A Novel Efficient Machine Learning Approach for Multiclass Classification of EEG Signal

Title:A Novel Efficient Machine Learning Approach for Multiclass Classification of EEG Signal

Volume: 11 Issue: 3

Author(s): Satyajit Anand*, Sandeep Jaiswal and Pradip Kumar Ghosh

Affiliation:

• Department of ECE, College of Engineering and Technology, Mody University of Science and Technology, Lakshmangarh,India

Keywords: Epilepsy, gaussian filter, scale invariant feature transform, harris operator, local binary pattern, multiclass least square support vector machine.

Abstract: Background: Epilepsy is one among the neurological disorders, which occurs due to temporary and irregular electrical disturbance in the brain. EEG signals are recorded from the patients with seizure. Gaussian filter is one of the pre-processing approaches, which is mainly applied to remove the unwanted signal and improve the signal quality. Harris operator is utilized for key point localization to remove the bad key points from the pre-processed signal. Here, the Scale Invariant Feature Transform (SIFT) is employed with the Local Binary Pattern (LBP) to identify the key points of the signal and these signal values are considered as the feature vector values, which are applied on Multiclass least square Support Vector Machine (MLS-SVM) with four different coding schemes: Error correcting code output code of MLS-SVM (MLS-SVM_ECOC), minimum output codes of MLS-SVM (MLS-SVM_MOC), one versus one of MLS-SVM (MLS-SVM_1v1), one versus all of ML-SVM (MLS-SVM_1vA) using kernel function to categorize epileptic seizure in Multiclass EEG signals by comparing the test features with trained features. The attained results demonstrated that the proposed technique outperformed others existing methods.

Methods: The noise is present in the real-time EEG signals. The noise is removed by applying Gaussian filter to the input signal. Key point is localized to generate the features by using the algorithm of SIFT. Then the local features are obtained by the method of LBP. Finally, the signals are categorized into a seizure and seizure-free signal by comparing the optimized features with trained features in SVM classifier. In this work, a novel classification algorithm i.e., Multiclass least square Support Vector Machine (MLS-SVM) is introduced with four coding schemes: MLS-SVM_ECOC, MLS-SVM_MOC, MLS-SVM_1v1, and MLS-SVM_1vA for multiclass classification of EEG signal. To solve the multiclass categorization in EEG signal, we reformulate the extracted features from the EEG signal to the set of binary classification by utilizing four different output coding schemes. The proposed technique is assessed using the performance metrics like accuracy and compared with other previously existing methods.

Results: The performance of the proposed technique is evaluated in terms of accuracy. In the suggested classifiers, we select 5-fold cross-validation. At every time, one fold is kept for testing set and remaining four folds are utilized for training set. The result shows that γ= 1 and σ2=10 is the best optimal parameter combination that produces the highest average classification accuracy 99.22% for MLS-SVM_1v1 as compared to the rest of the three proposed output scheme codings.

Conclusion: Epilepsy is the neurological disorder where the electrical movement of the brain is changed because of the person’s activity changes from the up-normal level. The symptoms of the seizure are unconsciousness and health issues. It may disturb the complete body or parts of the brain. During evaluation, distinct experimental measures are used to analyze the performance of the classifier. The accuracy of 99.29 % was reported by using the proposed method. From the results, it is validated that the anticipated MLS-SVM_1v1 offers the better results.

Cite this article as:

Anand Satyajit*, Jaiswal Sandeep and Ghosh Kumar Pradip, A Novel Efficient Machine Learning Approach for Multiclass Classification of EEG Signal, Recent Patents on Computer Science 2018; 11 (3) . https://dx.doi.org/10.2174/2213275911666180816105115