A New Filter Approach to Extract Relevant Features from Mass Spectrum Datasets

We propose an approach to extract relevant features from SELDI-TOF mass spectrum datasets. The proposed method can deal with both two-class and multiple-class problems. In the method, the relevance value of a feature representing how well the value of a feature helps to separate a sample from a given class was defined based on the difference between the numbers of samples in the given class with greater and less feature value than the sample. Using the relevance value as a basic factor, several ranked feature lists were established. Searching strategies to obtain optimal feature sets were also proposed by utilizing the relevance indices of features without using learning algorithms. The new method was applied to the three public mass spectrum datasets and showed better or comparable results than conventional filter methods.