Using independent component analysis & dynamical embedding to isolate seizure activity in the EEG

Abstract

We present a methodology for isolating the underlying seizure activity in multichannel scalp EEG. A number of seizure segments from various patients are extracted from the fetal EEG recorded in an epilepsy monitoring unit. We use the method of independent component analysis (ICA) in order to decompose the recorded scalp EEG into its underlying temporal and spatial components. Seizure-related activity in the independent components is identified by first performing a dynamical embedding on each component. Through a value linked to the dynamic complexity of the EEG segments it is possible to visually analyze the spatiotemporal components. In this proof of principle study, in 4 seizure EEGs analyzed we can identify what appear to be the relevant seizure components-there is more than one in each case. We identify seizure related activity by choosing those temporal components that depict a decrease in complexity around seizure onset, coupled with focal activity in the region grossly identified as being the origin from the raw scalp EEG. In addition to seizure related components, artifactual components are also adequately isolated by ICA. Decomposing the EEG in this way means that the scalp EEG can either be remapped using only the identified seizure components, or further in-depth analysis on the seizure can be undertaken on the spatiotemporal components directly. Although subjective, these preliminary results indicate that ICA coupled with complexity analysis may be beneficial in processing the epileptiform EEG prior to further in-depth analysis.