Single trial classification of magnetoencephalographic recordings using Granger causality

The use of Granger causality (GC) for studying dependencies in neuroimaging data has recently been gaining popularity. Several frameworks exist for applying GC to neurophysiological questions but many rely heavily on specific statistical assumptions regarding autoregressive (AR) models for hypothesis testing. Since it is often difficult to satisfy these assumptions in practical settings, this study proposes an alternative statistical methodology based on the classification of individual trials of data. Instead of testing for significance using statistics based on estimated AR models or prediction errors, hypotheses were tested by determining whether or not individual magnetoencephalography (MEG) recording segments belonging to either of two experimental conditions can be successfully classified using features derived from AR and GC concepts. Using this novel approach, we show that bivariate temporal GC can be used to distinguish button presses based on whether they were experimentally forced or free. Additionally, the methodology was used to determine useful parameter settings for various steps of the analysis and this revealed surprising insight into several aspects of AR and GC analysis which, previously, could not be obtained in a comparable manner. A final mean accuracy of 79.2% was achieved for classifying forced and free button presses for 6 subjects suggesting that classification using GC features is a viable option for studying MEG signals and useful for evaluating the effectiveness of parameter variations in GC analysis.

► Classification of short MEG trials using Granger causality. ► Improved AR modeling of MEG data from successful classification. ► Testing basic preprocessing steps for Granger causality and AR analysis of MEG data.