Evaluation of various classifiers performance on biomedical datasets

Often, an evaluation of a classifier is performed without deeper analysis. In this paper we decided to perform more rigorous evaluation. We present an evaluation of various classifier methods over biomedical data with orientation towards nature inspired methods. We have performed an experimental assessment of various traditional and nature inspired methods (41 distinct classifiers) over the total of 32 different biomedical datasets. We used 10-fold crossvalidation and for each experiment retrieved multiple objective parameters. The mean and best/worst-so-far values of the measures (accuracy, sensitivity, specificity, ...) have been statistically evaluated using the nonparametric Friedman test and post-hoc analyses. The ant-inspired ACO_DTree algorithm performed significantly better (alpha=0.05) in 29 experimental cases for the mean f-measure parameter and in 14 experimental cases for the best-so-far f-measure parameter. The top results have been obtained for certain subsets of the UCI database and for the dataset combining cardiotocography records and myocardial infarction records.