Feature selection using an ensemble of optimal wavelet packet and learning machine: Application to MER signals

This paper present a novel feature selection method based on a ensemble of optimal wavelet packet and learning machine (OWLM). First, the lifting schemes (LS) are used to construct a single decomposition stage of the wavelet transform (WT), these schemes are a fast and flexible implementation for the WT. Then the update and predictor operator of the LS are optimized by Genetic Algorithms (GA) and Lagrange's optimization. So far the ensemble of the LS and the GA is called optimal wavelet transform (OWT) and it is a signal-adapted decomposition. Then cascaded OWT's are used to implement the wavelet packet transform (WPT) which provides a tool for customizing the time-frequency plane were the signal's energy is projected. Finally the OWLM is constructed by the optimization of the time-frequency plane of the WPT by maximizing the class separability using the Davies-Bouldin (DB) criterion. Then energy features are computed from the OWLM space, also a principal component analysis (PCA) is computed in order to reduce complexity in the OWLM space. Three different basic classifiers are used to evaluate the performance of the proposed method including the linear Bayesian classifier, the k-NN classifier, and the quadratic Bayesian classifier. As example the OWLM is used over a microelectrode recording (MER) signals database from four different brain zones. This signals have strong non-stationary behavior and different techniques have been used to extract discriminant information from it. Based on experiments carried out, results show that the OWLM is superior to other commonly used features selection methods, such as the WT and the WPT. The best MER classification rate 99.14% is obtained by the OWLM in the database used.