Information Technology Reference
In-Depth Information
A generic overview of the proposed methodology emphasizing the various statis-
tical approaches is illustrated in Fig. 3.1. Different statistical decisions are possible
according to the profile of the data under examination. The first choice is based on
whether the data are normally distributed, whereas the second is based on the num-
ber of different groups under examination - i.e., whether two or more classes (tasks)
are being tested. A detailed view of feature selection and refinement blocks match-
ing our data characteristics is presented in Fig. 3.2. The steps involved, as well as
their implementation issues, are analyzed in the following sections.
Fig. 3.1: The proposed methodology uses significance-based statistics to reduce the
dimensionality of the problem and select the most salient and descriptive feature
vectors. Different statistical decisions are possible according to the profile of the
data under examination. If one is interested in discriminating two or more classes of
normally distributed data,
t
-test or analysis of variance (ANOVA) tests are appropri-
ate candidates, respectively. If the data is nonnormally distributed, Mann-Whitney
and Kruskal-Walls tests are the alternatives.
3.2.2 Feature Extraction (Step 1)
Over the past decade the WT has developed into an important tool for analysis of
time series that contain nonstationary power at many different frequencies (such
as the EEG signal), as well as a powerful feature extraction method [9]. There are
several types of wavelet transforms, namely the discrete (DWT) and the continuous
(CWT), which involve the use of orthogonal bases or even nonorthogonal wavelet
functions, respectively [8]. CWT is preferred in this approach, so that the time and
scale parameters can be considered as continuous variables. In the WT, the notion
