Extracting multi-mode ERP features using fifth-order nonnegative tensor decomposition

Abstract

Background: Preprocessed Event-related potential (ERP) data are usually organized in multi-way tensor, in which tensor decomposition serves as a powerful tool for data processing. Due to the limitation of computation burden for multi-way data and the low algorithm performance of stability and efficiency, multi-way ERP data are conventionally reorganized into low-order tensor or matrix before further analysis. However, the reorganization may hamper mode specification and spoil the interaction information among different modes.

New method: In this study, we applied a fifth-order tensor decomposition to a set of fifth-order ERP data collected by exerting proprioceptive stimulus on left and right hand. One of the latest nonnegative CANDECOMP/PARAFAC (NCP) decomposition methods implemented by alternating proximal gradient (APG) was employed. We also proposed an improved DIFFIT method to select the optimal component number for the fifth-order tensor decomposition.

Results: By the fifth-order NCP model with a proper component number, the ERP data were fully decomposed into spatial, spectral, temporal, subject and condition factors in each component. The results showed more pairs of components with symmetric activation region in left and right hemisphere elicited by contralateral stimuli on hand.

Comparison with existing method(s): In our experiment, more interesting components and coherent brain activities were extracted, compared with previous studies.

Conclusions: The discovered activities elicited by proprioceptive stimulus are in line with those in relevant cognitive neuroscience studies. Our proposed method has proved to be appropriate and viable for processing high-order EEG data with well-preserved interaction information among all modes.

Keywords: CANDECOMP/PARAFAC; Component number selection; Event-related potential; Multi-mode features; Nonnegative tensor decomposition.