Seven Epileptic Seizure Type Classification in Pre-Ictal, Ictal and Inter-Ictal Stages Using Machine Learning Techniques

doi:10.21203/rs.3.rs-2100432/v1

Download PDF

Research Article

Seven Epileptic Seizure Type Classification in Pre-Ictal, Ictal and Inter-Ictal Stages Using Machine Learning Techniques

https://doi.org/10.21203/rs.3.rs-2100432/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Based on the symptoms experienced during the episode and the Electroencephalograph (EEG) recording made during the inter-ictal phase, the doctor makes the epileptic seizure type diagnosis. The fundamental issue, however, is that patients frequently struggle to explain their symptoms in the absence of an observer and identify traces in inter-ictal EEG patterns.

Aims

This study examines electroencephalographic (EEG) signals from epileptic seizures in order to diagnose seizures in pre-ictal, ictal, and inter-ictal stages and to categorize them into seven groups.

Methods

For the investigation, a licensed dataset from Temple University Hospital was used. Seven different seizure types are pre-processed from the seizure corpus and divided into pre-ictal, ictal, and inter-ictal stages. K-Nearest Neighbor (KNN), Random Forest, and other machine and deep learning techniques were used to perform the multi-class categorization.

Result

With 20 channels and an 80 − 20 train-test ratio, multiclass classification of seven different types of epileptic seizures was accomplished. For the pre-ictal, ictal, and inter-ictal stages, weighted KNN achieved accuracy levels of 94.7%, 94.7%, 69.0% during training and 94.46%, 94.46%, and 71.11% during testing.

Conclusion

Seven epileptic seizure type classification using machine learning techniques carried out with MATLAB software and weighted KNN shows better accuracy comparatively.

Seizure classification

TUHEEG

ABSZ

CPSZ

FNSZ

GNSZ

SPSZ

TNSZ

TCSZ

Random Forest

KNN

EEG

Epilepsy is a brain disease characterized by repetitive seizures caused by erratic electrical discharge in the brain. Epilepsy is a chronic disease caused by excess electrical discharge in the brain causing involuntary behavioral changes such as loss of consciousness and convolutions. In low and middle-income countries, nearly 80% of epileptic patients live with three-fourths of these people facing either a treatment gap or a shortage of anti-seizure medicines. As such, the occurrence and frequency of epileptic events are unpredictable, making diagnosing and treating difficult. Most of the time patients visit the outpatient department (OPD) with a history [1] of epileptic episodes. Clinicians diagnose based on EEG recording taken at the time of hospital visit i.e. inter-ictal stage along with the onset symptoms observed. But, the main challenge is, the un-predictive nature of the occurrence of seizures leads most of the time patients to experience without any attendee, and because of this the onset symptoms could not be explained. At the same time, non-interactive epilepsy requires expertise to interpret inter-ictal EEG as looking the same as a normal person’s EEG and end up giving EEG reports as “normal”. This results in trial and error-based treatment. Neurologists prescribe some set of medicines predicting some type based on whatever information shared by the patient and attendee and then change the medicine in consecutive visits. As inter-ictal EEG looks the same as EEG of a non-epileptic epileptic person, also noise and various artifacts like the movement of body parts during acquisition put a challenge to even experts to diagnose from those readings. On the other hand, anti-seizure drugs (ASDs) have adverse side effects.

Seizure has four stages namely, pre-ictal, ictal, post-ictal, and inter-ictal. Pre-ictal is just before the occurrence of a after around 10 minutes of onset and lasts till the next occurrence of seizure. The pre-ictal stage usually involves dizziness, headache, and nausea and is followed by the stage of intense electrical activity in the brain called the ictal region. Then comes the post-ictal region where the patient returns to baseline conditions along with symptoms like disorientation, drowsiness, and headache. Figure 1 depicts the pre-ictal, ictal, and postictal stages of an epileptic seizure, where inter-ictal representation along with non-epileptic EEG is shown in Fig. 2.

With the advent of machine learning and Deep learning and its increasing popularity in healthcare application makes it possible to classify majorly a) seizure-seizure free and b) different types of seizure but it’s done in ictal period having high frequency EEG signal with spikes makes classification little simple. There are many works available in seizure-seizure free classification. Researchers use different pre-processing techniques for accuracy enhancement. In literature different mathematical models are used as pre-processing techniques like empirical mode decomposition (EMD) [2], ensemble EMD [3], dissimilarity based frequency distribution [4], based on high order statistical parameter tuning [5–6], feature extraction based on discrete wavelet transform (DWT) [7], composite multiscale dispersion entropy [8], etc. For the ever-growing demand for correct diagnosis, various algorithms are developed [9]. For segment wise seizure classification transfer function approach is adopted [10]. Increasing acceptability of artificial intelligence in the field of medical diagnostic, seizure classification from unprocessed EEG signals are carried out [11]. Various machine learning algorithms [12–13] like support vector machine (SVMs) [12], using auto regression feature [14], using Bayesian regularized shallow neural networks [15], bagging based ensemble framework [16], using extreme gradient based classifier [17], is used. Predominant usage of deep learning techniques [18–20] are witnessed. Based on domain invariant deep representation [21], autonomous deep feature extraction [22], convolutional neural networks (CNN) [22], recurrent neural networks (RNN) [24–25] etc. deep learning algorithms are used. To enhance accuracy one or more algorithms are clubbed together, such as CNN with long- short memory (LSTM) [26], CNN with empirical WT [27], etc.

Seizure type detection also slowly became research interest [28–29]. Different machine learning [30], deep learning methods [31–32] are explored with and without pre-processing of EEG signals. RNN [33], LSTMs, K nearest neighbor (KNN), SVMs [34] with different transforms like discrete cosine transform (DCT), Hilbert transform (HT), etc. are used as feature extractor based on their energy compaction [35], using recurrence plots [36],based on dictionary learning and sparse representations [37], using Random forest (RF) classifier [38], using different layers of CNN, transfer learning [39], using deep batch normalization [40], using neural memory networks [41], etc. are tried out with different researcher using cleaned and raw data from different datasets [42]. Slowly real time clinical diagnostics is tried out [43]. But most of the literature available classification is done in ictal stage EEG signal and very less exploration in inter ictal [44–45] or other stages [13] can be found because of its similarity with non-epileptic EEG.

It is clear from the literature that multi class seizure classification in pre-ictal, inter-ictal has not been paid much attention. Therefore, this paper focuses on epileptic seizure type classification based on pre-ictal, ictal and interictal stages using machine learning algorithms. To the best of the author’s knowledge, multi class classification for seven different types in pre-ictal, ictal and interictal is tried out for the first time.

In this paper the first section discusses introduction along with relevant supportive literatures, section 2 describes methodology and includes information about dataset, pre-processing techniques adopted and also the different machine and deep learning techniques used for classification. Section 3 deals with results and discussion followed by conclusion for the study.

The methodology followed in this work is as shown in Fig. 3

Table 1

Seizure types and description
Seizure code	Name	Description	Pre- Ictal		Ictal		Inter-ictal
Seizure code	Name	Description	Samples	NE	samples	NE	Samples	NE
^aABSZ	Absence Seizure	On EEG, absence discharges are noticed. Symptom: Loss of consciousness	271989	18	96568	56	745239	46
^aCPSZ	Complex Partial Seizure	Staring blank and daydreaming is the symptoms also loss of awareness regarding their surroundings.	317195	28	920447	52	753664	46
^bFNSZ	Focal Seizure	Starts in one part of the brain, but can spread to other parts i.e. become generalized.	426127	28	324887	55	786432	48
^bGNSZ	Generalized Non Specific Seizure	Generalized seizures which cannot be further classified. Symptoms: unconscious-ness with spasms, stiffening, etc.	323400	22	447487	54	712084	44
^aSPSZ	Simple Partial Seizure	Awareness is not affected. It affects body muscles.	298618	18	541405	52	786432	48
^aTNSZ	Tonic Seizure	Stiffening of muscles.	353294	28	1089507	52	786432	48
^aTCSZ	Tonic-Clonic Seizure	First stiffening followed by body jerking (Grand Mal)	411635	28	240279	45	700249	44
*NE = number of events

Seizure Types And Descriptions

An epileptic seizure is mainly having four major types, namely, Fibril, generalized, partial, and unclassified. They are subdivided into 20 different classes. In this work, seven types of seizures have been chosen, which have different symptoms during the onset of a seizure attack. Two manifestations are used i.e. electro clinical and electrographic. For representing Electro clinical, ‘a’ is used and for electrographic ‘b’ is used. Seizure types and description is narrated in Table 1.

Dataset

The dataset used in this work is taken from Temple University Hospital’s EEG Repository [42], specifically TUH EEG Seizure Corpus (TUSZ) as shown in Fig. 4. The repository contains 16986 sessions from 10874 unique subjects. The number of EEG channels varies from 20 to 31. The signals are of varying lengths (3,00,000 columns). The 10–20 electrode configuration was used.

Pre Processing

The obtained EEG recordings were then preprocessed be- fore feeding into machine learning models. Since each file had a varying number of channels (from 21 to 30 channels), the first 18 channels from all files were taken and labelled for the type of seizure. Duration of signal for ictal region was directly taken from the corpus information shown in Fig. 1. For inter-ictal regions, duration between the stop time of one event and start time of the next chronological event of the same file were taken as start and stop times. The duration of signal before the first event of each file is taken as the pre-ictal duration.

Feature extraction using Fast Fourier Transform (FFT)

For feature extraction of each signal, FFT as in Eq. (1) was performed on each signal and the power spectrum was analyzed. The frequencies in the signal with maximum power were extracted as features and appropriate labels (type of seizure each for ictal, preictal, and interictal) were added.

X(K)= $\sum _{n=0}^{N-1}x\left(n\right){e}^{\frac{2\pi Kn}{N}}$ (1)

The features extracted above were fed into multiple classification algorithms to find the models giving better results. Of these, the models showing better performance can then be tuned to improve performance. These models are explained as below.

Machine Learning Algorithms

A. Bagged Trees

To reduce variance of a statistical learning method, bagging or bootstrap aggregation, is a procedure in general. In this bagged tree (BT) method, using B bootstrapped training sets, B regression trees are constructed and the resulting predictions are averaged. Due to deep growth of trees and non-pruned nature, high variance with low bias can be observed in each individual tree, Towards reducing variance averaging the BT is used. The predicted value for an observation is the mode (classification) or mean (regression) of the trees. The difference between BT and Decision Trees(DT) is that BT may contain multiple trees, each having different features, terminal node counts, data, etc., where DT contains a single tree. Bagged Trees consist of the following steps.

Sample m data sets D₁, D_2, …., D_m from training set D with replacement.
For each D_j train a classifier h_j(x) where x represents features.
The final classifier is given by Eq. (2)

$$\text{h}\left(\text{x}\right) = \frac{1}{m} \sum _{j=1}^{m}{h}_{j}\left(x\right)$$

The advantages of Bagged Trees are:

It can reduce the variance within the algorithm (helpful with high-dimensional data).
It is easy to implement.
It forms a strong learner yielding better performance. Disadvantages of Bagged Trees are:
The algorithm is computationally intensive.
It is difficult to interpret (Low explaining ability).

B. K-Nearest Neighbors

The KNN algorithm works on the assumption that similar data points tend to be situated closer to each other in a plane, as seen in Fig. 5. This classification algorithm works by classifying data points based on their ‘K’ nearest neighbors. For example, if the value of K is taken to be 1, the data will be assigned the class of its closest neighbor. In Fig. 5, the blue hexagon is the test sample, two concentrated circles represent K = 3 and K = 5 nearest neighbors. In this example, K = 3 has 2 red balls close to the test sample whereas green only has one ball so, in this case, the test sample is classified as red. On the other hand, for K = 5, 3 green balls are present compared to 2 red balls. Therefore, for K = 5, the test sample will be classified as a green class. Because of this ambiguity, researchers prefer to have K = 1, most of the time showing better results, compared to any value of K. the nearest neighbors are found based on Euclidean distance as in Eq. (4).

The largest probability of an input x being assigned to a class is given as in Eq. (3):

$$P\left(y=j|X=x\right)=\frac{1}{K}{\sum }_{i\in A}^{}I\left({y}^{\left(i\right)}=j\right)$$

The distance metric is given by the Euclidean distance as follows (where p = 2):

$$d\left(x,{x}^{{\prime }}\right)=\sqrt{{\left({x}_{1}-{y}_{1}\right)}^{p}+{\left({x}_{2}-{y}_{2}\right)}^{p}+\dots +{\left({x}_{N}-{y}_{N}\right)}^{p}}$$

Commonly used distance metrics are the Euclidean, Manhattan and Minkowski metrics. d(p, q) is the distance between points p and q as shown in Eq. (5), Eq. (6) and Eq. (7) respectively. Here n is the number of dimensions of the train set and pi and qi are the values of the ith dimensions of data points p and q respectively.

Euclidean distance metric

$$d\left(p,q\right)= \sqrt{\sum _{i}^{n}{\left({p}_{i}- {q}_{i}\right)}^{2}}$$

Manhattan distance metric

$$d\left(p,q\right)= \sum _{i}^{n}|{p}_{i}- {q}_{i}|$$

Minkowski distance metric

$$d\left(p,q\right)= {\left({\sum }_{i}^{n}|{x}_{i}-{y}_{i}|\right)}^{1/p}$$

The merits of KNN are:

Implementation is simple.
To the noisy training data, is robust
For large training data, it can be more effective.

The demerits of the KNN algorithm are:

Determination of the value of K is complex sometimes.
High computation cost because for all the training samples the distance between the data points needs to be calculated.
Standardization is required for training sets.

C. Weighted KNN

The weighted K-Nearest Neighbors (WKNN) algorithm differs from the conventional KNN algorithm (as in Fig. 6) as weights are assigned to classes, and the decision is reached depending on the weights. The data points located close to the query point are assigned higher weights. The results obtained from this model are easy to interpret. The weighted Euclidean distance d between points p_i and q_i with weight w_i is calculated as in Eq. (8).

$$d\left(p,q\right)= \sqrt{{\sum }_{i}{w}_{i}{({p}_{i}-q)}^{2}}$$

The advantages of the weighted KNN algorithm are:

Gives importance to closer neighbors
No training period (Instance-based learning)
It is easy to implement

Disadvantages of weighted KNN are:

Sensitive to noise, missing data, and outliers
Does not work well for high-dimensional data
Needs feature scaling

D. Subspace KNN

In the random subspace method, the features (‘attributes’, ‘independent variables’, ‘predictors’) are sampled randomly unlike BT. In this, it replaces each learner. Informally, to avoid over-focusing on features by individual learners, which appear highly predictive/descriptive in the training set. But, at the same time, it fails to be as predictive for points outside that set. For this reason, for high-dimensional problems where the number of features is much larger than the number of training points, random subspaces are an attractive choice. The selection of subspace is mathematically shown in Eq. (9) and Eq. (10). The process is pictorially represented by Fig. 7.

Given a set of N points in an n-dimensional feature space {(${x}_{1},{x}_{2},\dots ,{x}_{n}$)| ${x}_{i}$ is real for all $1\le i\le n$} (9)
We consider the m-dimensional subspaces

{(${x}_{1},{x}_{2},\dots ,{x}_{n}$)|${x}_{i}=1$ for $\text{i}\in I,{x}_{i}=0 \text{f}\text{o}\text{r} \text{i}\notin I$} (10)

Where I, is an m-element subset of {1, 2… n} and m < n.

The advantages of Subspace KNN (SKNN) are:

Gives different results for each subspace
Avoids overfitting issues. better performance.

Disadvantages of Subspace KNN are:

Not suitable when data contains irrelevant features
Difficult to interpret.

E. Random Forest

A Random Forest (RF) algorithm can be used for classification as well as regression, as in the case of Support Vector Machines. The algorithm works by creating decision trees and assigning new data points to the category that wins the majority of votes, as described in Fig. 8. This model finds applications in many sectors and is often used for high-dimensional data.

The pre-processed signal passed through the classifier app for multi class classification without hyper parameter tuning. Seven class classification accuracy in percentage for pre-ictal stage is shown in Fig. 9, for ictal stage is shown in Fig. 10 and inter-ictal stage is shown in Fig. 11. The models selected based on accuracy more than 60%. The models selected are SVM Kernal, Cosine KNN (CO KNN) Cubic KNN (Cu KNN), Medium KNN (MKNN), Subspace KNN (SKNN), BT, WKNN, Fine KNN (FKNN).

This experiment used 80 − 20 train-test ratio. The maximum test accuracy obtained 77.4%, 81.8%, 68.7% for pre-ictal, ictal and inter-ictal stages respectively by BT model.

The train and test accuracy in percentage of all the three stages is shown in Fig. 12 and Fig. 13 respectively.

It can be clearly observed that Ictal shows better accuracy in both training and testing. Pre-ictal some of the models shows at par ictal accuracy where inter-ictal without hyper parameter tuning shows comparatively low accuracy.

RF is modeled and all other models are fine-tuned using hyper parameter aiming better accuracy for all the three stages. Table 2, Table 3 and Table 4 shown the fine-tuned results of train and test accuracy for pre-ictal, ictal and inter-ictal stages respectively.

Table 2

Seven Class Pre-ictal Classification
Model	Train Accuracy	Test Accuracy
WKNN	94.70	94.46
RF	75	76.8

Table 3

Seven Class ictal Classification
Model	Train Accuracy	Test Accuracy
WKNN	94.70	94.46
KNN	80.6	83.2
RF	75	76.8

Table 4

Seven Class Inter-ictal Classification
Model	Train Accuracy	Test Accuracy
WKNN	69.0	71.11
Ensemble BT	69.0	69.2
FKNN	65.1	65.2

The best-performing models were further fine-tuned for robust performance. Grid search, Random Search, and Bayesian Optimization techniques were utilized to find optimal hyperparameters. Weighted KNN, Bagged Trees, Random Forest, and Fine KNN were found to be the top models which improve accuracy on hyperparameter tuning. The hyper-parameters for these are outlined below

Weighted KNN: The Number of neighbors K, distance metric, and weight function (equal, inverse or squared inverse) parameters of WKNN are varied.
Fine KNN: In Fine KNN, the number of neighbors K is fixed to 1 while the other parameters are varied. Since Fine KNNs are not distance weighted, the only parameter left to vary is the distance metric.
Bagged Trees: Hyperparameters for bagged trees include minimum leaf size (minimum number of observations per leaf), number of trees in the ensemble, and whether or not the trees need pruning.
Random Forest: Maximum depth of trees in the forest, a minimum number of observations required in any node to split it, number of trees, maximum number of features provided, and fraction of dataset to be used in training, and some parameters that require tuning.

With hypermeter tuning the test accuracy is enhanced from 77.4–94.46% in the case of pre-ictal, in the case of ictal it improves from 81.8–94.46%, and in the inter-ictal stage accuracy increases from 68.7–71.11%. Though without fine-tuning BT performed best after hyperparameter tuning WKNN showed the best accuracy for all three stages.

The EEG signals of epileptic seizures are collected from the TUHEEG corpus [42]. With seven different types of seizure signals along with individual patient histories, a new project-based database is created. In this, all EEG signals consider the same number of channels. In the next step pre-ictal, ictal and interictal stages are extracted and made into three different sub-datasets, each of which has seven different classes. Seven class classifications of each of the stages are performed using a machine learning model and RF is modeled.

To improve the accuracy of models, hyperparameter tuning is performed. After fine-tuning WKNN gives the best result for all three stages. Accuracy for both pre-ictal and ictal shows 94.7% and 94.46% for train and test respectively. Where inter-ictal train and test accuracy were obtained as 69.0% and 71.1% respectively.

This experimental study can give promising aid to clinical diagnosis of epileptic seizure type and most of the patients come in the inter-ictal stage. The pre-ictal prediction of seizure can help the patients and clinicians to prevent the episode.

Ethics approval The study was carried out with an open source licensed database, so Ethical approval is not applicable.

Consent to participate Not applicable as a stored dataset is used.

Consent for publication No information which can identify the subjects is included in this article.

Conflict of interest None of the authors has any conﬂict of interest to disclose

Future scope

An epileptic seizure is a common neurological disorder and its occurrence is highly un-predictive. Therefore, most of the patients visit clinicians in an inter-ictal stage with vague details about the episode, making diagnosis extremely challenging. Proper study and research in inter-ictal definitely enhance the accuracy of diagnosis in clinical setup seizures.

This paper concentrates on seven different types of seizures, in the future researchers can explore more types of seizures incorporating the symptoms of each type.

Gao H, Du X, Li W, Liu W, Zhou, Dong, Seizure description accuracy and seizure classification results from patient histories, Epilepsy Research (2020) Vo. 167, DOI: 10.1016/j.eplepsyres.2020.106413.
Karabiber Cura OK, Atli S, Türe. Hatice Sabiha Akan, Aydin, Epileptic seizure classifications using empirical mode decomposition and its derivative, Biomed Eng Online (2020), 19, 1, DOI:10.1186/s12938-020-0754-y.
Dutta KK, Arokia KV,S, Swamy. Removal of Muscle Artifacts from EEG Based on Ensemble Empirical Mode Decomposition and classification of Seizure using Machine Learning Techniques, IEEE- International Conference on Inventive Computing and Informatics (ICICI 2017), DOI:10.1109/ICICI.2017.8365259.
Ech-Choudany Y, Scida D, Assarar M, Landré J, Bellach B, Morain-Nicolier F. Dissimilarity-based time–frequency distributions as features for epileptic EEG signal classification, Biomed Signal Process Control (2021), 64, DOI:10.1016/j.bspc.2020.102268.
Sharma R, Pachori RB, Sircar. Pradip, Seizures classification based on higher order statistics and deep neural network, Biomed Signal Process Control (2020), 59, DOI:10.1016/j.bspc.2020.101921.
Zhang S, Liu G, Cui XR, Cai W, Jing Hu, Qi XSun,YQiu,J, Yuan. A combination of statistical parameters for epileptic seizure detection and classification using VMD and NLTWSVM, Biocybernetics and Biomedical Engineering (2022), 42, 1, DOI:10.1016/j.bbe.2022.02.004.
Zarei A, Asl B, Mohammadzadeh. Automatic seizure detection using orthogonal matching pursuit, discrete wavelet transform, and entropy based features of EEG signals, Comput Biol Med (2021), 131, DOI:10.1016/j.compbiomed.2021.104250.
Sukriti Chakraborty M, Mitra. Debjani, Automated detection of epileptic seizures using multiscale and refined composite multiscale dispersion entropy, Chaos Solitons and Fractals (2021) 146, DOI:10.1016/j.chaos.2021.110939.
Alotaiby TN, Alshebeili SA, Alshawi T, EEG seizure detection and prediction algorithms: a survey. EURASIP J. Adv. Signal Process. 2014, 183 (2014). https://doi.org/10.1186/1687-6180-2014-183.
Yedurkar DP, Metkar SP, Stephan. Thompson, Multiresolution directed transfer function approach for segment-wise seizure classification of epileptic EEG signal. Cogn Neurodyn. 2022. DOI:10.1007/s11571-021-09773-z.
Priyasad D, Fernando T, Denman S, Sridharan. Sridha Fookes, Clinton, Interpretable Seizure Classification Using Unprocessed EEG with Multi-Channel Attentive Feature Fusion, IEEE Sens J (2021), 21, issue no. 17, DOI:10.1109/JSEN.2021.3090062.
Almustafa K, Mohamad. Classification of epileptic seizure dataset using different machine learning algorithms, Informatics in Medicine Unlocked(2020), Vol. 21, DOI: 10.1016/j.imu.2020.100444.
Dr AS. Kusumika Krori Dutta, Application of Machine Learning techniques in Electroencephalography signals, Brain & Behavior Computing, CRC press, (c) Taylor & Francis group(2021), ISBN: 978-1-003-09288-9.
Rajendran T, Sridhar KP, Vidhupriya P, Gayathri N, Anitha T. Epileptic seizure: Classification using autoregression features, Int J Curr Res Rev (2021) 13, 4, DOI:10.31782/IJCRR.2021.13429.
Yadav VP, Sharma K, Kumar. Variational mode decomposition-based seizure classification using Bayesian regularized shallow neural network, Biocybernetics and Biomedical Engineering (2021), 41, issue no. 2, DOI:10.1016/j.bbe.2021.02.003.
Alzami F, Tamamy AJ, Pramunendar, Ricardus Anggi Arifin, Zaenal, Fusion Of Bagging Based Ensemble Framework For Epileptic Seizure Classification, Transmisi (2020) Vol. 22, isuue no. 3, DOI: 10.14710/transmisi.22.3.102-106.
Panigrahi M, Behera DK, Patra K, Chandra. Epileptic seizure classification of electroencephalogram signals using extreme gradient boosting classifier, Indonesian J Electr Eng Comput Sci, (2022), 25, 2, DOI:10.11591/ijeecs.v25.i2.pp884-891.
Indira K, Dutta KK, Poornima S, Swamy Bellary SA. Deep Learning Methods for Data Science, Advanced Analytics and Deep Learning Models, John Wiley & Sons, Inc (2022), pp. 149–79, https://doi.org/10.1002/9781119792437.ch7.
Shoeibi A, Khodatars M, Ghassemi N, Jafari M, Moridian P, Alizadehsani R, Panahiazar M, Khozeimeh F, Zare AHosseini-Nejad, Khosravi H, Atiya A, Aminshahidi AF. Diba Hussain, Sadiq Rouhani, Modjtaba Nahavandi, Saeid Acharya, Udyavara Rajendra, Epileptic Seizures Detection Using Deep Learning Techniques: A Review, Int J Environ Res Public Health (2021), 18, 11, DOI:10.3390/ijerph18115780.
Daoud H, Bayoumi MA. Efficient Epileptic Seizure Prediction Based on Deep Learning. IEEE Trans Biomed Circuits Systems. 2019;13:5, pp. 804–13. DOI:10.1109/TBCAS.2019.2929053. issue no.
Cao X, Yao B, Chen B, Sun. Weifang Tan, Guowei, Automatic Seizure Classification Based on Domain-Invariant Deep Representation of EEG, Front NeuroSci (2021), 15, DOI:10.3389/fnins.2021.760987.
Woodbright M, Verma B, Haidar, Ali, Autonomous deep feature extraction based method for epileptic EEG brain seizure classification, Neurocomputing (2021), Vol. 444, DOI: 10.1016/j.neucom.2021.02.052.
Jia G, Lam HK, Althoefer, Kaspar. Variable weight algorithm for convolutional neural networks and its applications to classification of seizure phases and types, Pattern Recogn (2022) 121, DOI:10.1016/j.patcog.2021.108226.
Dutta KK, Poornima S, Sharma R, Nair D, Ploeger PG. Applications of Recurrent Neural Network: Overview and Case Studies, in Recurrent Neural Networks. CRC press; 2022. pp. 23–41. eBook ISBN9781003307822.
Dutta KK, Sridharan P, Bellary SArokiaS. Recurrent Neural Networks and Their Application in Seizure Classification, in Deep Learning in Visual Computing and Signal Processing, AAP Academic press(2022), Hard ISBN: 9781774638705, E-Book ISBN: 9781003277224.
Hussain W, Sadiq MT, Siuly S, Rehman A, Ur. Epileptic seizure detection using 1 D-convolutional long short-term memory neural networks, Appl Acoust (2021), 177, DOI:10.1016/j.apacoust.2021.107941.
Rout SK, Sahani M, Dora C, Biswal PK, Biswal. Birendra, An efficient epileptic seizure classification system using empirical wavelet transform and multi-fuse reduced deep convolutional neural network with digital implementation, Biomed Signal Process Control (2022) 72, DOI:10.1016/j.bspc.2021.103281.
Saminu S, Xu G, Shuai ZE, Kader IA, Jabire. Adamu Halilu Ahmed, Yusuf Kola Karaye, Ibrahim Abdullahi Ahmad, Isah Salim, A recent investigation on detection and classification of epileptic seizure techniques using EEG signal, Brain Sci (2021), 11, 5, DOI:10.3390/brainsci11050668.
Ein Shoka AA, Alkinani, Monagi H, El-Sherbeny AS, El-Sayed A, Dessouky, Mohamed M. Automated seizure diagnosis system based on feature extraction and channel selection using EEG, Brain Inf (2021) 8, 1, DOI:10.1186/s40708-021-00123-7.
Roy S, Asif U, Tang J, Harrer S, Seizure Type Classification Using EEG Signals and Machine Learning: Setting a Benchmark," 2020 IEEE Signal Processing in Medicine and Biology Symposium (SPMB), 2020, pp. 1–6, doi:10.1109/SPMB50085.2020.9353642.
Asif U, Roy S, Tang J, Harrer S, SeizureNet: Multi-Spectral Deep Feature Learning for Seizure Type Classification, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (2020), DOI: 10.1007/978-3-030-66843-3_8.
Dutta KK, Sharma S, Sasidharan A, Chetan S, Mukundan. Sleep monitoring wearables: Present to future, Wearable/Personal Monitoring Devices Present to Future, Springer, Singapore (2022), pp. 133–52. https://www.springerprofessional.de/en/sleep-monitoring-wearables-present-to-future/19797212.
Dutta KK, Multi-class time series classification of EEG signals with Recurrent Neural Networks, IEEE- International conference on cloud computing, data Science and Engineering, Confluence (2019), DOI: 10.1109/CONFLUENCE.2019.8776889.
Dutta KK, Indira DrK, Siri M. S, and Lakshminarayanan M. Multi-class Epileptic Seizure Classification Using Different Deep Learning Techniques- A Comparative Study, IJME (2022), Vol 7, issue no. 1, https://kalaharijournals.com/resources/IJME_Vol7.1_484.pdf.
Dutta KK, Manohar DPremila, Naaz F, Rajagopalan S. Epileptic Seizure Classification Based on Energy Compaction of Different Transformation Technique and Machine Learning Classifiers, IJME (2022),Vol 7, issue no. 1, https://kalaharijournals.com/resources/IJME_Vol7.1_485.pdf.
Khosla A, Khandnor P, Chand. Trilok, EEG-based automatic multi-class classification of epileptic seizure types using recurrence plots, Expert Syst (2022), 39, 5, DOI:10.1111/exsy.12923.
Peng H, Li C, Chao J, Wang T, Zhao, Chengjian Huo X, Hu. Bin, A novel automatic classification detection for epileptic seizure based on dictionary learning and sparse representation, Neurocomputing (2021), Vol. 424, DOI: 10.1016/j.neucom.2019.12.010.
Basri A, Arif. Muhammad, Classification of Seizure Types Using Random Forest Classifier, Advances in Science and Technology Research Journal (2021), Vol. 15, issue no. 3, DOI: 10.12913/22998624/140542.
Raghu S, Sriraam N, Temel Y, Rao SV, Kubben PL. EG based multi-class seizure type classification using convolutional neural network and transfer learning. Neural Netw. 2020;Vol. 124:202–12. DOI:10.1016/j.neunet.2020.01.017.
Purnomo A, Tjandrasa. Handayani, Epileptic Seizure Classification using Deep Batch Normalization Neural Network, Lontar Komputer: Jurnal Ilmiah Teknologi Informasi (2020) Vol. 11, issue no. 3, DOI: 10.24843/lkjiti.2020.v11.i03.p01.
Ahmedt-Aristizabal D, Fernando T, Denman S, Petersson L, Aburn MJ, Fookes, Clinton, Neural Memory Networks for Seizure Type Classification, Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society, EMBS (2020), DOI: 10.1109/EMBC44109.2020.9175641.
Seizure Corpus- Shah TUH-EEG, von Weltin V, Lopez E. McHugh S, Veloso J, Golmohammadi L, Obeid M, I., and Picone J. (2018). The Temple University Hospital Seizure Detection Corpus. Frontiers in Neuroinformatics. DOI: 10.3389/fninf.2018.00083.
Steriade C, Sperling MR, DiVentura B, Lozano M, Shellhaas RA, Kessler, Sudha Kilaru Dlugos, Dennis Dlugos, Dennis, Proposal for an updated seizure classification framework in clinical trials, Epilepsia (2022) Vol. 63 Issue 3, DOI: 10.1111/epi.17120.
Andrade-Machado RB, Cuartas V, Muhammad I, Khan. Recognition of interictal and ictal discharges on EEG. Focal vs generalized epilepsy, Epilepsy and Behavior (2021), 117, DOI:10.1016/j.yebeh.2021.107830.
Cao J, Grajcar K, Shan X, Zhao Y, Zou J, Chen L, Li Z, Grunewald R, Zis PD, Marco M, Unwin Z, Blackburn D, Sarrigiannis PG. Using interictal seizure-free EEG data to recognise patients with epilepsy based on machine learning of brain functional connectivity, Biomed Signal Process Control (2021) 67., DOI:10.1016/j.bspc.2021.102554.

Download PDF

Version 1

posted

You are reading this latest preprint version

Seven Epileptic Seizure Type Classification in Pre-Ictal, Ictal and Inter-Ictal Stages Using Machine Learning Techniques

Status:

Version 1

Abstract

Background

Aims

Methods

Result

Conclusion

Figures

Introduction

Methodology

Seizure Types And Descriptions

Dataset

Pre Processing

Feature extraction using Fast Fourier Transform (FFT)

X(K)= \(\sum _{n=0}^{N-1}x\left(n\right){e}^{\frac{2\pi Kn}{N}}\) (1)

A. Bagged Trees

B. K-Nearest Neighbors

C. Weighted KNN

D. Subspace KNN

E. Random Forest

Results

Discussion

Conclusion

Declarations

References

Status:

Version 1