Detection of epileptic seizures on the basis of Electroencephalogram (EEG) recordings is a challenging task due to the complex, non-stationary and non-linear nature of these biomedical signals. In the existing literature, a number of automatic epileptic seizure detection methods have been proposed that extract useful features from EEG segments and classify them using machine learning algorithms. Some characterizing features of epileptic and non-epileptic EEG signals overlap; therefore, it requires that analysis of signals must be performed from diverse perspectives. Few studies analyzed these signals in diverse domains to identify distinguishing characteristics of epileptic EEG signals. To pose the challenge mentioned above, in this paper, a fuzzy-based epileptic seizure detection model is proposed that incorporates a novel feature extraction and selection method along with fuzzy classifiers. The proposed work extracts pattern features along with time-domain, frequency-domain, and non-linear analysis of signals. It applies a feature selection strategy on extracted features to get more discriminating features that build fuzzy machine learning classifiers for the detection of epileptic seizures. The empirical evaluation of the proposed model was conducted on the benchmark Bonn EEG dataset. It shows significant accuracy of 98% to 100% for normal
Epilepsy is a serious chronic neurological disorder affecting over 50 million people of all ages around the globe [
The automatic epileptic seizure detection system is especially useful in countries where patients’ ratio is much greater than neurologists. Moreover, it can also act as an assistant in clinical practices and helpful in ambulatory settings to investigate long EEG recordings of patients. In recent literature [
The motivation behind the proposed research is many-fold because the existing methods confront several challenges that need to be tackled. First, EEG signals are highly non-stationary and non-linear in nature due to which signal characteristics vary over different seizure events within the same patient or in between two patients [
To address the aforementioned challenges, this paper presents a Fuzzy-based Epileptic Seizure Detection (FESD) framework. In this model, an improved feature extraction scheme is introduced that extracts a combination of temporal, spectral, non-linear, and pattern features from the sub-bands of EEG signals to identify the characterizing features of epileptic signals. In order to obtain the most significant and discriminating features among extracted feature vector, the FESD model introduces a feature selection strategy that first ranks and then selects top-ranked features. In this model, fuzzy logic-based machine learning algorithms are employed to deal with the issue of class overlapping. These algorithms build fuzzy classification models based on the selected features to detect epileptic seizure recordings.
The major contributions of this research work are mentioned below.
• A feature extraction method is proposed to improve the feature vector that extracts statistical features by temporal, spectral, non-linear, and pattern analysis of decomposed EEG signal sub-bands.
• In case of pattern features extraction, the One Dimensional-Local Binary Pattern (1D-LBP) based algorithm is proposed to know the morphological structure of epileptic EEG signals.
• A feature selection strategy for better classification is introduced that finds out the most distinguishing features by applying Information gain and Analysis of variance (ANOVA) statistical test.
• The proposed FESD employs fuzzy logic-based machine learning algorithms for EEG classification into three classes including normal, inter-ictal, and ictal.
The remaining paper is structured as follows. Section 2 briefly reviews the related work after categorizing on the basis of techniques used as machine learning algorithms and signal analyzing tools. The proposed FESD framework is elaborated in Section 3. The empirical evaluation of the proposed framework in terms of performance metrics is demonstrated in Section 4. Section 5 performs the comparison of the FESD model with the state-of-the-art literature works. Finally, Section 6 concludes this paper by summarizing the contributions and findings along with some future directions.
Before proceeding, some notations along with descriptions are provided in
In these methods, state-of-the-art machine learning algorithms such as Support Vector Machine (SVM), Random Forest (RF), and K-Nearest Neighbor (KNN) have been employed for epileptic seizure detection. Since extraction of the most appropriate and distinguishing features from EEG signals is an important task for epileptic seizure detection, these methods are further grouped into four types based on the signal analysis used for feature extraction. The four types are wavelet transform-based methods, non-linear analysis-based methods, multiple decomposition analysis-based methods, and non-decomposition analysis-based methods.
Term | Description |
---|---|
ANN | Artificial Neural Network |
ANOVA | Analysis of variance |
ApE | Approximate Entropy |
BPNN | Back Propagation Neural Network |
BRNN | Bayesian Regulation Neural Network |
BSSCCA | Blind Source Separation of Canonical Correlation Analysis |
CD | Correlation Dimension |
CHB-MIT | Children's Hospital of Boston-Massachusetts Institute of Technology |
CNN | Convolutional Neural Network |
DT | Decision Tree |
DT-CWT | Dual Tree- Complex Wavelet Transform |
DWT | Discrete Wavelet Transform |
EEG | Electroencephalogram |
EMD | Empirical Mode Decomposition |
EMG | Electromyogram |
EOG | Electrooculogram |
EWT | Empirical Wavelet Transform |
FD | Fractal Dimension |
FLDA | Fishers Linear Discriminant Analysis |
GA | Genetic Algorithm |
HE | Hurst Exponent |
HWPT | Harmonic Wavelet Packet Transform |
IEEG | Intracranial Electroencephalogram |
IQR | Inter-Quartile Range |
KNN | K-Nearest Neighbor |
LBP | Local Binary Patterns |
LDA | Linear Discriminant Analysis |
LDAG-SVM | Layered Directed Acyclic Graph- Support Vector Machine |
LLE | Largest Lyapunov Exponent |
LMD | Local Mean Decomposition |
LS-SVM | Least Square-Support Vector Machine |
LSTM | Long Short-Term Memory |
MLP | Multilayer Perceptron |
MSPCA | Multi-scale Principal Component Analysis |
NB | Naïve Bayes |
NNE | Neural Network Ensemble |
NNge | Non-Nested Generalized Exemplars |
PCA | Principal Component Analysis |
PE | Permutation Entropy |
PMRS | Pattern Match Regularity Statistic |
PSD | Power Spectral Density |
PSO | Particle Swarm Optimization |
RBF | Radial Basis Function |
RCNN | Recurrent Convolutional Neural Network |
RF | Random Forest |
RFEC | Rising and Falling Edge Count |
ROC | Receiver Operating Characteristic |
RVM | Relevance Vector Machine |
SAD | Sum of Absolute Difference |
SELM | Sparse Extreme Learning Machine |
Wavelet transform-based methods for EEG signal analysis are employed in [
By using wavelet transformation, a signal is decomposed into sub-signals to extract statistical features from the decomposed sub-bands of different frequency ranges. In order to classify EEG recordings into two classes, normal and epileptic, the authors in [
Wang et al. [
In seizure
To understand the non-linearity of EEG signals, EMD and Local Mean Decomposition (LMD) techniques were proposed. Riaz et al. [
To capture the pattern behavior of EEG signals for epileptic seizure detection, Shanir et al. [
Previous studies employed multiple decomposition-based methods for signals analysis to capture EEG information from different perspectives. In such methods, Ghayab et al. [
Similarly, Alickovic et al. [
Arunkumar et al. [
Similarly, Gu et al. [
In such methods, neural network-based machine learning algorithms were used for signal classification. These algorithms, such as Convolutional Neural Network (CNN) or Recurrent Convolutional Neural Network (RCNN), have the end-to-end structure in which no explicit feature extraction and selection was performed. However, the layered structure of the classifier automatically extracts robust features for EEG signal classification.
The recent deep learning-based methodologies have a noteworthy impact in detecting seizure EEG segments. In this regard, Ullah [
On the other hand, Acharya et al. [
In this paper, a novel Fuzzy-based Epileptic Seizure Detection (FESD) model is proposed to investigate EEG signals for epileptic seizure detection. The FESD method is comprised of six major phases Data Collection, Signal Pre-processing, Feature Extraction, Feature Selection, Classifier Building, and Classifier Evaluation, as illustrated in
In the current study, a benchmark Bonn dataset is used for experimentation and evaluation of the proposed model. This dataset was collected by the Epileptology Department of Bonn University, Germany, and freely available for educational purposes. It consists of 500 EEG recordings acquired from ten individuals, including five healthy volunteers and five epileptic patients [
The Bonn dataset is comprised of five subsets A, B, C, D, and E, in which every subset consists of 23.6 s long 100 EEG recordings. Among these subsets, A and B contain EEG recordings of normal individuals with awaken state of eyes open and close, respectively. The subsets C and D are seizure-free EEG data collected from intracranial electrodes placed within and opposite to epileptogenic zone, respectively. In subset E, EEG recordings are ictal or seizure is a seizure or ictal EEG data obtained by placing electrodes in epileptogenic zone intra-cranially. The description of the Bonn dataset is summarized in
The collected EEG signals are contaminated with different types of artifacts that create noise in original recordings. These artifacts originate due to the movements of body limbs (EMG), eyes blinking (EOG), beating the heart muscles or Electrocardiogram (ECG), electrode movement, power lines impairment, and some environmental factors as well. The artifacts are eliminated by applying the Butterworth bandpass filter. This filter allows passing a certain range of frequencies while discards the remaining. In this work, Butterworth filter bandwidth is configured as 0.1–60 Hz. The reason behind choosing this bandwidth is that the most frequencies of EEG signals lie in this range. According to the configured bandwidth setting, the Butterworth filter allows passing only those frequency components of EEG signal that fall in the range of 0.1 to 60.0 Hz while the remaining that beyond this range are suppressed.
This paper introduces an improved method to extract such features about EEG signals that are better representative of distinct seizure EEG recordings from non-seizure ones. The method is comprised of three steps in which firstly, generate signal sub-bands by decomposition. Secondly, extract the statistical features from generated sub-bands to obtain varying trends and complexity of the signal. Thirdly, the most discriminating features are selected from extracted ones that lead to omit redundant and less important features. Three steps of the proposed feature extraction method are elaborated in the following sub-sections.
In the first step of the feature extraction method, Discrete Wavelet Transform (DWT) technique [
Subsets
Individual type
Health state
Electrode type
Electrode placement
A
Normal
Awaken with eyes open
Surface
International 10–20 system
B
Normal
Awaken with eyes closed
Surface
International 10–20 system
C
Patient
Seizure-free
Intracranial
Within epileptogenic zone
D
Patient
Seizure-free
Intracranial
Opposite to epileptogenic zone
E
Patient
Seizure
Intracranial
Within epileptogenic zone
To enhance EEG characterization, salient feature extraction is very important. In second step of feature extraction phase, four types of features temporal, spectral, non-linear, and patterns are extracted from decomposed sub-bands to achieve effective classification accuracy. These features capture different types of information about EEG sub-signals, such as temporal features collect time domain information and spectral features give statistical information about the frequency components. Similarly, non-linear features reflect the signal complexity and irregularity, and pattern features report the signal trends and their variations.
Temporal Features
These features extract time domain statistical features of EEG signals such as amplitude minimum, maximum, median, mean, mode, and standard deviation. In the current study, minimum amplitude [
Spectral Features
One fundamental characteristic of signals is the frequency that informs about the rhythms of a signal. The spectral features encircle information about frequency components that constitute the signal and other fractal features such as signal power and energy. EEG signals have transitory characteristics and non-stationary nature. Thus signal analysis becomes less significant by only considering the time-domain features; therefore, spectral features are extracted as well [
c) Non-linear Features
Since EEG signals are non-linear by nature, therefore capturing the EEG signals’ non-linearity is significant. In the proposed FESD model, non-linear features such as approximate entropy (
Pattern Features
In order to have information about the changing trends of EEG signals, pattern features have significant importance. This paper presents an algorithm to extract pattern features by utilizing 1D-LBP technique [ One-Dimensional Local Binary Pattern
This technique is used to get pattern information from one-dimensional data such as time series. In the mathematical formulation given by [
For the current study, nine consecutive samples of time series are considered at a time, and an 8-bit binary code is obtained for each of the central samples.
Algorithm1 Description
The proposed algorithm categorizes patterns of binary codes into four groups; no-change, one-change, two-change, and multi-change based on the number of bit-transitions. No-change designates those binary patterns in which the entire binary code contains either all zeros or all ones. One-change category groups those patterns where only one transition happens, either from zero to one or one to zero. In the two-change group, the binary code pattern shows two transitions of zero and one. However, the remaining patterns with more than two binary transitions in their codes are placed in the multi-change group.
According to Algorithm1, the input consists of EEG signal time series and three lists of decimal pattern ids with the varying number of transitions. Among these lists, a0 contains ids of no-change patterns, a1 is the list of one-change pattern ids and a2 includes ids, of two-change patterns. The algorithm outputs five pattern features, including four histograms
In this feature selection phase, the most distinguishing features are chosen from the feature vector obtained by the proposed feature extraction method. Two major purposes of this step are the removal of redundant as well as low ranked features and selection of discriminating features. The extracted feature vector contains 14 features, including three temporal, three spectral, three non-linear, and five pattern features. In order to shortlist the extracted feature vector and select the most distinctive ones, a feature selection strategy is employed.
The proposed feature selection strategy includes InfoGain computation and application of ANOVA statistical test. Under this strategy, first InfoGain ranks the extracted features with respect to their usefulness and distinguishing ability. InfoGain is calculated by using
The top-ranked features are selected for ANOVA test that further purify the selection by choosing the most significant features among top-ranked features.
In the classifier building phase, the classification of EEG signals into three different classes, namely normal, inter-ictal, and ictal classes not only differentiate normal from epileptics but also seizure (ictal) and seizure-free (inter-ictal) states of epilepsy. In real-life classification cases, fuzzy logic-based algorithms are very useful where classes are overlapping. One of these scenarios is the classification process of EEG signals in which the algorithms must be capable of understanding the human brain EEG signals based on their characterizing features. For example, in EEG recordings, seizure signals have some features similar to non-seizure EEG segments and vice versa [
It is an extension of Repeated Incremental Pruning to Produce Error Reduction (RIPPER) [
In
In this classifier, the similarity of
In
In this algorithm, FNN is combined with fuzzy rough approximations [
VQNN is derived from making variations in FRNN algorithm [
It is a classifier for inducing a rule-based inference engine from data based on the fuzzy lattice framework [
In this phase, classifier evaluation is performed by 70 and 30 percent split of input EEG data for training and testing purposes, respectively. The training data is used to train the fuzzy logic-based five fuzzy classifiers, namely, FURIA, FLR, FNN, FRNN, and VQNN. These classifiers build their classification models on the basis of selected feature vectors and classify test data into normal, inter-ictal, and ictal classes. The built models are evaluated by means of three performance metrics: accuracy, sensitivity, and specificity. These performance measures are defined in the following formulas.
In above
This section summarizes classification results obtained by using the Bonn dataset [
In this study, the Bonn dataset, including five subsets of EEG signals, is used to perform experiments. The Python programming language, along with MySQL database, is used for feature extraction. Moreover, fuzzy Waikato Environment for Knowledge Analysis (WEKA) toolbox that provides a collection of fuzzy machine learning algorithms along with traditional classification algorithms is utilized not only for classifier building and evaluation but for feature selection as well. In this paper, fuzzy logic-based five fuzzy classifiers, FNN, FRNN, VQNN, FURIA, and FLR, are used to evaluate ten classification cases in terms of accuracy, sensitivity, and specificity performance measures.
In order to evaluate the proposed FESD model, ten classification cases are performed to distinguish seizure EEG signals from non-seizure ones. These classification cases are A-E, B-E, AB-E, C-E, D-E, CD-E, ABCD-E, AB-CD-E, A-D-E, and A-B-C-D-E. Among two-class classification cases, A-E, B-E, and AB-E refer to the normal
This study is working on seven such cases in which two-class classification is discussed. These classification cases are further categorized into the following three subtypes.
Normal
This type of two-class classification includes A-E, B-E, and AB-E classification cases. Inter-ictal
Classifier
A-E
B-E
Accuracy (%)
Sensitivity (%)
Specificity (%)
Accuracy (%)
Sensitivity (%)
Specificity (%)
FURIA
100
100
100
96.67
100
93.33
FNN
96.67
93.33
100
91.67
93.33
90.00
FRNN
100
100
100
100
100
100
VQNN
100
100
100
100
100
100
FLR
100
100
100
96.67
100
93.33
In C-E, D-E, and CD-E classification cases, EEG segments are classified into inter-ictal and seizure classes. The results of these cases are outlined in
Classifier | C-E | D-E | ||||
---|---|---|---|---|---|---|
Accuracy (%) | Sensitivity (%) | Specificity (%) | Accuracy (%) | Sensitivity (%) | Specificity (%) | |
FURIA | 96.67 | 100 | 93.33 | 100 | 100 | 100 |
FNN | 91.67 | 90.00 | 93.33 | 85.00 | 83.33 | 86.67 |
FRNN | 98.33 | 96.67 | 100 | 98.33 | 96.67 | 100 |
VQNN | 98.33 | 96.67 | 100 | 100 | 100 | 100 |
FLR | 96.67 | 100 | 93.33 | 98.33 | 100 | 96.67 |
Classifier | AB-E | CD-E | ||||
---|---|---|---|---|---|---|
Accuracy (%) | Sensitivity (%) | Specificity (%) | Accuracy (%) | Sensitivity (%) | Specificity (%) | |
FURIA | 98.89 | 97.06 | 100 | 94.44 | 100 | 91.07 |
FNN | 90.00 | 76.47 | 98.21 | 91.11 | 91.18 | 91.07 |
FRNN | 100 | 100 | 100 | 98.89 | 97.06 | 100 |
VQNN | 100 | 100 | 100 | 98.89 | 97.06 | 100 |
FLR | 100 | 100 | 100 | 93.33 | 94.12 | 92.86 |
Non-seizure
Only one classification case ABCD-E is concerned with seizure
Classifier | ABCD-E | ||
---|---|---|---|
Accuracy (%) | Sensitivity (%) | Specificity (%) | |
FURIA | 98.67 | 93.75 | 100 |
FNN | 90.67 | 68.75 | 96.61 |
FRNN | 99.33 | 100 | 99.15 |
VQNN | 98.67 | 100 | 98.30 |
FLR | 97.33 | 93.75 | 98.30 |
The classification of EEG into three classes is a difficult task as compared to two-class classification cases. In this study, A-D-E and AB-CD-E are three-class classification cases. These cases refer to classification between normal, inter-ictal, and seizure classes.
Classifier | A-D-E | |||
---|---|---|---|---|
Accuracy (%) | Sensitivity (%) Inter-ictal | Sensitivity (%) Ictal | Specificity (%) | |
FURIA | 96.67 | 88.46 | 100 | 100 |
FNN | 76.67 | 57.69 | 91.17 | 76.67 |
FRNN | 97.78 | 100 | 97.05 | 96.67 |
VQNN | 97.78 | 96.15 | 97.05 | 100 |
FLR | 84.44 | 92.31 | 97.05 | 63.33 |
Classifier | AB-CD-E | |||
---|---|---|---|---|
Accuracy (%) | Sensitivity (%) |
Sensitivity (%) |
Specificity (%) | |
FURIA | 93.33 | 88.89 | 90.62 | 100 |
FNN | 65.33 | 57.14 | 68.75 | 72.72 |
FRNN | 94.67 | 93.65 | 96.87 | 94.54 |
VQNN | 94.00 | 88.89 | 96.87 | 98.18 |
FLR | 89.00 | 95.24 | 93.75 | 80.00 |
In this case, EEG signals are classified into five classes A, B, C, D, and E. Thus, empirical evaluation of A-B-C-D-E is presented in
Classifier | A-B-C-D-E | |
---|---|---|
Accuracy (%) | Sensitivity (%) Ictal | |
FURIA | 77.33 | 90.62 |
FNN | 52.00 | 68.75 |
FRNN | 82.00 | 96.87 |
VQNN | 80.67 | 100 |
FLR | 73.33 | 96.87 |
This section analyzes the effectiveness of different types of features in the proposed feature vector. In order to perform this analysis, experiments were conducted with temporal, spectral, non-linear, and pattern features separately for ten classification cases. The comparison of experimental results for different feature types using the FRNN classifier is shown in
This section analyzes the effectiveness of different types of features in the proposed feature vector. In order to perform this analysis, experiments were conducted with temporal, spectral, non-linear, and pattern features separately for ten classification cases. The comparison of experimental results for different feature types using the FRNN classifier is shown in
This section analyzes the significance of the proposed feature selection strategy by conducting experiments with and without it.
This section is dedicated to comparing the proposed FESD methodology with recent literature works of epileptic seizure detection using the Bonn dataset.
In prior studies [
The comparative analysis in
The current work analyzed the proposed FESD model in terms of ten classification cases of the Bonn dataset. Each classification case was implemented with five fuzzy classifiers individually using 70/30 split ratio of the data. The empirical results demonstrated that FRNN and VQNN showed remarkable performance; however, FRNN proved the best. The proposed feature extraction and feature selection strategies were also analyzed experimentally with FRNN to be the best ones. To analyze feature extraction method, the proposed model was implemented with different feature types as well as proposed feature vectors separately. The percent accuracies depicted the superiority of the proposed feature vector as compared to the different feature types. In order to analyze the effectiveness of the proposed feature selection strategy on classification accuracy, experiments were conducted with and without feature selection. The obtained results confirmed the importance of feature selection strategy in order to achieve better classification accuracy for epileptic seizure detection.
[Ref.] | Method | Classification case | Accuracy (%) |
---|---|---|---|
[1] | DWT + PSO-SVM | A-E | 99.38 |
[2] | DWT + ICFS + RF | A-E | 98.45 |
[4] | DWT + Hilbert Envelope + KNN | A-E | 99.70 |
[7] | TQWT + Kraskov Entropy + LS-SVM | CD-E | 97.75 |
EMD + Hilbert Transform + SVM | A-E |
99.00 |
|
[21] | (DWT+FFT+EMD) + (PCA+ANOVA) + SVM | ABCD-E | 99.25 |
3 P-1D-CNN | B-E |
99.6 |
|
[27] | CNN | B-D-E | 88.67 |
DWT + Genetic Algorithm +ANN | C-E |
95.00 |
|
DWT + InfoGain + FRNN | A-E |
100 |
In addition, the proposed epileptic seizure detection framework was also evaluated with 10-fold cross validation to compare recently introduced literature in terms of percent accuracy for the Bonn dataset. This comparative analysis proved the outstanding performance of the proposed FESD model for epileptic seizure detection.
Epilepsy is a common mental disorder that disrupts the normal mental activity. It not only negatively affects the brain functionality but also physically disturbs the patient. In this research work, we have studied and critically evaluated the machine learning-based epileptic seizure detection methods in order to point out existing challenges and limitations. After a critical analysis, it has been concluded that existing approaches have flaws to cater with the unobvious behavior of EEG signals due to its non-stationary nature. So, it is needed to work on an automatic system for feature extraction, selection, and classification tasks to better capture the morphology of non-linear and non-stationary EEG signals. For this purpose, the FESD model was proposed that introduced an improved features extraction method and a feature selection strategy for building classification models in fuzzy classifiers. In comparison to the existing techniques, the proposed model achieved better accuracy results for ten classification cases along with FRNN and VQNN fuzzy classifiers. The results provided 100% accuracy for normal
In this work, we have focused on the classification of individuals into normal, inter-ictal, and ictal classes but did not consider pre-ictal and post-ictal states of the patient. The proposed approach can be improved by taking into account the pre-ictal and post-ictal regions in EEG recordings. Furthermore, the proposed model was evaluated on the small single-channel dataset, which in future can be evaluated on a large-scale real-world dataset. In addition, one of the future directions can be to introduce a framework by which normal individuals and epileptic patients could be further categorized into different severity levels of disorder for their better medical treatment.