Using a smart method for automatic diagnosis in medical applications, such as sleep stage classification is considered as one of the important challenges of the last few years which can replace the time-consuming process of visual inspection done by specialists. One of the problems regarding the automatic diagnosis of sleep patterns is extraction and selection of discriminative features generally demanding high computational burden. This paper provides a new single-channel approach to automatic classification of sleep stages from EEG signal. The main idea is to directly apply the raw EEG signal to deep convolutional neural network, without involving feature extraction/selection, which is a challenging process in the previous literature. The proposed network architecture includes 9 convolutional layers followed by 2 fully connected layers. In order to make the samples of different classes balanced, we used a preprocessing method called data augmentation. The simulation results of the proposed method for classification of 2 to 6 classes of sleep stages show the accuracy of 98.10%, 96.86%, 93.11%, 92.95%, 93.55% and Cohen's Kappa coefficient of 0.98%, 0.94%, 0.90%, 0.86% and 0.89%, respectively. Furthermore, comparing the obtained results with the state-of-the-art methods reveals the performance improvement of the proposed sleep stage classification in terms of accuracy and Cohen's Kappa coefficient.
Common spatial pattern (CSP) is shown to be an effective pre-processing algorithm in order to discriminate different classes of motor-based EEG signals by obtaining suitable spatial filters. The performance of these filters can be improved by regularized CSP, in which available prior information is added in terms of regularization terms into the objective function of conventional CSP. Variety of prior information can be used in this way. In this paper, we used time correlation between different classes of EEG signal as the prior information, which is clarified similarity between different classes of signal for regularizing CSP. Furthermore, the proposed objective function can be easily extended to more than two-class problems. We used three different standard datasets to evaluate the performance of the proposed method. Correlation-based CSP (CCSP) outperformed original CSP as well as the existing regularized CSP, Principle Component Cnalysis (PCA) and Fisher Discriminate Analysis (FDA) in both two-class and multi-class scenarios. The simulation results showed that the proposed method outperformed conventional CSP by 6.9% in 2-class and 2.23% in multi-class problem in term of mean classification accuracy.
The incremental least-mean-square (ILMS) algorithm is a useful method to perform distributed adaptation and learning in Hamiltonian networks. To implement the ILMS algorithm, each node needs to receive the local estimate of the previous node on the cycle path to update its own local estimate. However, in some practical situations, perfect data exchange may not be possible among the nodes. In this paper, we develop a new version of ILMS algorithm, wherein in its adaptation step, only a random subset of the coordinates of update vector is available. We draw a comparison between the proposed coordinate-descent incremental LMS (CD-ILMS) algorithm and the ILMS algorithm in terms of convergence rate and computational complexity. Employing the energy conservation relation approach, we derive closed-form expressions to describe the learning curves in terms of excess mean-square-error (EMSE) and mean-square deviation (MSD). We show that, the CD-ILMS algorithm has the same steady-state error performance compared with the ILMS algorithm. However, the CD-ILMS algorithm has a faster convergence rate. Numerical examples are given to verify the efficiency of the CD-ILMS algorithm and the accuracy of theoretical analysis.