A Recurrent Probabilistic Neural Network with Dimensional Reduction and Its Application to Time Series EEG Discrmination

This paper proposes a novel reduced-dimensional recurrent probabilistic neural network, and tries to classify electroencephalography (EEG) during motor images. In general, a recurrent probabilistic neural network (RPNN) is a useful tool for pattern discrimination of biological signals such as electromyograms (EMGs) and EEG due to its learning ability. However, when dealing with high dimensional data, RPNNs usually have problems of heavy computation burden and difficulty in training. To overcome these problems, the proposed RPNNs incorporates a dimension-reducing stage based on linear discriminant analysis into the network structure, and a hidden Markov model (HMM) and a Gaussian mixture model (GMM) are composed in the network structure for time-series discrimination. The proposed network is also applied to EEG discrimination using Laplacian filtering and wavelet packet transform (WPT). Discrimination experiments of EEG signals measured during calling motor images in mind were conducted with four subjects. The results showed that the proposed method can achieve relatively high discrimination performance (average discrimination rates: 84.6± 5.9%), and indicated that the method has possibility to be applied for the human-machine interfaces.