Classifying breast invasive carcinoma subtypes: A feature selection-based machine learning approach

Breast invasive carcinoma (BRCA) is one of the most common breast cancer types with multiple distinct subtypes, each with a distinguished morphological structure. Identifying discriminating genes to classify the subtypes plays an essential role upon the pathogenesis and treatment of carcinomas. In the present work, we propose a machine learning framework to classify BRCA subtypes based on the gene expression profiles of differentially expressed (DE) genes identified from TCGA BRCA data. We have utilized several feature selection algorithms to discover numerous key discriminating DE genes to classify BRCA subtypes through different classification models. The efficacy of the classification framework is evaluated through several performance metrics like accuracy, Matthews correlation coefficient (MCC), logistic loss, precision, recall, and f1-score. Results show that the JMIM feature selection algorithm, in conjunction with the SVM-RBF and the soft voting classifiers, outperforms the other algorithms with high precision. Moreover, we have identified several hub genes most significantly associated with patients' survival in BRCA.