Window size analysis in ARMA-modeled spectral biomarkers for epileptic children

Abstract

Autoregressive Moving Average (ARMA) models are suitable for modeling processes whose frequency spectrum exhibits both sharp peaks and deep nulls. In this analysis an ARMA model was used to model EEG data and estimate its time-frequency spectrum. Spectral features were then extracted and their suitability as biomarkers for epileptic children whose EEG is clinically diagnosed as normal is analyzed. Furthermore, variations in classification are investigated when features are extracted from different window sizes. Results show that through windowing, a maximum classification score of 97.9% is achieved with an improvement of up to 18.3% over the non-windowed case.