Improving Signal Separability and Inter-Session Stability for a Brain-Computer Interface by Time-Series-Prediction-Preprocessing

This paper presents a preprocessing procedure for improving the separability of electroencephalogram (EEG) signals recorded from subjects for a right/left motor imagery based brain-computer interface (BCI). The EEG data is preprocessed utilizing a recently proposed time-series-prediction (TSP) technique. Two neural networks (NNs) are trained to perform one-step-ahead predictions of the EEG time-series data where one NN is trained to predict right motor imagery signals and the other left motor imagery signals. The NNs are used in a procedure referred to as neural-time-series-prediction-preprocessing (NTSPP) where signals are fed into both NNs and two new signal types are produced i.e. the predicted signals (Ys) or the prediction error signals (Es). In this investigation the well known adaptive autoregressive modeling (AAR) technique is used to extract features from the Es and Ys signals. Classification is performed using linear discriminant analysis (LDA). This NTSPP procedure is tested offline on three subjects and classification accuracy (CA) rates approach 98%. The approach shows significant potential for improving robustness and feature stability across sessions and a clearly distinguishable improvement in performance is observed when features are extracted from the NTSPP signals compared to those extracted from the original signals (Os)