NUST Institutions Library Catalogue

Classification of neural correlates of hand motions from EEG signals recently gained attention of
researchers for the development of BCI systems for persons suffered from stroke, spinal cord
injury who are not able to do voluntary movements or person with amputated arm or legs.
Commonly LDA, SVM and K-NN models are used for the classification of hand motions, CNN
and hybrid models are also used but most methods include the complex methods or preprocessing of EEG data and extraction of time or frequency domain features from the preprocessed signals which is a time consuming and lack flexibility because the EEG signals vary
from human to human. In this thesis a Deep CNN model for end-to-end0learning of neural
corelates for reach and grasp actions is introduced, aiming to increase rate of recognition &
balanced classification0accuracy throughout all the subjects. A new model of CNN for
movement classification is proposed that can also be used on the edge devices because of its
smaller size for the development of BCI systems. In the proposed model separable convolutional
blocks are used which reduce the number of parameters and hence the size of model also
decreases. The dataset that is used for the testing of model is BNCI Horizon 2020 Reach and
Grasp action dataset that is publicly available dataset. The dataset is also tested on 3 machine
learning models LDA, SVM and K-NN are used, in which input is given in the form of time
domain feature set the average accuracies achieved on these models are 60.77 (±3.80 STD),
66.73 (±2.86 STD), and 79.81 (±3.11 STD) respectively on the unseen dataset. Then the dataset
is tasted on proposed Deep learning model along with DeepConvNet and EffNet models. The
proposed model achieves the average classification accuracy of 92.44 (±4.13 STD), 92.9 (±4.23
STD) and 81.7(±5.68 STD). The model proposed achieves the same accuracy as DeepConvNet,
but the size of proposed model is far smaller than the DeepConvNet model. Results shows that
the proposed model shows the improved results with less variation of results within the subjects.
Which will become helpful in the creation of real time BCI systems.