Predicting Working Memory performance based on specific individual EEG spatiotemporal features

Abstract

Working Memory (WM) is a limited capacity system for storing and processing information, which varies from subject to subject. Several works show the ability to predict the performance of WM with machine learning (ML) methods, and although good prediction results are obtained in these works, ignoring the intersubject variability and the temporal and spatial characterization in a WM task to improve the prediction in each subject. In this paper, we take advantage of the spectral properties of WM to characterize the individual differences in visual WM capacity and predict the subjects performance. Feature selection was implemented through the selection of electrodes making use of methods to treat unbalanced classes. The results show a correlation between the accuracy achieved with an Regularized Linear Discriminant Analysis (RLDA) classifier using the power spectrum of the EEG signal and the accuracy achieved by each subject in the behavioral experiment response of a WM task with retro-cue. The proposed methodology allows identifying spatial and temporal characteristics in the WM performance in each subject. Our methodology shows that it is possible to predict the WM performance in each subject. Finally, our results showed that by knowing the spatiotemporal characteristics that predict WM performance, it is possible to customize a WM task and optimize the use of electrodes for agile processing adapted to a specific subject. Thus, we pave the way for implementing neurofeedback through a Brain-Computer Interface.