Loading...

Classification of Motor Imagery in Electroencephalogram Signal Based on Spatio-temporal Feature selection Using Elastic Net

Noei, Shahryar | 2017

645 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 49868 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Jahed, Mehran
  7. Abstract:
  8. Motor imagery causes Event related Synchronization/Desynchronization (ERS/ERD) in Electroencephalogram (EEG) signal. These potentials can be used as an input for a Brain Computer Interface (BCI) system. To do so, it is necessary for these inputs to be correctly classified. The quality of classification is severely affected by the features extracted. Common Spatial Patterns (CSP) algorithm is often used for this task. Some of this method disadvantages are neglecting non-stationary properties of EEG signal and its proneness to overfitting. Additionally, its success is highly dependent on the frequency band that the algorithm is performed in. The most suitable sub band is interchangeable between different subjects. In this work, the usage of Filter Bank Common Spatial Patterns in different sub bands as spatial features in conjugation with the EEG time series as temporal features, for considering the EEG non-stationary properties, is proposed. Since using these features makes the classification task a high dimensional one, a slightly modified elastic net method is proposed for feature selection. This algorithm is specifically designed for high dimensional problems. Moreover, Stochastic Gradient Boosting method is used for the classification of the non-stationary features. Since this algorithm tries to improve the classification result by the means of reducing its bias, it is prone to overfitting. To solve this problem, the Bagging method is proposed to reduce the variance. In the end, the proposed method is validated using two different datasets. The first dataset is for the BCI competition IV, dataset IIb, which consists of 9 different subjects. The mean accuracy obtained on the train data, using 10×10-fold cross validation method is 82.5±7, and for the test data is 83.8±8 (Kappa coefficient of 0.68±0.17). The second dataset is gathered by the writers of this thesis and it consists of 4 different subjects. The mean accuracy obtained on this data using 10×10-fold cross validation is 77.1±15. Obtained results show that the performance of the proposed method outperforms most of common methods and methods proposed by other papers. This improvement is significance in users that spatial features were not able to produce desirable results solely
  9. Keywords:
  10. Motor Imagery (MI) ; Elastic Network Model ; Brain-Computer Interface (BCI) ; Electroencphalogram ; Spatiotemporal Patterns ; Common Spatical Patterns ; Gradiant Boosting

 Digital Object List

 Bookmark

No TOC