Motor imagery based on adaptive parameterless empirical wavelet transform and selective integrated classification

Abstract

Improving recognition rate of motor imagery (MI)-related electroencephalography (EEG) is of great importance for numerous brain computer interface (BCI) applications. However, the performance of a typical BCI system greatly relies on the effectiveness of the extracted features from raw EEG signals and the ability of the classifier to correctly identify different MI patterns. Therefore, in this paper, a new recognition method based on adaptive parameterless empirical wavelet transform (APEWT) and selective integrated classification model is proposed to enhance the classification accuracy of MI-related EEG signal. First, the APEWT is used to decompose EEG signals from different MI patterns into several intrinsic mode functions (IMFs), each of which contains different rhythm information over different frequency bands. Then several related modes are optimally selected based on the correlation coefficients calculated between each IMF component and the original signal to reconstruct EEG signals. Next, in order to further extract useful pattern information from both the time domain and frequency domain, the energy spectrum features of multiple time segments from the reconstructed signals and marginal spectrum features of different frequency bands corresponding to those selected modes are investigated, respectively. Finally, the extracted multiple features from time domain and frequency domain are input into the selective integrated classification model to build an MI recognition system. The selective integrated classification model consists of several extreme learning machines (ELMs) as the basic classifiers, different weights are assigned, respectively, to ELM basic classifiers based on their corresponding classification performances, and several basic ELM classifiers with good performances are selected to construct the final integrated model. The proposed method is evaluated on two public datasets, including BCI Competition Ⅱ dataset Ⅲ and BCI Competition IV dataset 2 b, and is compared with four different combination methods where different features in time domain or frequency domain in the feature extraction stage and different ELMs based classification models are considered. Experimental results demonstrate that the proposed method outperformed four combination methods and the existing methods recently reported in the literature using the same datasets in terms of classification accuracy and area under the ROC curve receiver operating characteristic metric. Specifically, our proposed method achieves the highest average classification accuracy (89.95%) in the compared methods, which indicates its better classification performance and generalization capability. In addition, the proposed method exhibits high computational efficiency, thus providing a new solution for online recognition of MI-related BCI and having the potential to be embedded in a practical system for controlling an external device.