Creating a Nonparametric Brain-Computer Interface with Neural Time-Series Prediction Preprocessing

The issue of subject-specific parameter selection in an electroencephalogram (EEG)-based brain-computer interface (BCI) is tackled in this paper. Hjorth- and Barlow-based feature extraction procedures (FEPs) are investigated along with linear discriminant analysis (LDA) for classification. These are well-known nonparametric FEPs but their simplicity prevents them from matching the performance of more complex FEPs. Neural time-series prediction preprocessing (NTSPP) has been shown to enhance the separability of both time- and frequency-based features and is used in this work to improve the applicability of these FEPs. NTSPP uses a number of prediction modules (PMs) to perform m-step ahead prediction of EEG time-series recorded whilst subjects perform motor imagery-based mental tasks. Depending on the PMs, the NTSPP framework normally requires subject-specific parameters to be predefined. In this work each PM is a self-organizing fuzzy neural network (SOFNN). The SOFNN has a self-organizing structure and good nonlinear approximation capabilities however; a number of parameters must be defined prior to training. This is problematic therefore the practicality of a general set of parameters, previously selected via a sensitivity analysis (SA), is analyzed. The results indicate that a general set of NTSPP parameters may provide the best results and therefore a fully nonparametric BCI may be realizable