Extraction of common task features in EEG-fMRI data using coupled tensor-tensor decomposition

Abstract

The fusion of simultaneously recorded EEG and fMRI data is of great value to neuroscience research due to the complementary properties of the individual modalities. Traditionally, techniques such as PCA and ICA, which rely on strong strong nonphysiological assumptions such as orthogonality and statistical independence, have been used for this purpose. Recently, tensor decomposition techniques such as parallel factor analysis have gained more popularity in neuroimaging applications as they are able to inherently contain the multidimensionality of neuroimaging data and achieve uniqueness in decomposition without imposing strong assumptions. Previously, the coupled matrix-tensor decomposition (CMTD) has been applied for the fusion of the EEG and fMRI. Only recently the coupled tensor-tensor decomposition (CTTD) has been proposed. Here for the first time, we propose the use of CTTD of a 4th order EEG tensor (space, time, frequency, and participant) and 3rd order fMRI tensor (space, time, participant), coupled partially in time and participant domains, for the extraction of the task related features in both modalities. We used both the sensor-level and source-level EEG for the coupling. The phase shifted paradigm signals were incorporated as the temporal initializers of the CTTD to extract the task related features. The validation of the approach is demonstrated on simultaneous EEG-fMRI recordings from six participants performing an N-Back memory task. The EEG and fMRI tensors were coupled in 9 components out of which 7 components had a high correlation (more than 0.85) with the task. The result of the fusion recapitulates the well-known attention network as being positively, and the default mode network working negatively time-locked to the memory task.

Early soft and flexible fusion of electroencephalography and functional magnetic resonance imaging via double coupled matrix tensor factorization for multisubject group analysis

“…Furthermore, in the oddball paradigm, our model obtains activated areas congruent with the literature and provides also frequency information that is not possible with the usual analysis in the framework of GLM. We need also to stress out that the use of tensor-tensor approaches (Chatzichristos et al, 2018;Jonmohamadi et al, 2020) could not have been applied in none of the use cases, since the subjects do not have the same time courses. Hence, DCMTF is the only method available (to the best of the authors' knowledge) that can be used in a (semi-) blind context using raw data and not predefined features in both of the analyzed use cases.…”

“…One can note that the studies employing tensor decompositions in multisubject fusion applications are using only block-event designs for this reason (Chatzichristos et al, 2018;Jonmohamadi et al, 2020).…”

Low-Rank Tensor Patching Based on Convolutional Sparse Coding for Communication Data Repair

“…In recent years, the research area of low-rank tensors has been extended to numerical analysis, computer vision, machine learning, signal and image processing, etc. [27][28][29][30][31]. erefore, it is worthwhile to pay attention and explore how to perform efficient and accurate tensor complementation to recover the information of each dimension of the incomplete tensor.…”

Functional, structural, and phenotypic data fusion to predict developmental scores of pre-school children based on Canonical Polyadic Decomposition

“…A number of studies have carried out coupled matrix-tensor decompositions of EEG and MRI data successfully. Simultaneous recording EEG and fMRI were jointly analysed based on canonical polyadic decomposition (CPD), also known as PARAFAC [37], in [38,39,40,41] to study neural activity. CPD was also used to fuse EEG and fMRI with the purpose of characterising neurological disease such as schizophrenia [42] and epilepsy [43,44].…”