Classification of Motor Imagery based EEG Signals Using Deep Learning Architecture

Brain-computer interfaces (BCIs) aim to translate brain activity into control signals for external devices. Motor imagery (MI), the mental rehearsal of movements, has been studied extensively in BCIs as a means of decoding user intent from neural activity. Electroencephalogram (EEG) recordings provide a non-invasive interface for monitoring MI by detecting changes in brain activation. Accurately classifying MI from EEG signals is essential for developing practical MI-based BCIs. However, existing machine learning techniques require extensive preprocessing and handcrafted feature engineering, limiting their accuracy and efficiency. In this work, we propose a deep learning approach for end-to-end MI classification from raw EEG signals without manual feature extraction. Specifically, we employ a long short-term memory (LSTM) network to automatically learn spatiotemporal features from the EEG data. The extracted features are then fed into a multilayer perceptron (MLP) to perform classification. We demonstrate that our proposed framework achieves significantly higher accuracy in classifying right-hand and foot MI compared to conventional machine learning classifiers. Our results highlight the potential of deep learning to advance MI-based BCIs by learning highly complex representations of EEG signals in an end-to-end fashion. The proposed LSTM-MLP model thus represents a promising step towards more sophisticated BCIs that can translate raw neural data into control signals without the need for handcrafted features.