Development of a Deep Learning Model for Motor-Imagery Classification

EEG-based intention recognition systems, also termed as EEG-based motor imagery (MI) systems, have shown promise as assistive devices e.g., neuro-robotics and as an improvement of the quality of life (Q-O-L) for the physically challenged individuals. The system enables non-muscular communications and non-linear controls that can be utilized further for Brain Machine Interfaces (BMIs). In this study where EEG signals from 109 subjects are used which are then decomposed into alpha, beta, theta, and delta based on their frequency ranges using low band filter l(n) and high band filters h(m). To extract distinguished sequential features from signals, multi tasks LSTM are employed. Finally, the Long Short-Term Memory (LSTM) is applied to softmax layers for motion intention classification of fists, feet, and eyes. This manuscript aims to conduct ample experiments on publicly available EEG benchmark dataset, decomposing the signals according to their frequency bands and applying Multiple Task Deep Recurrent Neural Network to identify multiple intentions. Experiments reveals that proposed network gives promising accuracy of 96.35% with an error rate of 3.65%.