Application of Machine Learning Models to the Detection of Breast Cancer

Abstract

This work aims to build a binary breast cancer classifier algorithm based on the blood test and anthropometric data (age, body mass index, glucose, insulin, homeostasis model assessment, leptin, adiponectin, resistin, and monocyte chemotactic protein-1) of 116 subjects. For this study, a performance comparison of the following machine learning models was performed: decision tree, random forest, K-nearest neighbors, artificial neural networks, vector machines of support, and logistical regression. The methodologies used in the data were as follows: k‐fold cross‐validation (k = 10); splitting data into 80% training and 20% testing. For the first, the mean of accuracy and sensitivity were evaluated in the second, values of accuracy, sensitivity, specificity, and area under some tests. In addition, most mammograms are performed on benign tumors. With this, it is clear that these exams can use other tools to assist in decision-making, and machine learning can offer great utility and good cost/benefit in the diagnostic process of breast cancer. Many research papers for breast cancer biomarkers have been reported over the years. The present work will analyze the potential quantitative variables: age, receiver operating characteristic curve. Furthermore, the p value, Pearson correlation coefficient, and, depending on the input variable, the test only with variables with a significance threshold of 5% are computed from the normal distribution assessment (calculated from Kolmogorov–Smirnov test (KS test)) which were as follows: glucose, insulin, resistin, and homeostasis assessment model. As the best final classifier, the random forest was used in the training/test method and with nine variables, with 83.3% accuracy, 100% sensitivity, 64% specificity, and 0.881 of area under the curve.