Abstract : Electroencephalographic (EEG) recording provides a powerful measure of neural dynamics underlying human cognition, such as working memory. However, the analysis of multidimen-sional EEG data is challenging because it requires the modeling of temporal and spatial correlations in order to determine the EEG features most predictive of memory performance. Standard techniques, such as generalized estimating equations (GEE), can select features and account for sample correlation. However, they cannot explicitly model how a dependent variable relies on features measured at different memory stages and scalp locations. We propose an approach to automatically and simultaneously determine both the relevant spatial features and relevant temporal points that impact the response of a memory task. The proposed model can still correct for the non-i.i.d nature of the data, similar to GEE, by estimating the within-individual correlations. Our approach decomposes model parameters into a summation of two components and imposes separate block-wise LASSO penalties to each component when building a linear model in terms of multidimensional EEG features. An accelerated gradient descent algorithm is developed to efficiently solve the related optimization problem. We identified that the influential factors for working memory between healthy subjects and schizophrenia patients differ in frequency bands, scalp positions and information processing stages.
