Analysis of wavelet-filtered tonic-clonic electroencephalogram recordings \ O. A. Rosso I A. Figliola I J. Creso I E. Serrano 2 llnstituto de C&lculo, Facultad de Ciencias Exactas y Naturales, Universidad de Buenos Aires, Argentina 2Departamento de Matem~ticas, Facultad de Ciencias Exactas y Naturales, Universidad de Buenos Aires, Argentina Abstract--EEG signals obtained during tonic-c/onic epileptic seizures can be severely contaminated by muscle and physiological noise. Heavily contaminated EEG signals are hard to analyse quantitatively and also are usually rejected for visual inspection by physicians, resulting in a considerable loss of collected information. The aim of this work was to develop a computer-based method of time series analysis for such EEGs. A method is presented for filtering those frequencies associated with muscle activity using a wavelet transform. One of the advantages of this method over traditional filtering is that wavelet filtering of some frequency bands does not modify the pattern of the remaining ones. In consequence, the dynamics associated with them do not change. After generation of a "noise free" signal by removal of the muscle artifacts using wavelets, a dynamic analysis was performed using non-linear dynamics metric tools. The characteristic parameters evaluated (correlation dimen- sion D 2 and largest Lyapunov exponent 21) were compatible with those obtained in previous works. The average values obtained were: D2-4.25 and 21-3.27 for the pre-ictal stage; D2-4.03 and 21-2.68 for the tonic seizure stage; D2-4.11 and 21-2.46 for the clonic seizure stage. Keywords--EEG, Epileptic seizures, Wavelet analysis, Non-linear dynamics metric tools Med. Biol. Eng. Comput., 2004, 42, 516-523 J 1 Introduction THE ELECTRO-ENCEPHALOGRAM (EEG) can be roughly defined as the mean electrical activity of the brain at different sites on the head. Using an empirical basis, EEG patterns are correlated with functions, dysfunctions and diseases of the central nervous system. Although visual inspection of the EEG is quite useful, it remains subjective and hardly allows any systematisation. To overcome this fact, quantitative EEG (qEEG) analysis provides objective measures that can reflect important characteristics of the brain activity (NIEDERMEYER and LOPES DA SILVA, 1982). A scalp EEG signal is essentially a non-stationary time series that presents artifacts mainly due to eye movements, muscle and heart activity. Artifacts make it difficult to analyse scalp EEG signals using mathematical tools. Sometimes, artifacts are present only for a few seconds, and the small contaminated epoch can be neglected without much information about the signal being lost. In other cases, however, they conceal almost the whole signal, and no information about the underlying brain activity can be extracted. A tonic-clonic (TC) epileptic seizure is characterised by violent muscle contractions. Massive initial tonic spasms are followed some seconds later by a clonic phase with violent flexor Correspondence should be addressed to Dr Osvaldo A. Rosso; emaih oarosso@fibertel.com.ar Paper received 20 October 2003 and in final form 3 March 2004 MBEC online number: 20043900 © IFMBE: 2004 516 movements and characteristic rhythmic spasms that last until the end of the seizure. During these seizures, artifacts related to muscle contractions are especially troublesome, because they reach very high amplitudes (NIEDERMEYER and LOPES DA SILVA, 1982). in fact, they limit the traditional visual analysis to the pre- and post-ictal periods and, furthermore, they restrict the application of some mathematical methods. Analysis of the brain activity during these seizures has previously been performed only in special cases, for example, in patients treated with curare (an inhibitor of the muscle responses) (GASTAUT and BROUGHTON, 1972) or by eliminating the high- frequency muscle activity by the use of traditional filters (GOTMAN et al., 1981). However, traditional filtering methods (bandpass filters based on Fourier transform) have several disadvantages: they can introduce spurious oscillations and ringing effects and, more importantly, they can also affect the morphology of the remaining signal (QUIAN QUIROGA et al., 2001). One of the new approaches to qEEG analysis is the application of concepts related to non-linear dynamics theory. Since the pioneering work of BABLOYANTZ et al. (1985), many researchers have begun to study the dynamics of brain activity by reconstructing the orbits of the system in the phase space and measuring non-traditional parameters, such as the correlation dimension D2, the largest Lyapunov exponent 21 or Kolmogorov entropy K 2. These non-linear metric invariants proved to be very useful in the characterisation of the brain dynamics in different states (BA~AR and BULLOCK, 1989; BA~AR, 1990; ELBERT et aL, 1994; PRITCHARDand DUKE, 1995). For example, using EEG recordings from surgical implanted electrodes, LEHNERTZand Medical & Biological Engineering & Computing 2004, Vol. 42