Information processing in biomedical applications

Abstract

To classify biomedical data is to find a mapping from patterns to a set of classes (e.g., disease states). Patterns are represented by features (e.g., metabolite concentrations) and class labels are assigned using a reference test (e.g., an expert's analysis of "normality"). This process often suffers from three significant challenges: voluminous features; pattern paucity; and reference test imprecision. Three computational intelligence based techniques, which exploit the notion of information granulation, are presented to address these challenges. Fuzzy quantile encoding replaces a feature with its membership values in a fuzzy set collection describing the feature's interquantile range. Class label adjustment compensates for reference test imprecision by adjusting design set class labels using a fuzzified similarity measure based on robust measures of class location and dispersion. Stochastic feature selection is a strategy where instances of classifiers are presented with feature regions sampled from an ad hoc cumulative distribution function. These techniques as well as their application to several classification problems in the biomedical domain will be discussed. © 2009 Springer-Verlag Berlin Heidelberg.