Predicting Breast Cancer via Supervised Machine Learning Methods on Class Imbalanced Data

A widespread global health concern among women is the incidence of the second most leading cause of fatality which is breast cancer. Predicting the occurrence of breast cancer based on the risk factors will pave the way to an early diagnosis and an efficient treatment in a quicker time. Although there are many predictive models developed for breast cancer in the past, most of these models are generated from highly imbalanced data. The imbalanced data is usually biased towards the majority class but in cancer diagnosis, it is crucial to diagnose the patients with cancer correctly which are oftentimes the minority class. This study attempts to apply three different class balancing techniques namely oversampling (Synthetic Minority Oversampling Technique (SMOTE)), undersampling (SpreadSubsample) and a hybrid method (SMOTE and SpreadSubsample) on the Breast Cancer Surveillance Consortium (BCSC) dataset before constructing the supervised learning methods. The algorithms employed in this study include Naive Bayes, Bayesian Network, Random Forest and Decision Tree (C4.5). The balancing method which yields the best performance across all the four classifiers were tested using the validation data to determine the final predictive model. The performances of the classifiers were evaluated using a Receiver Operating Characteristic (ROC) curve, sensitivity, and specificity.