Estimation of temporal scales of variation in long-term scalp electroencephalograms from epilepsy patients

Long-term neurophysiological recordings, such as scalp encephalograms (EEG), have been routinely used in studies that aim to characterize dynamic changes in brain activity associated with normal biological processes, such as sleep, but are also becoming increasingly common for clinical evaluation of patients with neurological disorders, such as epilepsy. Analysis of non-stationary recordings from multiple days poses new signal processing challenges, in regard to algorithm efficiency and computational cost, as well as adequate dimensionality data reduction. We compared four approaches for estimating the underlying temporal dynamics of long-term recordings from patients with medically refractory epilepsy: (i) model order selection using the minimum description length principle, (ii) approximate entropy, (iii) mutual information, and (iv) Detrended Fluctuation Analysis (DFA). Individual approaches were found to be sensitive only to specific scales of variation. Approximate entropy and mutual information were sensitive to local dynamics, whereas dynamic model order estimation captured only slowly varying dynamics. DFA was sensitive to multiple temporal scales.