Advanced breast cancer prediction using Deep Neural Networks integrated with ensemble models

Abstract

Breast cancer (BC) is a fatal illness that affects millions of people every year. After lung cancer, BC illness is one of the world's major causes of death for women. A breast cell-derived malignant tumor is referred to as BC. Both developed and developing countries are struggling with this widespread cancer. Machine learning (ML) and Deep Learning (DL) have appeared as effective technologies in BC predictions with the highest accuracy in the past years due to their robust taxonomy and diagnostic capabilities. This paper introduces a novel Deep Neural Networks-based Stacking Ensemble Model (DNN-SEM) enhanced with a hybrid stacking ensemble model (SEM) and Extra Tree Classifier (ETC) technique to extract the most essential features from the suggested BC datasets. The proposed DNN-SEM integrates Deep Belief Network (DBN) and Artificial Neural Network (ANN) as level-1 models, referred to as SEM-DBN and SEM-ANN, respectively. The level-1 models are designed using four traditional ML algorithms, including XGBoost Classifier (XGBC), Logistic Regression (LR), Random Forest (RF), and Support Vector Machine (SVM), which are designed as level-0 models. The proposed DNN-SEM model is trained using four BC datasets, namely Diagnostic Wisconsin Breast Cancer Dataset (WBCD) (Dataset-I), Coimbra Breast Cancer Dataset (CBCD) (Dataset-II), Original Wisconsin Breast Cancer Dataset (WDBC) (Dataset-III), and Prognostic Wisconsin Breast Cancer (WBCP) (Dataset-IV). The efficacy of the proposed DNN-SEM is assessed through established evaluation metrics, including accuracy, sensitivity, specificity, Matthew's correlation coefficient (MCC), F-score, confusion matrix, and ROC curves. To analyze the efficiency of the DNN-SEM, its performance is compared with the proposed single classifiers, ensemble, and state-of-the-art models present in the literature. The results demonstrate that DBN-SEM achieves the highest accuracy of 99.62 %, with the lowest error rate. The proposed DBN-SEM and ANN-SEM achieved promising accuracy scores against level-0 and state-of-the-art methods.