Classification and Detection of Epileptic Seizures Based on Statistical Features and Autoregressive Model

Epilepsy is a prevalent neurological disorder affecting millions worldwide, characterized by recurrent, unpredictable seizures stemming from abnormal brain activity. A neurological condition called epilepsy is marked by frequent, spontaneous seizures brought on by atypical brain activity. This is a long-term, central nervous system disorder that can present in different ways and to different degrees. Sudden, excessive electrical discharges in the brain's neurons create seizures in epilepsy, which momentarily impair normal brain function. From minor, transient alterations in awareness or feeling to severe convulsions and loss of consciousness, these disruptions can result in a wide spectrum of symptoms. In order to better understand the complex mechanisms behind epileptic seizures, this research will examine the intricacies of neuronal hyperexcitability, abnormal synchronisation, and the interaction between hereditary and environmental factors that influence seizure start and propagation. For the management and treatment of epilepsy, early epileptic episode recognition is essential. To create effective seizure detection models, a number of various ML techniques have recently been applied. This introduction focuses on the use of MTE(1) [1], SE(3) [2], and AR(4) [3] model features in combination with a number of well-known classification algorithms, including the LGBM, Decision Tree, Gradient Boost, K-Nearest Neighbours (KNN), Random Forest, Logistic Regression, and XGBoost (XGB) classifiers. MTE, a component that analyses signal energy variations, provides insight into the peculiar energy EEG patterns associated with seizures. The detection of seizure activity is aided by higher entropy levels, which sample signal complexity and irregularity. The AR model [3] describes the temporal correlations and patterns in EEG signals related to seizures. Machine learning systems can accurately discriminate between seizure and non-seizure states by combining these traits. Some of the most widely used algorithms for classification: (1) LGBM: The gradient-boosting framework LightGBM is renowned for its outstanding accuracy and efficiency. To improve predictive performance, it constructs a group of decision trees with improved gradient boosting technique. (2) Decision Tree: A well-liked classifier, decision trees divide data according to feature thresholds. Decision trees can successfully identify seizures by learning from the available characteristics and building a model that resembles a tree. (3) Gradient Boost: Gradient boosting is an ensemble method that combines the decision trees and other weak learners to produce a powerful prediction model. It enhances the performance of decision trees as a whole, addressing the flaws of individual decision trees. (4) K-Nearest Neighbours (KNN) classifies data points according to how close they are to nearby points. KNN can determine seizure patterns based on their resemblance to recognised seizure instances by computing the distances between data points in the feature space. (5) Random Forest: Using many decision trees to build predictions, Random Forest is an ensemble learning technique. By combining the findings of different trees, it reduces overfitting and offers reliable classification. (6) Logistic Regression: The connection between characteristics and the binary outcome (seizure or no seizure) is modelled using logistic regression. It calculates the probability of being in each class and labels each one according to a selected threshold. (7) XGBoost (XGB) is an optimised gradient boosting system that makes use of parallel processing and tree pruning methods. By supporting weak models, it produces quick and precise predictions. These algorithms can be taught to accurately categorise incoming EEG signals by using labelled data during training, where the features derived from EEG signals during seizure and non-seizure periods are used as input. These algorithms provide a comprehensive approach to epileptic seizure identification, permitting early intervention and better patient care. They combine MTE, SE, and AR model features for classifications. It is important to note that the properties of the dataset, the necessary computations, and the desired performance all influence the choice of an algorithm. Researchers and medical professionals should evaluate and compare the effectiveness of several algorithms to choose the best strategy for their particular application. The two main research objectives of this paper are: (1) Development of a model based on the extraction of features from MTE, SE, and AR. (2) EEG signal classification and epileptic seizure detection. This EEG database was obtained by Andrzejak et al. [4] and is publicly accessible through the University of Bonn. It consists of the five datasets A, B, C, D, and E. There are 4097 samples total across each of the 100 single-channel EEG segments in each dataset, each lasting 23.6 seconds. Section I describes the introduction of the proposed method and dataset description and preprocessing. Section II is related to the work on the detection of Epileptic Seizures using various classification techniques. Section III consists of the proposed work based on the extraction of MTE(1), SE(3), and AR(4) features. Sections IV and V are based on comparative analysis and results, respectively.

1.1 Dataset description and preprocessing

The datasets consist of five sets (denoted as A-E) provided by University of Bonn, each containing 100 single-channel EEG segments of 23.6 seconds. The details of each set are as follows:

Set A (marked as O): EEG segments recorded from five healthy volunteers with eyes open.

Set B (marked as Z): EEG segments recorded from the same healthy volunteers with eyes closed.

Set C (marked as N): EEG segments recorded from seizure-free intervals in the hippocampal formation of patients.

Set D (marked as F): EEG segments recorded during seizure-free intervals from the epileptogenic zone.

Set E (marked as S): EEG segments recorded during seizure activity from the epileptogenic zone.

Preprocessing the EEG datasets required two steps. A band-pass filtering (0.5-50Hz) was first used to get rid of high-frequency noise and low-frequency drifts. After that, automated methods were employed to eliminate artefacts such as muscular contractions and eye blinks.

The proposed work is divided into three subsections, as Figure 1 illustrates. (i) Energy, randomness, and pattern recognition are employed to recover features; (ii) significant features are chosen; and (iii) features are used to identify epileptic episodes. The extracted features are classified into general categories such as MTE, SE, and AR models, The Block diagram of the proposed work includes:

EEG Samples Block: EEG Samples are extracted from 500 EEG signals i.e., Z, O, N, F, S.

Extraction Block: Extraction of EEG features related to Energy, randomness, and pattern recognition includes: MTE, SE, and AR models.

Classification Block: Classification Block to identify epileptic episodes from EEG samples based on the characteristics of extracted features.

1.png

Figure 1. Proposed work

3.1 Feature extraction and selection

In this section, MTE, SE, and AR models are used to extract features from EEG signals. AR models examine temporal dependencies, SE quantifies signal complexity, and MTE records abrupt changes in signal energy. This step is required to describe various elements of EEG dynamics connected to seizure activity. After that PCA for feature selection is used to minimize dimensionality and preserve the most informative features. Furthermore, cross-validation and statistical analysis aid in determining the most discriminative characteristics needed to create a seizure detection model.

3.2 MTE

A feature for signal analysis is the MTE, a measure of signal energy that catches sudden shifts. The following mathematical formula can be used to get the MTE of a signal ×1(n) at time index n:

$M T E=(1 / N) \times \sum_{n=1}^{N-1}\left(x 1(n)^2-x 1(n-1) \times 1(n+1)\right)$     (1)

N represents the length of the signal ×1(n). A single value was calculated for each channel from a total of 100 channels. The higher value of MTE represents Epileptic Seizure.

3.3 SE

The conditional likelihood that two sequences that are similar at point 'm' will stay similar at point 'N' is represented by the negative logarithm SampEn(m,r,N), where self-matches are not included in the probability measure.

$\operatorname{SampEn}(m, r, N)=-\ln \left(\frac{B^m(\boldsymbol{r})}{A^m(\boldsymbol{r})}\right)$     (2)

N denotes data point in a time series EEG data as x1(n)=x1(1), x1(2), x1(3)...x1(N), and B^m(r) reflects the likelihood that two sequences will match for m+1 points, whereas A^m(r) represents the likelihood that two sequences will match for m points. Take the m vector defined as:

$x_m 1(i)=[x l(i) . x l(i+1) \ldots \ldots \ldots x l(i+m-1)]$, for $1 \leq i \leq N-m+1$     (3)

Beginning with the i^th sample, these vectors reflect m subsequent ×1 values. Three variables affect SE: (1) the pattern length parameter (m); (2) the threshold parameter (r); and (3) the number of sampling points (N; N=500).

Steps for the calculation of SE

In the proposed work the value of SE is determined in the following steps as shown in Figure 2:

(1) Taking the number of sampling points (N=500).

(2) Initialize the value of m from 6 to 7 and i from 1 to N-m+1.

(3) Calculate Conditional Probability $\mathrm{B}^{\mathrm{m}}(\mathrm{r})$ and $\mathrm{A}^{\mathrm{m}}(\mathrm{r})$ for m=7 and m=6 respectively according to the value of threshold parameter(r).

(4) Calculate SE [2] as:

$\operatorname{SampEn}(6, r, 500)=-\ln \left(\frac{B^m(r)}{A^m(r)}\right)$     (4)

2.png

Figure 2. Steps for calculation of SE

3.4 AR model

The procedure for estimating the AR parameters of a set of N data points using Burg's technique is as follows:

(1) Initialize:

a. Select a desirable value for the AR model's p order.

b. To hold the AR parameters, a, make an array of length p+1.

(2) Calculate the autocorrelation coefficients:

a. Determine the data sequence's autocorrelation coefficients, or r, for lags 0 to p.

b. The variance of the data sequence is the sole component of the autocorrelation coefficient at lag 0.

(3) Initialize the forward and backward prediction errors:

a. Make two arrays, E and A, each of length p+1, to hold the forward and backward prediction errors, respectively.

b. Assign the values E(0) and A(0) to r(0).

(4) Iterate using Burg's algorithm:

a. For m=1 to p, in order: Utilizing the following equation, determine the reflection coefficient, k(m) as Eq. (5):

$k(m)=\left(r(m)-\sum_{i=1}^{m-1}(\boldsymbol{a}(\boldsymbol{i}) \times \boldsymbol{r}(\boldsymbol{m}-\boldsymbol{i}))\right) / \boldsymbol{E}(\boldsymbol{m}-\mathbf{1})$     (5)

b. Using the reflection coefficient, update the AR parameters as (6):

$a(m)=k(m)-\sum_{i=1}^{\boldsymbol{m}-\mathbf{1}}(\boldsymbol{a}(\boldsymbol{i}) * \boldsymbol{k}(\boldsymbol{m}-\boldsymbol{i}))$     (6)

c. The forward and backward prediction errors should be updated as Eq. (7) and Eq. (8):

$E(m)=E(m-1)^*\left(1-\boldsymbol{k}(\boldsymbol{m})^2\right)$     (7)

$A(m)=A(m-1)^*\left(1-\boldsymbol{k}(\boldsymbol{m})^2\right)$     (8)

d. The estimated coefficients of the AR model are represented by the resulting AR parameters, a(1), a(2), ... , and a(p).

The AR parameters and prediction errors are updated together with the reflection coefficients iteratively by Burg's algorithm until convergence or the desired degree of accuracy is reached. Depending on the specific application or implementation, the initiation and convergence criteria may change.

Table 1 lists a total of eight features. For seizure detection, MTE, SE (SE1, SE2, and SE3), and the AR model (AR1, AR2, AR3, and AR4) are utilized as inputs for the classifiers NB, DT, RF, Extra Tree, KNN, Logistic Regression, Ada-Boost, Gradient-Boost, XGB, and LGBM. According to the performance criteria presented in Table 2, significant features might be chosen. As seen in Figure 3, it is depicted as a heatmap. Table 3 lists the amount of EEG sample properties and some of its fundamental characteristics.

Table 1. Number of features

Table 2. Important features

3.png

Figure 3. Number of parameters

Table 3. Features and its basic characteristics