Detecting hyperventilation through efficient ARMA modeling of ECoG data of epileptic patients

In the context of epileptic seizure prediction-a task strongly pursued in recent years-it is important to clearly define a pre-ictal (pre-seizure) state so as to allow an alarm signal to be given to the patient. Currently the most reliable indicator for pre-ictal states is the correlation dimension calculated from recordings of the Electrocorticogram (ECoG) which, as reported by Lehnertz et al. (1998), drops down up to 20 minutes prior to the seizure onset. However, this is not only the case prior to a seizure but for example also during hyperventilation. In this respect auto-regressive-moving-average (ARMA) modeling can serve as a discriminating measure so as to not confuse hyperventilation with pre-ictal states. The approach taken in this paper determines a multivariate parsimonious ARMA model by extracting the prevalent modes related to the spectral peaks of a standard AR model for the given multichannel data. Several signal processing spectral smoothing procedures are adopted in dealing with the patient data. Measurements from interhippocampal depth electrodes as well as subdural strip electrodes in the temporal lobe, from both the left and right hemisphere are considered