Results

The 5 seconds, quasi-stationary epochs of different portions of the intracranial EEG recordings of complex partial seizure from patients undergoing monitoring with subdural grid arrays were analyzed Fig 1. Matching pursuit time-frequency decomposition of residuals, after subtracting AR portion of the signal, showed much more transient components of the higher frequencies than in the raw data analyses. As an illustration of the method analyses for one 5 sec epoch are shown. Fig. 2 shows a 5 sec epoch of intracranial recording from one grid electrode during seizure initiation. The high amplitude low frequency component as well as high frequency low amplitude component are visible. Fig. 3 shows a three dimensional representation of the time-frequency distribution of the energy of this signal. The Matching Pursuit algorithm decomposed this epoch into 164 Gabor waveforms representing 99.8 % of energy of the signal. The remaining 0.2 % was discarded by the algorithm as not coherent with the signal. The decomposition is dominated by high energy low frequency components with a relatively small number of higher frequency components. Fig. 4 shows residuals left after fitting the AR model to the signal shown in Fig. 2. Fig. 5 shows results of MP analysis of these residuals. In this case the MP algorithm decomposed signal into 103 coherent Gabor waveforms representing 88.5 % of total energy of the residuals. The pattern of time-frequency distribution is different than in Fig. 3. The waveforms are more evenly distributed on the time-frequency plane and more high frequency components are visible. This allows for analysis of patterns of low amplitude high frequency components and possible correlation of these patterns with space-time properties of seizure dynamics.