CNS 2014 Abstract: Nikolaidis, A., Reggente, N. et al
Using the the task-dependent functional connectivity across nodes within established cortical network from both before and after videogame training, we were able to asses the “plasticity” of these networks as a function of training. The plasticity of these networks, combined with graph theoretical metrics were used as features in a leave-one-out Ridge Regression that was able to account for upwards of 80% of the variance in individual difference scores in learning. Each network contributed varying levels of accuracy to classification depending on its involvement in the subject’s instructed priorities within the task. For example, plasticity in the Cingulo-Opercular network preferentially predicted(upwards of 55% of variance accounted for) learning in a training strategy that relied more heavily on executive control of attention and goal directed behavior.
Predicting individual differences in cognitive gains from videogame training using machine learning analyses of fMRI functional connectivity patterns
Aki Nikolaidis1, Nicco Reggente2, Drew Goatz3, Kathryn Hurley3, Andrew Westphal2, Arthur F. Kramer1; 1University of Illinois, Urbana Champaign, Beckman Institute, 2University of California, Los Angeles, Psychology Department, 3University of Illinois, Bioengineering Department
One of the important questions in cognitive training, and learning and memory more broadly, is how pre-existing individual differences in brain connectivity influence the effect of training. In this study, we use the fMRI functional connectivity of multiple networks, including the frontal-parietal and motor networks, to predict individual differences in learning over the course of 30 hours of cognitive training with the Space Fortress videogame. We used various metrics of functional connectivity and graph theory-derived parameters from 45 young adult participants as features to train adaptive multivariate regression models. Using a leave-one-participant-out cross-validation procedure, we find that we can predict a significant percentage of the variance in learning performance (defined as pre-post differences in Space Fortress score). By analyzing the performance of different regression models, we find that distinct brain networks contain different types of information regarding individual differences in learning rate. Furthermore, using both support vector regression and ridge regression we demonstrate how different feature and model parameters have important effects on model performance, and we consider how these parameters may have limited previous research using such techniques. We discuss implications of our results for cognitive training, as well as the continued use of machine learning and graph theoretical analyses in cognitive neuroscience.