Enhanced Accuracy for Motor Imagery Detection Using Deep Learning for BCI

Abstract

The Brain-Computer Interface (BCI) is a system where the human brain and a hardware device interact instantly. This transfers the brain data recorded directly to the computer that can be used for external system control. There are four key components of the BCI method, namely the acquisition of signals, preprocessing of acquired signals, extraction, and classification of features. In conventional machine learning algorithms, the accuracy achieved is negligible and not up to the mark to classify motor imagery data for multiple classes. The main explanation for this is that features are manually selected, and we are unable to get certain features that result in greater precision. For classifying the multi-class motor imagery (MI) data, we have implemented deep learning algorithms in this work. Two different approaches have been explored in this study: Artificial Neural Network (ANN) and Long Short-Term Memory (LSTM). We evaluate the accuracy of classification on two datasets, i.e. Competition for BCI III, dataset IIIa and competition for BCI IV, dataset IIa. The results showed that deep learning algorithms provide higher accuracy outcomes than conventional machine learning algorithms. LSTM outperforms the ANN and the deep learning classifier gives 96.2 percent average classification accuracy.