Objective: To test the influence of common electrophysiological preprocessing on the estimation of transfer entropy.

Background: Transfer entropy (TE) provides one possibility to study temporal causal relationships between affected brain regions and muscles by using a model-free implementation of Wiener’s causality principle. Beside its model-freeness TE has the advantage of inferring directed information flow from nonlinear systems. However, the influence of standard electrophysiological preprocessing like filtering or downsampling on the estimation of TE still remains elusive. For classical autoregressive implementations of Granger causality it is known that filtering has negative effects on the estimation of the causal relationships. Since it has been shown that TE and Granger causality are equivalent for jointly Gaussian distributed systems it is possible that similar negative effects arise in TE estimation.

Methods: In order to analyze the influence of preprocessing, different filter settings and downsampling factors were tested in a simulation framework. To account for different data types such as electroencephalography, local field potential or single-cell recordings, we used a linear model and two nonlinear models with sigmoid coupling functions.

Results: For nonlinear coupling and progressively lower filter cut-off frequencies (320 Hz, 160 Hz, 80 Hz), a monotonic increase from 0 % to up to 72 % false negative connections and up to 26 % false positive connections was observed. In contrast, for the linear model only for the 160 Hz low-pass filter up to 1.7 % false negative and 14.4 % false positive connections were found. Independent of filtering or model type, the modeled interaction delays were robustly estimated, with the highest delay deviation being approximately three samples. If the downsampling factor is so large that the resulting sampling period exceeds the highest interaction delay, an increase of up to 74 % false negatives was observed.

Conclusions: TE is highly suited to complement model based causality measures, especially when inferring directed information flow in nonlinearly coupled neural networks. However, if the type of network dynamics is unknown, filtering and downsampling are not advised when estimating TE.

« Back to 2016 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/how-does-preprocessing-affect-estimation-of-causal-information-transfer-in-the-human-brain/