An Enhanced Feature Selection Approach for Breast Cancer Prediction Using a Hybrid Framework

Breast Cancer (BC) is a prevalent disease worldwide, particularly affecting women. Developing accurate classification models is a significant challenge in medical informatics. However, machine learning techniques have effective categorization models and data mining from repositories often involves high dimensionality, incorrect labeling, and class imbalance. These issues can adversely affect the performance of most statistical learning approaches for outcome prediction. The aim of this research is to introduce an innovative approach for Feature Selection (FS) that combines Maximum Likelihood Logistic Regression (MLLR) with Noise Filtering (NF) and Data Balance (DB) techniques. Our proposed method involves several steps improving breast cancer prediction accuracy. Firstly, we utilized multiple iterations of the Iterative Noise Filtering based on the Fusion of Classifiers (INFFC) algorithm to reduce class noise present in cancerous datasets. Next, we addressed the issue of class imbalance by using a modified Synthetic Minority Over-sampling Technique (SMOTE) that employs random linear interpolation between malignant instances and their nearest neighbors to create new synthetic malignant samples. Subsequently, we applied the Maximum Likelihood Logistic Regression (MLLR) approach and determined the statistical significance of each feature's estimated coefficients with the Wald test at a given confidence interval. Additionally, we compared the effectiveness of our proposed approach with several well-known feature selection methods such as Correlation Feature Selection (CFS), Chi-Square (CS), Principal Components Analysis (PCA), and ReliefF (RF, RFW) on the Wisconsin breast cancer dataset Prognostic (WPBC). Finally, we utilized five different classifiers—Naïve Bayes (NB), Multi-Layer Perceptron (MLP), K-Nearest Neighbors (KNN), AdaBoostM1, and Bagging—to assess their performance using the area under the receiver operating characteristic curve (AUC) metric. The experimental results indicated the Maximum Likelihood Logistic Regression (MLLR) approach consistently outperformed other feature selection (FS) methods in all cases. Among the standard FS techniques, ReliefF (RF) produced the best area under the receiver operating characteristic curve (AUC) scores. Our findings demonstrated that combining MLLR with noise filtering and data balancing can effectively select relevant features and improve classifier performance in binary classification problems. Furthermore, based on the results of five different classifiers, our proposed technique outperformed the compared methods in terms of prediction accuracy. We used ROC curve to evaluate the result performance.