Investigation of improved evolutionary feature selection techniques for biomedical applications

Abstract

In the hypothesis of pattern recognition, the classification of medical datasets is becoming a challenging task due to the large number of features and training data limitations. Redundancies present in these irrelevant features affect the overall classification accuracy. Such insurmountable issues to classification can be overcome by Feature Selection (FS) techniques performed through combinatorial optimization methods such as Genetic Algorithm (GA). However, there are many limitations in basic GA like premature convergence, low degree of solution accuracy due to fixed crossover and mutation rates, low convergence capacity and lack of population diversity. These limitations restrict the genetic search from reaching global optimal solution and thereby reducing their efficiency as a FS method. To overcome these issues, this thesis suggests an effective wrapper framework with some acceptable solutions i.e. firstly, by adding fitness scaling technique (called sigma scaling) along with the Stochastic Universal Sampling (SUS) selection function. This scaling technique has enabled the genetic search to re-adjust the fitness values of the population, which has prevented GA from reaching premature convergence. Secondly, the degree of solution accuracy is improved by adaptively changing the crossover and mutation rates based on the population fitness. Further, the masking concepts of crossover and mutation enhanced the population diversity and increased the convergence capacity as well. To measure the quality of the chromosomes, three different objective functions namely Classification Accuracy, Geometric Mean and a weighted aggregation of Geometric mean and sum of selected features have been employed. The classifiers employed along with ten-fold cross validation method have served as an evaluator of the proposed Improved Genetic Algorithm (IGA) algorithm. The experimental results are presented in terms of several performance measures like Positive prediction, Negative prediction, Sensitivity, Specificity, Accuracy, F-measure, AUC, Kappa statistic and G-mean. Through the proposed IGA method, promising classification accuracy has been obtained for all the six benchmark datasets i.e. 100% for MEEI dataset, 99.49% for PD dataset, 84% for CAD dataset, 99.24% for ES dataset, 94.34% for BT dataset & 94% for CTG dataset. Also, a reduced feature subset has been attained for all these datasets as well i.e. 8 features for MEEI & PD dataset, 3 out of 9 features for CAD dataset, 14 out of 34 features for ES dataset, 3 out of 13 features for BT dataset and 6 out of 22 features for CTG dataset has been obtained espectively. On the whole, experimental results show that the proposed algorithm has evolved a feature subset with a smaller number of features and higher classification performance than using all the features. All the algorithms used in these experiments were simulated using MATLAB 2011a.