Improving breast cancer prediction using a pattern recognition network with optimal feature subsets

AIM: To predict the presence of breast cancer by using a pattern recognition network with optimal features based on routine blood analysis parameters and anthropometric data. METHODS: Sensitivity, specificity, accuracy, Matthews correlation coefficient (MCC), and Fowlkes-Mallows (FM) index of each model were calculated. Glucose, insulin, age, homeostatic model assessment, leptin, body mass index (BMI), resistin, adiponectin, and monocyte chemoattractant protein-1 were used as predictors. RESULTS: Pattern recognition network distinguished patients with breast cancer disease from healthy people. The best classification performance was obtained by using BMI, age, glucose, resistin, and adiponectin, and in a model with two hidden layers with 11 and 100 neurons in the neural network. The accuracy, sensitivity, specificity, FM index, and MCC values of the best model were 94.1%, 100%, 88.9%, 94.3%, and 88.9%, respectively. CONCLUSION: Breast cancer diagnosis was successfully predicted using only five features. A model using a pattern recognition network with optimal feature subsets proposed in this study could be used to improve the early detection of breast cancer.