GENERALIZED EEG-FMRI SPECTRAL AND SPATIOSPECTRAL HEURISTIC MODELS René Labounek 1,2,3 , David Janeček 1 , Radek Mareček 2 , Martin Lamoš 1,2 , Tomáš Slavíček 1,2 , Michal Mikl 2 , Jaromír Baštinec 4 , Petr Bednařík 2,3,5 , David Bridwell 6 , Milan Brázdil 2 , Jiří Jan 1 1 Department of Biomedical Engineering, Brno University of Technology, Brno, Czech Republic 2 Central European Institute of Technology, Masaryk University, Brno, Czech Republic 3 Center for Magnetic Resonance Research, University of Minnesota, Minneapolis, USA 4 Department of Mathematics, Brno University of Technology, Brno, Czech Republic 5 Division of Endocrinology and Diabetes, University of Minnesota, Minneapolis, USA 6 Mind Research Network, Albuquerque, NM 87106, USA ABSTRACT The aim of the current study is visualization of task-related variability in EEG-fMRI data, performed as a blind-search analysis without stimulus timings, using a methodology that is based on Kilner’s et al. heuristic approach [2]. We show that filters of the relative EEG spectra with different fre- quency responses visualize different task-related brain net- works. The effect is more pronounced within an event- related oddball paradigm (i.e. detecting rare visual targets) than within a block-design semantic decision paradigm (i.e. detecting semantic errors). The mutual information between different EEG-fMRI activation maps calculated with filters of different frequency responses appears stable between the different paradigms. We also introduce preliminary results implementing the heuristic analysis with spatiospectral EEG components, where the filter response has two dimensions and depends on frequency and channels. Index TermsSimultaneous EEG-fMRI, heuristic model, GLM, ICA 1. INTRODUCTION In the current study we extended current models used to analyze simultaneously measured EEG-fMRI signals. The ultimate goal of work was to reconstruct task-related net- works from EEG-fMRI data without prior knowledge of stimulation timing (see also [1-3]). Recently, we proposed that relative (i.e. normalized) EEG power [p(ω), eq. 1] is more powerful than absolute power for visualizing task-related EEG-fMRI networks [1]. The strongest task-related correlates between relative EEG power and fMRI-BOLD [4] signal were observed within the α (8-12Hz) and γ (20-40Hz) bands. () = () ∫ ()  (1) In eq. 1, s(ω) represents the power spectral density for frequency ω in the acquisition time duration of the n-th fMRI scan. The numerator and denominator fluctuate over time, sometimes asynchronously across frequencies. Kilner et al. [2] introduced the EEG-fMRI heuristic model in 2005 (eq. 2). This approach fuses the data based on models of trans-membrane voltage and current changes, i.e. directly at the level of neuronal changes. [ ] 2 ∝ (1 + ) 2 ∫ 2 () ∫ 2 () (2) This simplest model declares that changes in BOLD signal b are proportional to neuronal activation a which is proportional to changes in root mean square frequency of whole normalized (relative) EEG power spectrum p(ω). The character ~ indicates variables during increased activity, while variables without ~ represent signal values during rest. Rosa et al. (2010) [3] considered the denominators in equation (2) to have a constant baseline, and incorporated general linear model (GLM) fitting (eq. 3). ∝ √∫  2 () (3) Fig. 1 shows the function ω 2 (in eq. 3) is only different frequency response of the relative EEG spectra filter in contrast to the response of a band of interest. In our previous study [1], we calculated task-related ac- tivation maps for different frequency bands of interest, sug- Fig. 1. Different weighting functions g(ω) for classic heuristic model and different frequency bands of interest 978-1-4799-2349-6/16/$31.00 ©2016 IEEE 767