A wrapper-based feature selection approach to investigate potential biomarkers for early detection of breast cancer

Breast cancer (BC) biomarkers can radically improve the early detection in patients and, as a result, reduce mortality rate, whether for detecting individuals at increased risk of developing cancer or in the screening process. Finding a successful biomarker for breast cancer would be a fast and low-cost first solution to predicting BC, and it could potentially lead to a decline in the global BC mortality rate. However, biomarker exploration translates into the role of feature ranking and selection in machine learning terminology. This study explores the influence of using a particular biomarker or combinations of different biomarkers as predictors for breast cancer. Three different classification algorithms were integrated with a sequential backward selection model: support vector machine (SVM), random forests (RF), and Decision Trees (DTs). The result shows that the optimal set of biomarkers comprises Glucose, Resistin, homo, BMI, and Age using the SVM model. The sensitivity and specificity were 0.94 and 0.90, respectively and the 95% confidence interval for the AUC was [0.89, 0.98]. The result indicates that Glucose, Resistin, homo, BMI, and Age combined can serve as a crucial BC biomarker in BC screening and detection.