EEG signals classification based on autoregressive and inherently quantum recurrent neural network

Abstract

This paper shows a novel hybrid approach using an Auto-Regressive (AR) model and a Quantum Recurrent Neural Network (QRNN) for classification of two classes of Electroencephalography (EEG) signals. The QRNN-AR has been shown to be capable to capture and quantify the uncertainty inherently in EEG signals because it uses fuzzy decision boundaries to partition the feature space. Two diverse element extraction techniques were used to extract the features from EEG signals; AR coefficients are processed with Levinson-Durbin algorithm and mean square error. AR provides a better frequency resolution and good spectral data of short EEG segment. The QRNN trained by the back propagation algorithm is compared with Quantum Neural Network (QNN) and Quantum Wavelet Neural Network (QWNN). The average accuracy of the proposed QRNN model is 88.28452% at 6 seconds. The Accuracy to Time Ratio (ATR) value is 14.714086, which shows the superiority of the proposed model. Experimental results demonstrate that the QRNN-AR gives the highest overall accuracy and short processing time. In addition, the structure of the proposed method is more reliable.