Assessment of changing interdependencies between human electroencephalograms using nonlinear methods

We investigate the problems that might arise when two recently developed methods for detecting interdependencies between time series using state space embedding are applied to signals of different complexity. With this aim, these methods were used to assess the interdependencies between two electroencephalographic channels from 10 adult human subjects during different vigilance states.

The significance and nature of the measured interdependencies were checked by comparing the results of the original data with those of different types of surrogates. We found that even with proper reconstructions of the dynamics of the time series, both methods may give wrong statistical evidence of decreasing interdependencies during deep sleep due to changes in the complexity of each individual channel.

The main factor responsible for this result was the use of an insufficient number of neighbors in the calculations. Once this problem was surmounted, both methods showed the existence of a significant relationship between the channels which was mostly of linear type and increased from awake to slow wave sleep.

We conclude that the significance of the qualitative results provided for both methods must be carefully tested before drawing any conclusion about the implications of such results.