Abstract
Accurate classification of biomedical data is often confounded by potentially imprecise class labels assigned by an external reference test. We present a gradation method using fuzzy set theory and a dispersion-adjusted similarity measure to assign, for each pattern in a design set, a degree of belongingness to each class. After training a classifier using this adjusted design set, its performance is measured using a validation set of patterns with their original class labels. We empricially demonstrate the effectiveness of this method using three publicly available biomedical datasets. Using the same classifier, we benchmark the results against the original datasets without gradation.