| Summary: | Electroencephalography (EEG) is noninvasive and it is an effective tool to understand the complex nature of the brain. Its analysis by visual inspection is tedious and costly, and that is why many researchers in recent years resort to the use of computer-assisted diagnosis methods. In this work, we propose a new basis of wavelets constructed with Laguerre polynomials inspired by the similarities and differences between the Mexican hat wavelets family and the Gaussian wavelets family. The constructed wavelets are applied in epileptic seizure detection of EEG signals with the help of artificial neural networks and support vector machines for classification. We use a benchmark database, which consists of five hundred signals divided into five different classes, to solve eight clinically relevant classification problems with temporal and spectral features extracted from a continuous wavelet transform (CWT) of the EEG signals and the constructed wavelets. The spectral representation shows an intensive intermediary frequency (10–30 Hz) activity for seizure signals, while the other signals are characterized by a small high frequency (30–40 Hz) activity. We verify that, for any EEG classification problem, the model made up of CWT-PCA-SVM with a quadratic kernel is a classic case of overfitting irrespective of the wavelets used. The results obtained with wavelets constructed by Laguerre polynomials with a three-, five- and ten-fold cross-validation are better than several state-of-the-art classification methods in terms of classification accuracy. Keywords: Wavelets, Laguerre polynomials, Epileptic seizure detection, Pattern recognition neural networks, Support vector machines
|
|---|
