A new strategy for model order identification and its application to transfer entropy for EEG signals analysis

The background objective of this study is to analyze electrenocephalographic (EEG) signals recorded with depth electrodes during seizures in patients with drug-resistant epilepsy. Usually, different phases are observed during the seizure evolution, including a fast onset activity. We aim to ascertain how cerebral structures get involved during this phase, in particular whether some structures "drive" other ones. Regarding a recent theoretical information measure, namely the transfer entropy (TE), we propose two criteria, the first one is based on Akaike's information criterion, the second on the Bayesian information criterion, to derive models' orders that constitute crucial parameters in the TE estimation. A normalized index, named partial transfer entropy (PTE), allows for quantifying the contribution or the influence of a signal to the global information flow between a pair of signals. Experiments are first conducted on linear autoregressive models, then on a physiology-based model, and finally on real intracerebral EEG epileptic signals to detect and identify directions of causal interdependence. Results support the relevance of the new measures for characterizing the information flow propagation whatever unidirectional or bidirectional interactions.