Generating Brain-maps with ICA Source Estimates from Hybrid Optimizer Using Spectral Screened EEG Data

Abstract

Electroencephalography (EEG) electrodes tend to sum the biologically plausible brain sources’ contributions, which depend on the strengths, orientations, and locations of the sources, due to the effect of volume conduction in the scalp. To separate the underlying sources, performing the independent component analysis (ICA) decomposition has been established to be the most appropriate technique available so far, provided the sources are linearly mixed without differential time delays in the recorded EEG signal. Towards this, the EEGLAB open source toolbox makes use of 20 variations of ICA for blind deconvolution of the actual EEG signal; the brain-maps generated using the unmixed EEG signal have been proved to be very useful in anatomical, functional, and pathological diagnosis. The ICA algorithms reported in the literature in the context of processing the EEG data, as far as we are aware, yield only local optimal solution for the multi-extremal optimization problem, wherein the independence criterion is the objective function. In this paper, we propose a hybrid optimizer — a combination of global and local optimizers — which is capable of producing near-global-optimal estimate of the ICA weight matrix, to reconstruct the separated EEG data. In the proposed scheme, unlike the conventional ICA algorithms adopted in the EEGLAB, the use of the Lie group techniques, to implicitly impose the orthonormality constraint among the IC estimates during the iterative steps of the optimization process, improves the IC estimation accuracy. To circumvent the computational complexity involved in the implementation of the global optimizer, the EEG data are first spectral screened to form a set of proto-type data vectors. In our experiments, this preprocessing step not only relieves the computational load, but also enhances the accuracy of the IC estimates, due to the redundancy removal in the original data. To conclude, we have demonstrated that the brain-maps generated using the hybrid optimizer in conjunction with the Lie group techniques, whose input is spectral screened prior to the IC estimation, remain far superior to those generated using the conventional approaches, in terms of the interpretability of the separated EEG data for diagnostic purposes.