EEG analysis in the field of neurology is customarily done using frequency domain methods like fast Fourier transform. A complex biomedical signal such as EEG is best analysed using a time-frequency algorithm. Wavelet decomposition based analysis is a relatively novel area in EEG analysis and for extracting its subbands. This work aims at exploring the use of discrete wavelet transform for extracting EEG subbands in encephalopathy. The subband energies were then calculated and given as feature sets to SVM classifier for identifying cases of encephalopathy from normal healthy subjects. Out of various combinations of subband energies, energy of delta subband yielded highest performance parameters for SVM classifier with an accuracy of 90.4% in identifying encephalopathy cases.
Electroencephalogram (EEG) is a signal which represents the electrical activity of millions of neurons in the brain. The signal is acquired from the surface of the scalp. Since it reflects the neuronal activity of the cerebral cortex, it is used in the diagnosis of diseases which involves the function of the cortical neurons. The “EEG picture” of a disease is often a visual waveform or an abnormal frequency or a hypersynchrony or abnormalities in waveform amplitude.
EEG signals are nonstationary; i.e., the frequency components present in the signal vary with time [
Biomedical signals can be analysed in a better manner using time-frequency analysis [
In our study, we have explored the application of discrete wavelet transform in EEG analysis in cases of encephalopathies. Encephalopathy is a disease of the brain due to malfunction or structural changes resulting from metabolic disorders due to organ dysfunction, chemicals, medications, or injuries [
Demir et al. observed reduction in the alpha, asynchronous slow waves, focal slow activities, triphasic waves, burst-suppression pattern, and generalized or focal spike-sharp activities in encephalopathic EEG [
The EEG data needed for this analysis was collected from patients of encephalopathy and healthy individuals from EEG lab of Neurology Department, Government Medical College, Thiruvananthapuram, Kerala. We studied a sample consisting of 232 EEG epochs of 15 encephalopathic patients and 218 EEG epochs of 12 normal healthy subjects. Encephalopathic cases included hepatic and uremic encephalopathy.
Patients with structural pathology, infections of the CNS, and cerebral vascular insult (confirmed by neuroimaging or other investigations) and patients with clinical picture suggestive of metabolic encephalopathy but without obvious metabolic disturbances detected in the necessary biochemical investigations and metabolic encephalopathy occurring in the background of another neurological illness causing cognitive dysfunction or a degenerative condition were excluded from our study. Normal healthy controls of the study include patients with single episode of syncope, who are clinically found to be normal and whose seizures and structural lesions were ruled out.
EEG epochs of 12-second duration were saved, from the artefact-free region of the recording under the supervision of two neurologists. EEG signals were recorded in EEG machine using NicVue software, in international 10-20 electrode system with 21-channel recording with average reference montage setting.
This is a novel approach proposed by Selesnick et al. combining low pass filtering and sparse filtering [
(a) Raw EEG signal without denoising; (b) EEG signal after LPF-TVD denoising.
EEG signal without denoising
EEG signal after LPF-TVD denoising
It is done by formulating the l1 norm of derivative of x which represents the signal having a sparse derivative. Noisy signal is represented by y=x+w. The optimization problem can be written as
Here, majorization-minimization algorithm (MM Algorithm) proposed by Figueiredo et al. [
The technique of time-frequency analysis has been utilised in EEG analysis in many studies. EEG of epileptic patients was analysed using DWT and transient features like epileptic spikes were identified in time-frequency domain [
DWT makes computation easier and faster by avoiding the redundant data which was processed in continuous wavelet transform. DWT of a signal s[n] was taken by passing it through a series of low pass and high pass filters to analyse low frequency and high frequency components, respectively (see Figure
General structure of discrete wavelet transform.
Here, g[n] and h[n] represent the impulse response of low pass and high pass filters, respectively. The outputs of filters are given in (
After downsampling, sample number decreases to half and scale is doubled [
The EEG epochs were subjected to a filtering process using a combined technique of low pass filtering and total variation denoising proposed by Ivan W Selesnick [
The EEG waves are conventionally classified into delta (less than 4 Hz), theta (4 to 7 Hz), alpha (8 to 13 Hz ), and beta waves (13 to 30 Hz), based on their frequency [
Application of DWT to generate subbands of EEG. Different levels of decomposition of DWT are shown. A6, D6, D5, and D4 yield delta, theta, alpha, and beta waves, respectively. fm: maximum frequency content of the EEG signal.
Thus, the EEG was decomposed into its subbands and their energies were calculated from the wavelet coefficients [
Energy of delta subband:
Energy of theta, alpha, and beta subbands:
As
Various subband energies (expressed as percentage of total energy called relative energy) were given as features to SVM for classifying EEGs of encephalopathic patients from that of normal healthy subjects. SVM is employed in our study as many studies reported good results for support vector machine (SVM) classification and even higher classification accuracy than neural networks in various neurological disorders [
Flow diagram of the study.
Discrete wavelet transform (DWT) was performed on the data. EEG epochs of both normal and encephalopathic patients were decomposed into subbands, namely, delta, theta, alpha, and beta using DWT. Here, as the sampling rate was 500 Hz, maximum frequency was taken to be 250 Hz. Therefore 6-level decomposition was carried out using fourth-order Daubechies (db4) as the mother wavelet. It was selected as mother wavelet because it offers maximum correlation with the EEG signal. Thus the subbands were generated by 6-level decomposition using DWT (see Figure
This is similar to the result given by Kaplan [
Values of mean, standard deviation, and standard error of subbands energies of normal and encephalopathic EEG.
| | | | |
---|---|---|---|---|
| Normal | 218 | 14.96 | 6.85 |
Encephalopathy | 232 | 2.89 | 2.71 | |
| ||||
| Normal | 218 | 30.60 | 6.03 |
Encephalopathy | 232 | 9.64 | 7.68 | |
| ||||
| Normal | 218 | 19.71 | 6.47 |
Encephalopathy | 232 | 19.20 | 11.78 | |
| ||||
| Normal | 218 | 31.39 | 10.34 |
Encephalopathy | 232 | 67.73 | 19.34 |
Various EEG subbands obtained using DWT: (a) normal EEG; (b) EEG in encephalopathy.
Distribution of EEG subband energies in encephalopathy and healthy groups.
Independent sample t-test was done to identify the subband energies which showed significant difference so that they can be potentially used for classifying between encephalopathy group and normal. Delta subband energy (
We have implemented an SVM classifier for the diagnosis of encephalopathy based on the energies of subbands of EEG signal. We used a subset of data to train and subsequently the rest of the data were tested (see Table
Data set for training and testing.
| | | |
---|---|---|---|
| 100 | 100 | 200 |
| |||
| 132 | 118 | 250 |
The features used for classification were energies of all subbands, i.e., delta, theta, alpha, and beta, energies of delta, alpha, and beta individually (as these were found to be significantly different in the statistical tests).
Accuracy, sensitivity, and specificity are mainly used as performance parameters for the classifier. Sensitivity is the ability of the test to find out the diseased cases correctly (TP/TP+FN). Specificity is the ability of the classifier to find out the normal cases rightly (TN/TN+FP). Accuracy may be described as the ability of the classifier to distinguish diseased and normal cases correctly (TP+TN/TP+TN+FP+FN). Test statistics of the classifier are given in Table
Test statistics of SVM classifier based on different feature sets based on subband energies of EEG.
| | | | |
---|---|---|---|---|
| 86 | 121 | 115 | 94 |
| ||||
| 118 | 105 | 110 | 118 |
| ||||
| 0 | 13 | 8 | 0 |
| ||||
| 46 | 11 | 17 | 38 |
Performance parameters of SVM classifier based on different subband energies of EEG. Accuracy, sensitivity, and specificity are colour coded in the bar diagram.
Normal EEG background activity in an person who is awake is in the alpha range in the posterior head region. The initial EEG changes in encephalopathy are mild slowing of background which is reactive to external stimuli, followed by intermittent polymorphic delta activity. With worsening encephalopathy, there is continuous polymorphic delta activity persisting >80% of the record which is unreactive to external stimuli, with absent posterior dominant background [
The “splitting” of an EEG waveform into its various frequency subbands maybe best performed using discrete wavelet transform compared with the customarily used frequency domain approach like fast Fourier transform. Our study concludes the relevance of wavelet decomposition in EEG analysis where time localisation of frequency components of the signal is possible. After applying LPF-TVD filtering, the EEG subbands were extracted using DWT and their energies were calculated. Statistical tests conducted revealed significant difference in delta, alpha, and beta between encephalopathy and normal EEG. Implementation of SVM classifier gave higher performance parameters for classifying the two groups when delta alone or alpha alone were taken as the features. The results correlate with the explanation of loss of normal alpha rhythm and prominence of delta rhythm during encephalopathy. This work can be extended for identifying various stages and severity of encephalopathy. This study provides a complete framework for the automated diagnosis of encephalopathy based on subband energies of EEG.
As per the ethical committee guidelines of the institution, where the study was conducted, it is not permitted to share the EEG data and patients’ confidential information in a public repository.
The authors declare that there are no conflicts of interest regarding the publication of this paper.
The authors acknowledge the authorities of Trivandrum Medical College for giving the permission for conducting the study.