Time-frequency structure of epileptic seizures

Occurrence of epileptic seizures is usually visible in raw EEG, especially if recorded from intracranial electrodes (ECoG). A detailed time-frequency description can serve as a basis for their classification and for identification of different dynamic states during evolution and propagation of seizures, which was demonstrated in [6] by means of MP estimates.

Figure 5: Bottom: intracranial recording of an epileptic seizure, horizontal axis in seconds, sampling 100 Hz. Middle time-frequency distribution of energy was obtained as follows: signal was decomposed by MP with 17 different realizations of a stochastic dictionary ($3*10^6$ atoms), and resulting energy distributions were averaged. Picture illustrates different dynamical states identified in [6]: near 15th second, the period of seizure initiation transforms into transitional rhythmic activity, then before 30th second the organized rhythmic activity with 8 visible harmonics becomes dominant, to dissolve into intermittent bursting activity near 90th second. Data and interpretation courtesy of prof. P. J. Franaszczuk and prof. G. K. Bergey. Top--the same distribution in 3 dimensions; presentation as in Figure 2.

Figure 5 presents one recording from [6] and time-frequency distribution of its energy, obtained by averaging decompositions in different realizations of stochastic dictionary (this procedure is presented on simulated signal in Figure 2 plots g and h). Stochastic dictionaries were not available when the original paper ([6]) was published. Their application allowed for detection of higher number of harmonics and smoother representation of changing frequencies.