Exploiting dimensionality reduction and neural network techniques for the development of expert brain–computer interfaces

Abstract

Background:Analysis and classification of extensive medical data (e.g. electroencephalography (EEG) signals) is a significant challenge to develop effective brain–computer interface (BCI) system. Therefore, it is necessary to build automated classification framework to decode different brain signals. Methods:In the present study, two-step filtering approach is utilize to achieve resilience towards cognitive and external noises. Then, empirical wavelet transform (EWT) and four data reduction techniques; principal component analysis (PCA), independent component analysis (ICA), linear discriminant analysis (LDA) and neighborhood component analysis (NCA) are first time integrated together to explore dynamic nature and pattern mining of motor imagery (MI) EEG signals. Specifically, EWT helped to explore the hidden patterns of MI tasks by decomposing EEG data into different modes where every mode was consider as a feature vector in this study and each data reduction technique have been applied to all these modes to reduce the dimension of huge feature matrix. Moreover, an automated correlation-based components/coefficients selection criteria and parameter tuning were implemented for PCA, ICA, LDA, and NCA respectively. For the comparison purposes, all the experiments were performed on two publicly available datasets (BCI competition III dataset IVa and IVb). The performance of the experiments was verified by decoding three different channel combination strategies along with several neural networks. The regularization parameter tuning of NCA guaranteed to improve classification performance with significant features for each subject. Results:The experimental results revealed that NCA provides an average sensitivity, specificity, accuracy, precision, F1 score and kappa-coefficient of 100% for subject dependent case whereas 93%, 93%, 92.9%, 93%, 96.4% and 90% for subject independent case respectively. All the results were obtained with artificial neural networks, cascade-forward neural networks and multilayer perceptron neural networks (MLP) for subject dependent case while with MLP for subject independent case by utilizing 7 channels out of total 118. Such an increase in results can alleviate users to explain more clearly their MI activities. For instance, physically impaired person will be able to manage their wheelchair quite effectively, and rehabilitated persons may be able to improve their activities.