Ponente
Descripción
The study of event-related potentials (ERP) has become increasingly popular in the last decades due to its excellent temporal resolution. Still, the waveforms extracted from these paradigms are excessively entangled in the time domain, which causes that the use of deconvolution methods be slow or difficult to converge. The use of techniques that assume variability in the response across subjects, like the linear mixed effect models (LME), offers many improvements for EEG modeling due to its higher power to detect variability across subjects and groups. Still, its use in real data requires deconvolution methods for continuous EEG data, which have a high computational cost.
We develop method and software for efficient and realistic overlap correction combined with LME for the deconvolution of ERPs. Our approach is based on the Expectation-Maximization (EM) algorithm, which is conceptually simple and has a fast initial linear convergence algorithm.
We use the EM algorithm to explore the performance of two different types of models with different levels of complexity. i) The simplest model assumes the average group components as fixed effects with a single random effect absorbing all variability. Therefore, it is an LMM effect with deconvolution; ii) A more refined model retains the fixed effect with deconvolution but introduces intersubject variability of components as a direct random effect. The smoothness of the fixed and random ERP components is controlled with a penalty function. Additionally, we explicitly model individual spontaneous background EEG by adding another random effect. Thus, we deal with the dual random effects (LMM-DRE) model.
We introduce several technical innovations to speed up convergence and reduce run times to evaluate these two models via simulations. We use proximal operators for the calculus of the initial approximation of the EM and estimate the background EEG activity for each individual using the cosine discrete transform.
We used the simpler convolutional LMM to study ERP data from the Barbados Nutrition Study (BNS). The BNS is a 50-year-long longitudinal study of the effects, only during the first year of life, of Protein Energy Malnutrition (PEM). The ERP experiment was a Go/No Go inhibition task, an experiment designed to detect the long-term effects of PEM on cognitive processes such as attention. The deconvolution LMM took out overlapping responses to the stimuli, motor responses, and endogenous components linked to cognitive processes specific to each group showing differences between the PEM and control group, confirming other results from our group on the long-term effects of malnutrition.
