidistant from both sides of the box, the distribution is symmetrical. If the median is closer to the bottom of the box and the whisker is shorter at the lower end of the box, the distribution is positively skewed. Finally, if the median is closer to the top of the box and the whisker is shorter at the upper end of the box, the distribution is negatively skewed. Considering these characteristics of the box plot, distributions and distortions between average ACC values can be visually analysed. Examining the variables given in Figure 14, it is observed that all the median values are close to the bottom of the box, and again they are quite short on their whiskers. Consequently, it can be said that the distribution is positively skewed. Besides, examining the graphic in general, it is quite obviously observed that the frequency of SWs is the highest in channels 3, 4, 8, 13, and 18. Interquartile range give us information about how the data is dispersed among each sample. In other words, as the box gets longer, the data shows more dispersal. On the other hand, the data is dispersed less as the box get smaller. When we look at Figure 14, the box showing the 8th channel is longer than the other boxes. In this case, it is seen that there are more SWs dispersals on this channel. The whiskers at the ends of each box give us information about the general spread. Looking at the whiskers in this figure, a wider spread was observed in channels 4 and 13, especially in the 8th channel. In other words, the data are observed as more dispersed in these channels.

Besides, it was examined in which areas of the brain SWs are seen more frequently, according to the results. The results of the analysis achieved in this context are given in Table 13. Examining the table, it was observed that slow waves emerge as generalized rather than focal or lateralized. It was determined that especially the multifocal occurrence, coded as D (at least 2 focus, each in different hemispheres), is more than the others. That being the case, it was determined that SWs are observed as multifocal in EEG recordings and, as discussed earlier, they are most frequently observed in the occipital area that includes both sides of the brain.

14.png

Figure 14. The indication of the average ACC values obtained with Boxplot (SUH dataset)

Table 13. The frequencies of SW values by the brain regions for the SUH dataset

Statistical performance measurement results are presented in Table 14 in its most general form. All TP, TN, FP, FN values observed for the SUH dataset, and the results of the classifier metrics obtained with these values are seen in detail. The lowest ACC value is 92.7%, and the highest ACC value is 98.5%. Looking at the PPV values in the table, the rate of correct detection was determined as 90.9% of those determined as SW. It was observed that the method performed according to the results has a high rate of success.

15.png

Figure 15. The graph of ACC on the SUH dataset

The main reason for using the F-Score value is not to make an incorrect model selection in unevenly distributed datasets. Besides, F-Score is a significant statistical metric when a measurement metric that includes not only FN or FP but also all error costs is needed. When the F1 score is 1, the model is considered perfect, and when it is 0, it is considered unsuccessful. Looking at Table 14, it is seen that F-score values are very close to 1 (average 92%).

As it is seen in Figure 15, the classification success achieved for each data (22 subjects) is generally above 90%. The general average ACC value is 96.5% (Table 15). The method performed according to these results has a high rate of success in detecting SWs.

The results achieved were analyzed graphically with polynomial fitting (Figure 16). The curve fitting process was conducted with the Linear model Ploy3. As can be seen in the Figure 16, some data such as suh-06, suh-09, and suh-11 stay out of the fitting curve. However, most of the data collected as a result of the best model and the curve fitting process are appropriate for the model. Looking at the Root Mean Square Error (RMSE) value of the method performed, it has a small error rate of 0.01598.

The method performed on the second dataset (Bonn dataset) was implemented using the same values (7 value for MPP and average <=3). All average SW values obtained on the Bonn dataset are presented in Table 16. The averages of 1 and above 1 from the average SW values obtained for the SUH dataset were taken into account. Channels above 1 and 1 in the Bonn dataset must have SWs. However, when the table is examined, all the average values are below 1. This proves that there is no SW in the Bonn dataset (normal data - A) used in the study. Consequently, thanks to the method performed, while the SUH dataset had a high SW rate of 96.5%, a 100% success rate (ACC) was achieved in the Bonn dataset without SW.

Table 14. Experimental results using the proposed method on the SUH dataset (PPV, SEN, SPE, F-score, ACC)

Table 15. Overall statistics in the SUH dataset

16.png

Figure 16. Graph of overall ACC on SUH dataset

Table 16. Average of SW on Bonn dataset