Blind Visualization of Task-Related Neural Networks from Simultaneous EEG-fMRI Data

Abstract

Introduction: Several stable EEG spatiospectral patterns were reported to be present in the EEG spectra [1,2]. Since the fixed canonical hemodynamic response function (HRF) was used in most similar studies utilizing spectral models [3-5], and the variable HRF function was utilized in studies investigating spatiospectral patterns [6], a direct comparison between spectral and spatiospectral models is still missing. The purpose of this study is to compare the efficacy of the spatiospectral and the spectral models with the overall goal to improve Visual Oddball Task (VOT) network visualization blindly and directly from acquired data.

Materials and Methods: Simultaneous EEG-fMRI data was obtained from 21 healthy subjects, each with four sessions of visual oddball paradigm consisting of 15% targets, 70% frequents, and 15% distractors1. For the spatiospectral approach, same spatiospectral group-ICA decomposition of EEG spectra was used as in reference 1 and the stable δ4 band pattern (as named in reference 6) demonstrating significant relation to the task [1,2] was selected. For the spectral approach, the signal in spectral domain was filtered by spectral average over the leads of the δ4 pattern. As in reference 6, EEG-fMRI data fusion at single-subject level was performed with linear regression calculating with variable HRF. The output was averaged across individuals with a one-way ANOVA test providing EEG-fMRI F-maps above the threshold p<0.001. Suprathreshold volume and descriptive parameters were extracted.

Results and Discussion: Based on visual inspection, the activation sites were consistent among approaches. The absolute power spectral model provided the highest volume of activations with highest statistical significance; relative power spatiospectral approach provided the 2nd largest and most significant results; absolute power spatiospectral approach demonstrated slightly lower activation volumes and significance than the previous one; the relative power spectral approach showed the lowest strength of the results. The fact that the spectral filter was designed based on a priori observation of the stable spatiospectral patterns limits its observed strength.

Conclusions: The first preliminary full comparison between spectral and spatiospectral models for EEG-fMRI data fusion was utilized for one stable spectral/spatiospectral pattern of VOT dataset. The current results show that the spectral approach utilizing absolute power fluctuations demonstrates the most significant outputs. More experimental trials need to be tested to ensure the result consistency over other optimizing approaches.

References: 1. Labounek et al. 2018 Brain Topography, 31(1), pp. 76-89; 2. Labounek et al. 2019 IFMBE Proceedings 68(2), pp. 125-128; 3. Rosa et al. 2010 NeuroImage 49, pp. 1496-1509; 4. Sclocco et al. 2014, Frontiers in Hum. Neurosci. 8, pp. 186; 5. Labounek et al. 2015 J of Neurosci. Meth. 245, pp. 125-136; 6. Labounek et al. 2019 J of Neurosci. Meth. 318, pp. 34-46