Breast cancer predictive analytics using supervised machine learning techniques

Abstract

Breast cancer unarguably has been the very prominent disease amongst women as well as the next most dangerous after lung cancer. Early diagnosis and prevention is of paramount importance. Several methods such as micro-array analysis and network analysis have been proffered but they are somewhat expensive and time consuming. There is a need to develop an automated system based on Machine learning techniques to detect breast cancer early. Benign and Malignant tumors were classified using Logistic Regression (LRO), Bayes Network (BNK), Multilayer Perceptron (MLP), Sequential Minimal Optimization (SMO), J48, Naive Bayes (NBS) and Instance Based Learner (IBK) algorithms, which were implemented in Waikato Environment for Knowledge Analysis (WEKA). The breast cancer database for this study was collected from the University of Wisconsin Hospitals, published on California College, Irvive (UCI) website. The five most critical performance metrics when selecting an algorithm in model building in the health related domain are Area Under the ROC Curve (AUC), Receiver Operating Characteristics Curve (ROC), Mean Absolute Error (MAE), Accuracy and Kappa Statistic. In relation to the results of Accuracy, Precision and Kappa Statistic which were evaluated and compared, BNK has best predictive accuracy of 97.14%, followed by SMO with 96.71%, then LOR with 96.57%. On the other hand, LOR has the highest AUC of 99.3%, followed by BNK with 99.2%, then SMO with 96.5%. Beyond accuracy, AUC should be keenly considered in algorithm selection and model building. Therefore, Logistic Regression should be chosen as the best classifier instead of Bayes network for breast cancer optimal prediction.