TY - GEN
T1 - Common Spatial Pattern Based Data Augmentation Technique for Decoding Imagined Speech
AU - Panachakel, Jerrin Thomas
AU - Ganesan, Ramakrishnan Angarai
AU - Ananthapadmanabha, T. V.
N1 - Publisher Copyright:
© 2021 IEEE.
PY - 2021
Y1 - 2021
N2 - Modern machine learning techniques require a huge amount of training data for satisfactory performance. This limits the application of these techniques in classifying EEG signals due to the non-availability of huge datasets. One way to mitigate this problem is by using data augmentation techniques to increase the number of training samples. In this paper, a novel common spatial pattern (CSP) based data augmentation technique is proposed. The efficiency of the proposed method is demonstrated by training a deep neural network (DNN) on the augmented dataset for decoding imagined speech from EEG. Using CSP, nine EEG channels that best represent the underlying cortical activity corresponding to the imagination of the words 'in' and 'cooperate' are identified, and discrete wavelet transform (DWT) features are extracted for each of these channels. Treating the selected EEG corresponding to each imagined word as an independent sample helps in providing enough samples to train the DNN. Maximum voting is applied to the results of individual feature vectors of each trial to obtain the predicted class label. We have obtained accuracies exceeding change level accuracies across subjects, which indicates that the network is able to generalize well.
AB - Modern machine learning techniques require a huge amount of training data for satisfactory performance. This limits the application of these techniques in classifying EEG signals due to the non-availability of huge datasets. One way to mitigate this problem is by using data augmentation techniques to increase the number of training samples. In this paper, a novel common spatial pattern (CSP) based data augmentation technique is proposed. The efficiency of the proposed method is demonstrated by training a deep neural network (DNN) on the augmented dataset for decoding imagined speech from EEG. Using CSP, nine EEG channels that best represent the underlying cortical activity corresponding to the imagination of the words 'in' and 'cooperate' are identified, and discrete wavelet transform (DWT) features are extracted for each of these channels. Treating the selected EEG corresponding to each imagined word as an independent sample helps in providing enough samples to train the DNN. Maximum voting is applied to the results of individual feature vectors of each trial to obtain the predicted class label. We have obtained accuracies exceeding change level accuracies across subjects, which indicates that the network is able to generalize well.
KW - Brain-computer interface
KW - Common spatial pattern
KW - Deep learning
KW - Imagined speech
KW - Spatial filter
UR - https://www.scopus.com/pages/publications/85123387472
U2 - 10.1109/CONECCT52877.2021.9622684
DO - 10.1109/CONECCT52877.2021.9622684
M3 - Conference contribution
AN - SCOPUS:85123387472
T3 - Proceedings of CONECCT 2021: 7th IEEE International Conference on Electronics, Computing and Communication Technologies
BT - Proceedings of CONECCT 2021
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 7th IEEE International Conference on Electronics, Computing and Communication Technologies, CONECCT 2021
Y2 - 9 July 2021 through 11 July 2021
ER -