BCI applications based on artificial intelligence oriented to deep learning techniques

Abstract

A Brain-Computer Interface, BCI, can decode the brain signals corresponding to the intentions of individuals who have lost neuromuscular connection, to reestablish communication to control external devices. To this aim, BCI acquires brain signals as Electroencephalography (EEG) or Electrocorticography (ECoG), uses signal processing techniques and extracts features to train classifiers for providing proper control instructions. BCI development has increased in the last decades, improving its performance through the use of different signal processing techniques for feature extraction and artificial intelligence approaches for classification, such as deep learning-oriented classifiers. All of these can assure more accurate assistive systems but also can enable an analysis of the learning process of signal characteristics for the classification task. Initially, this work proposes the use of a priori knowledge and a correlation measure to select the most discriminative ECoG signal electrodes. Then, signals are processed using spatial filtering and three different types of temporal filtering, followed by a classifier made of stacked autoencoders and a softmax layer to discriminate between ECoG signals from two types of visual stimuli. Results show that the average accuracy obtained is 97% (+/- 0.02%), which is similar to state-of-the-art techniques, nevertheless, this method uses minimal prior physiological and an automated statistical technique to select some electrodes to train the classifier. Also, this work presents classifier analysis, figuring out which are the most relevant signal features useful for visual stimuli classification. The features and physiological information such as the brain areas involved are compared. Finally, this research uses Convolutional Neural Networks (CNN) or Convnets to classify 5 categories of motor tasks EEG signals. Movement-related cortical potentials (MRCPs) are used as a priori information to improve the processing of time-frequency representation of EEG signals. Results show an increase of more than 25% in average accuracy compared to a state-of-the-art method that uses the same database. In addition, an analysis of CNN or ConvNets filters and feature maps is done to and the most relevant signal characteristics that can help classify the five types of motor tasks.