Application of Berkner transform to the detection and the classification of transients in EMG signals Berkner Decomposition and classification results Sylvain Meignen, Pierre-yves Guméry and Valérie Perrier LMC-IMAG Laboratory, Mosaic team and TIMC-IMAG Laboratory Introduction Berkner transform [1] is an approximation of the continuous wavelet transform which preserves most of its characteristics. We use it to build a model for the detection and the classification of non stationary events in EMG signals of the genioglossus muscle. We finally show an illustration of the detection on a real EMG signal. Conclusion This poster shows how to use maxima lines made by the extrema of the Berkner transform to build an efficient transient detector for EMG signals of the genioglossus muscle. The particularity of the method over existing detectors using wavelets is that the time information in well preserved through the use of maxima lines. References 1. K. Berkner, R.O. Wells, A new hierarchical scheme for approximating the continuous wavelet transform. 2. L. Senhadji, J.J. Bellanger, G. Carraut, Détection temps échelle d’événements Paroxytiques intercritique en encéphalogramme, Traitement du signal, Vol. 12, #4, 1995, p S. Meignen, P-Y Guméry, Application de la transformée de Berkner à la détection et à la classification de transitoires dans les EMG réflexes, in review Detection of transients with maxima lines Presentation of the signals We present a general model for signals definition : where is the background activity, are useful transients that we wish to detect whereas the remainder are artefacts. Reconstruction with extremaExamples of maxima lines of Berkner Decomposition Properties of binomial sequences We use such a discrete filter translated and divided by 2 to get a sequence that corresponds to the filtering of a discrete signal f. Such a sequence enables perfect reconstruct of any signals of. We have shown [3] that in the case r = 1, we have a very simple reconstruction formula : These coefficients are chi square distributed with degrees of freedom that are functions of p [2]. In practice, these degrees of freedom are estimated over parts of the signal that do not contain useful transients but maybe artefacts. The details of the construction is given in [3]. It is worth noting that not only this detector separates background activity from transients, but also naturally eliminates artefacts associated to low frequencies. Once we have separated the background activity from the transients, we discriminate useful transients from the remaining artefacts as the latter are isolated whereas the former are associated to a train of transients. This is measured by the proportion of points that are transients in the neighbourhood of each transient. This approach is relevant since, in our case, the transients cannot be separated from the artefacts on a frequency basis. Moreover, the advantage of this method over existing methods is that as even if we collect information at different scales as we use maxima lines the precise time location of the useful transients is preserved. As a result of the central limit theorem, this sequence converges to the normal distribution when n tends to infinity as follows : We here admit that it is possible to build a sequence that approaches the derivative of order r of the Gaussian function. It is proved [1] that such a sequence satisfies : Transients detection algorithm We make the hypothesis that different kinds of transients do not appear simultaneously. It has been shown that the extrema of the sequence make up so called maxima lines when increases. These maxima lines contain useful information about the high frequency content of the signal, which is grouped into the variable : The extrema of the sequence are representative of the high frequency components of the signal. If we only use them to reconstruct the signal, we notice that the quadratic reconstruction error is driven by the extrema for small n (see reconstruction with extrema).