| dc.description.abstract | In the first part of the thesis, the results of our studies on the classification of phonological categories in imagined words are presented. We have investigated whether there are any statistically significant differences in the mean phase coherence values in six cortical regions when the participants imagined uttering nasal and bilabial consonants. The cortical regions considered are prefrontal, premotor, motor, somatosensory, sensorimotor and auditory cortices, which are normally involved in speech production. We have observed statistically significant differences in the MPC values in all the six cortical regions in the beta band and in the motor cortex alone in the gamma band. The results obtained support the dual stream prediction model for imagined speech. Further, we have tried to classify the speech imagery EEG epochs based on the phonological category of the prompt using a shallow neural network. We have obtained an accuracy of 84.9% in the classification task when beta band MPC values are used. This is around 12% higher than the benchmark result on this dataset. The accuracies of our model dropped to 59.0% and 69.1%, respectively when only alpha or gamma band MPC values are used. The fact that EEG carries correlates of the phonological categories of the imagined phonemes can help in designing better prompts for a speech imagery-based BCI system.
In part two of the thesis, we have investigated as to whether we can classify the imagined prompt from the EEG recorded during speech imagery. For this, we have developed three different architectures. All these architectures address the problem of lack of sufficient training data. In the first one, which we call the CSP-based architecture, the issue of limited training data is addressed by treating the features extracted from selected EEG channels as distinct inputs to the classifier. Common spatial pattern (CSP) is used for channel selection. The primary classifier is a deep neural network (DNN) with four hidden layers whereas the secondary classifier is a majority voting classifier. The second model, which we call the TL-based architecture handles the problem of limited training data by using data augmentation and transfer learning (TL). In this architecture, MPC and MSC values from the alpha, beta and gamma bands are arranged as a 3D data, and are input to a ResNet50-based classifier, where ResNet50 is used as a fixed feature extractor. The third one, which we call the LSTM-based architecture employs overlapping window-based data augmentation to increase the amount of data, which is possible because the ASU dataset involves repeated imagination by the subjects. The LSTM-based architecture uses CSP for feature extraction, linear discriminant analysis (LDA) for dimensionality reduction and long short-term memory (LSTM) network as the primary classifier and a majority voting classifier as the secondary classifier.
All the three architectures have achieved above chance-level accuracies on the publicly available ASU speech imagery EEG dataset. The TL-based and the LSTM-based architectures have achieved average accuracies of 92.8% and 85.2%, respectively, for the classification of "short-long" words, which are better than the state-of-the-art. Although the LSTM-based architecture has lower accuracy than the TL-based architecture, the former can classify a 5 s EEG epoch in less than 110 ms, making it a suitable algorithm for online BCI systems. We have also performed ablation studies to identify the optimal number of EEG channels in the case of the CSP-based architecture and the optimal EEG frequency band in the case of the TL-based and the LSTM-based architectures. The optimal number of EEG channels in the case of CSP-based architecture is nine whereas the optimal EEG band for both the TL-based and the LSTM-based architectures is the gamma band.
In the last part of the thesis, we have proposed three architectures for classifying the altered state of consciousness during Rajayoga meditation from the resting state. Classification of Rajayoga meditation is challenging since it is probably unique in that the practitioners meditate with their eyes open. The first model, which we call the CSP-LDA architecture, uses CSP for feature extraction and LDA as the classifier. The second one, which we call the CSP-LDA-LSTM architecture is similar to the LSTM-based architecture used for decoding imagined speech. The third model, which we call the SVD-DNN architecture uses singular value decomposition (SVD) for choosing the relevant subspace of the signal and DNN for classification. The best intra-subject classification performance of 98.2% is obtained using the CSP-LDA-LSTM architecture whereas the best inter-subject performance of 96.4% is obtained using the SVD-DNN architecture. Both these architectures are able to capture subject-invariant features and can be deployed for grading the depth of meditation and for classifying other altered states of consciousness such as disorders of consciousness and hypnosis. | en_US |
