Abstract
The objective of this study is to compare the effectiveness of some standard feature reduction and classification techniques (including principal component analysis, PCA; multilayer perceptrons, MLPs; and nearest neighbor classifiers, k-NN) against several proposed variants for the analysis of high-dimensional biomedical spectral data. First, the original feature space is augmented by nonlinear transformations of the original features. We present an extension of sub-pattern PCA (SpPCA) proposed by Chen, which exploits sub-patterns (rather than complete original patterns) in the process of dimensionality reduction. Comprehensive experiments demonstrate the effectiveness of SpPCA over standard PCA. SpPCA leads to the development of individually reduced subspaces and, because of the local nature of this processing, the effectiveness of the reduction and classification processes may be enhanced. With respect to classifier topologies, we present and contrast two common classifiers, MLP and k-NN, as well as several fusion strategies. Finally, some general design guidelines are offered. The experimental framework uses biomedical data acquired from magnetic resonance spectrometers.