A Leakage-Free Rank-Based Ensemble Feature Selection Pipeline for High-Dimensional Gene Expression Data: A Case Study on Prostate Cancer

Accurately classifying prostate cancer from high-dimensional gene expression data remains challenging due to a small number of samples and a large number of features. We study the performance of Analysis of Variance (ANOVA), Relief Feature Selection (ReliefF), Minimum Redundancy Maximum Relevance (mRMR), and an ensemble proposed model that combines the outputs of the three methods. We applied a strictly leakage-free ۵-fold cross-validation pipeline on a publicly available prostate cancer microarray dataset consisting of ۱۰۲ samples and ۲,۱۳۵ genes. Feature selection was performed using only training folds, followed by Random Forest (RF) classification with fixed hyperparameters. Results are reported as balanced accuracy (BAC) and area under the ROC curve (AUC), and statistical differences between methods were assessed using Wilcoxon signed-rank tests. All methodologies were able to achieve similarly high training performance (BAC≈۰.۹۷۶-۰.۹۷۸; AUC ۰.۹۹۷-۰.۹۹۸), strong test performance (BAC ۰.۹۱۳-۰.۹۲۳; AUC ۰.۹۶۳-۰.۹۶۷), and thus generalized well without overfitting. None of these differences were statistically significant, indicating that an ensemble approach does not confer additional predictive benefit. These results show that several feature selection strategies can reliably identify discriminative genes and achieve comparable classification performance in small-sample, high-dimensional gene expression datasets.

کلیدواژه ها:

Feature selection ، Ensemble Learning ، Gene Expression Data ، Prostate Cancer ، Data Leakage