A Novel Multi-period Multivariate Multi-scale Phase Locking Value and its Application

Abstract

Motor Imagery EEG or ECoG is the most popular driving signal in brain computer interface based rehabilitation system. Empirical Mode Decomposition (EMD) can be employed in feature extraction, which only a single scale IMF is considered by using Phase Locking Value (PLV), leading to the loss of phase information. In this paper, a Multi-period Multivariate Multi-scale PLV (MMMPLV) is proposed based on Noise-Assisted Multivariate EMD (NAMEMD). The selected multi-channel MI-ECoG are decomposed simultaneously by NAMEMD to obtain the multivariate multi-scale IMFs, and their length are divided into many periods. Then the PLV of pair-wise IMFs at the same scale are calculated in each time period for any two-channel MI-ECoG signals. The resulted MMMPLV are constructed as phase features. Furthermore, the phase features generated by MMMPLV and the spatial features extracted by Common Spatial Subspace Decomposition (CSSD) algorithm are fused in series, yielding a new feature extraction method, denoted as MMMPC. Experiments were conducted on a public database, and the Support Vector Machine (SVM) is used to classify the combined features. The experiment results of 9-fold Cross Validation (CV) show that the proposed method yields relative higher classification accuracy and better stability compared with the other synchronization methods and classical feature extraction methods.