The GLM-Spectrum: A multilevel framework for spectrum analysis with covariate and confound modelling

Abstract

Spectrum estimators that make use of averaging across time segments are ubiquitous across neuroscience. The core of this approach has not changed substantially since the 1960s, though many advances in the field of regression modelling and statistics have been made during this time. Here, we propose a new approach, the General Linear Model (GLM) Spectrum, which reframes time averaged spectral estimation as multiple regression. This brings several benefits, including the ability to do confound modelling, hierarchical modelling and significance testing via non-parametric statistics.We apply the approach to a first-level EEG resting-state dataset alternating between eyes open and eyes closed resting-state. The GLM-Spectrum can model both conditions, quantify their differences, and perform denoising through confound regression in a single step. This application is scaled up from a single channel to a whole-head recording and, finally, applied to quantify age differences across a large group-level dataset. We show that the GLM-Spectrum lends itself to rigorous modelling of within- and between-subject contrasts as well as their interactions, and that the use of model-projected spectra provides an intuitive visualisation. The GLM-Spectrum is a flexible framework for robust multi-level analysis of power spectra, with adaptive covariance and confound modelling.

Is high-frequency activity evidence of an anterior temporal lobe network or micro-saccades?

“…A parsimonious explanation is that these data derive entirely from eye muscle, some components of which were not picked up on the EOG recordings. The signals are not only high-frequency, but also broadbandconsistent with previous reports of muscle activity (Jerbi et al, 2009;Muthukumaraswamy, 2013) and a recent study which indicates a strong relationship between EEG data and EOG recordings high frequencies (Quinn et al, 2022).…”

“…A parsimonious explanation is that these data derive entirely from eye muscle, some components of which were not picked up on the EOG recordings. The signals are not only high-frequency, but also broadband – consistent with previous reports of muscle activity (Jerbi et al, 2009; Muthukumaraswamy, 2013) and a recent study which indicates a strong relationship between EEG data and EOG recordings high frequencies (Quinn et al, 2022). The location, at the anterior temporal pole, is also consistent with the location of the lateral rectus muscle, previously identified as a source of high-frequency electrical activity (Carl et al, 2012; Jerbi et al, 2009).…”

Comparison between EEG and MEG of static and dynamic resting-state networks

“…Group-averaged values of each network feature were simply taken from the regression estimates of the fitted GLMs. For a comprehensive overview of the GLM and its application to neuroimaging data, readers should consult [37].…”

Extensive MEG time-series phenotyping unveils neural markers predictive of age

“…We assessed if sex and total intracranial volume have potentially confounded the prediction analyses as these variables have contributed to variance in spectral MEG features in previous work (e.g. 3,4,76,77 ). We implemented a model-agnostic approach using the partial confounder test from the mlconfound package (https://mlconfound.readthedocs.io) that tests the conditional (in-)dependence between the predicted values and each potential confounder 41,78 .…”