Statistical Test-Based Feature Selection and Classification Techniques for Breast Cancer Data

Breast cancer is a disease that affects the majority of women and it is the second most common cause of death among women globally. Medical scientists have proven that there are a vast number of genes that are responsible for breast cancer. Among them, all genes are not equally responsible. Therefore, the most relevant and informative genes are needed to find out to control the disease. The objectives of our study are: (i) To find the most informative and significant genes using different statistical test-based feature selection techniques (FST) as well as find the best classifier and (ii) To validate our experimental results using a simulated dataset. The breast cancer dataset is a benchmark dataset provided by Kent Ridge Biomedical Data Repository, USA. In our study, we have used different statistical test-based feature selection techniques such as the t-test and Wilcoxon signed rank sum (WCSRS) test. Naïve Bayes (NB), Adaboost (AB), linear discriminant analysis (LDA), artificial neural network (ANN), k-nearest neighbor (KNN), and random forest (RF) are treated as classification techniques. Our analysis included 24,188 genes and 97 patients. Among them, 46 patients were with cancer and 51 were in control. We considered 70% of the dataset as a training set and the rest is a test set and repeated this procedure about 1000 times. Among all the combinations of FST and classification techniques t-test-based Naive Bayes classifier gives us the highest classification accuracy. The analysis of our study indicates that the integration of t-test-based FST and Naïve Bayes classifier produces the maximum classification accuracy.