Abstract:
Objective. The classification and features extraction is extremely important part of brain
computer interface (BCI) systems. During the previous few years, deep learning techniques
have been employed on different types of data for features extractions, and classification
purposes. However, few people have employed these deep learning techniques on BCI
applications. Approach The goal is to acquire features using the deep learning techniques,
which eventually enhance the model classification performance. In this research, we
investigated at the integration of CNN and LSTM. The proposed CNN architecture comprises
of four block for the purpose of features extraction. Each block comprises of a Convolutional
Layer, Batch Normalization Layer (BNL), Rectified Linear Unit (RELU) activation layer and
one Max-Pooling layer. To begin, we converted motor imagery EEG signals into 2D images
using the Continuous wavelet transform. These images are fed in to a proposed framework.
CNN Block are used to extract robust features and LSTM architecture for classification
purposes. Main results. Our model was assessed on the dataset III of BCI competition II where
we achieved 96.43% and kappa value reached 0.9286. We also assessed the performance of
our suggested model on Dataset 2b from BCI Competition IV where it achieved average
accuracy of 86.83% and average kappa value of 0.742.
Significance. Our proposed method has shown that deep learning technique gave remarkable
classification results compared to traditional approaches.
