Time series analysis of magnetoencephalographic data during copying

We used standard time series modeling to analyze magnetoencephalographic (MEG) data acquired during three tasks. Each task lasted 45 s, for a total data acquisition period of 135 s. Ten healthy human subjects fixated their eyes on a central blue point for 45 s (fixation only, "F" task). Then a pentagon (visual template) appeared surrounding the fixation point which simultaneously became red (fixation + template, "FT" task). After 45 s, the fixation point changed to green, which was the "go" signal for the subjects to begin continuously copying the pentagon for 45 s using a joystick and without visual feedback of their movement trajectory (fixation + template + copying, "FTC" task). MEG data were acquired continuously from 248 axial gradiometers at a sampling rate of 1017.25 Hz. After removal of cardiac artifacts and rejection of records with eyeblink artifacts, a Box-Jenkins autoregressive integrative moving average (ARIMA) analysis was applied to the unsmoothed, unaveraged MEG time series for model identification and estimation within 25 time lags (approximately 25 ms). We found that an ARIMA model of 25th order autoregressive, first order differencing, and first order moving average (p=25, d=1, q=1) adequately modeled the series and yielded residuals practically stationary with respect to their mean, variance, and autocorrelation structure. These "prewhitened" residuals were then used for assessing pairwise associations between series using crosscorrelation analysis with +/-25 time lags (approximately +/-25 ms). The cross-correlograms thus obtained revealed rich and consistent patterns of interactions between series with respect to positive and/or negative correlations. The overall prevalence of these patterns was very similar in the three tasks used, and, for particular sensor pairs, they tended to be preserved across tasks.