MIDNN- a classification approach for the EEG based motor imagery tasks using deep neural network

In recent times, Motor Imagery (MI) tasks have gained great attraction among researchers in the field of Brain-Computer Interface (BCI). The MI tasks are the core field of the human brain in which the person imagines the movement of the body parts without performing actual movement of the body. MI tasks cause the activation of lakhs of neurons in the brain to interact with each other. The activation of the neurons generates electrical signals that can be captured through electroencephalogram (EEG) devices. The MI-EEG-based signals can move the external devices such as a wheelchair, moving cursors, etc., and hence, are very helpful to design and develop personal assistants for the disabled person for interaction and communication to the outside world. In this paper, MI-EEG data for left-hand(LH) and right-hand(RH) movements are recorded using a multi-channel EEG device. Further, a Deep Neural Network (DNN) model (MIDNN) is proposed for the binary-class classification of the collected dataset. The performance of the proposed model has been tested on the BCI benchmark dataset BCI competition III V dataset for LH and RH MI tasks.The fifth order low pass Butterworth filter is used to denoise the raw signals and then decomposed into six frequency sub-bands of (0.5–4) Hz, (4–8) Hz, (8–12) Hz, (12–16) Hz, (16–24) Hz, and (24–40) Hz using Butterworth bandpass filter of same order. The sub-bands are used to extract features from MI-EEG signals of LH and RH movement using welch power spectral density (PSD). The accuracy obtained by the MIDNN model is around 70% on the local dataset and 72.51% on the BCI dataset using PSD as features from each channel for classification of LH and RH tasks. To further improve the performance of the model, the spectral features from the estimated PSD of each of the six sub-band are obtained in the form of band power. The accuracy obtained by the same MIDNN model using band power as features is 88.89% on the local dataset and 82.48% on the V dataset of BCI competition III. The proposed MIDNN model acheived a significant increase in classification accuracy by 13.7% and 26.9% on BCI and Emotiv dataset respectively.