Uso da análise de discriminante linear em conjunto com a transformada wavelet discreta no reconhecimento de espículas

Abstract

Researchers have concentrated efforts in the past 20 years, by applying the wavelet transform in processing, filtering, pattern recognition and classification of biomedical signals, in particular signals of electroencephalogram (EEG) containing events characteristic of epilepsy, the spike. Several families of mother-wavelets were used, but there are no consensus about which mother-wavelet is the most adequate for this purpose. The signals used have a wide range of events. The literature reports EEG signals sampled from 100 to 600 Hz with spikes ranging from 20 to 200 ms. In this study we used 98 wavelets. The EEG signals were sampled from 200 Hz up to 1 kHz. A neurologist has scored a set of 494 spikes and a set 1500 non-spike events. This study starts evaluating the amount of wavelet decompositions required for the detection of spikes, followed by detailed analysis of the combined use of mother-wavelets of the same family and among families. Following is analyzed the influence of descriptors and the combined use of them in spike detection. The results of these studies indicate that it is more appropriate to use a set of mother-wavelets, with many levels of decomposition and with various descriptors, instead of using a single mother-wavelet or a specific descriptor for the detection of spikes. The selection of this set of wavelets, decomposition level and descriptors allows to obtain high levels of detection according to the computational load desired or computing platform available for implementation. This study reached performance levels between 0.9936 to 0.9999, depending on the computational load. Other contributions of this study refer to the analysis of the border extension methods for spike detection; and the influences of the EEG signal sampling rate in the classifier performance, each one with significant results. Also shown are: a new spike detection architecture by making use of linear discriminant analysis; and the presentation of a new descriptor, the centred energy, based on the response of the coefficients of decomposition levels of the wavelet transform, able to improve the discrimination of spike and non-spike events.