A comparative study identifies random forest with minimum redundancy maximum relevance feature selection as a superior transcriptomic classifier for gastric adenocarcinoma diagnosis

Abstract The high-dimensional nature of transcriptomic data complicates the early diagnosis of gastric adenocarcinoma. While machine learning offers promise, the optimal synergy between feature selection and classifiers is unclear. To define this, we performed a systematic comparison using 1,132 tumor and 131 normal gastric tissue samples from public microarrays. Five feature selection methods (Minimum Redundancy Maximum Relevance or MRMR, F-Test, Chi², Variance Threshold, and random forest importance) and nine classifiers (RF, XGBoost, AdaBoost, SVM, KNN, DT, NB, RUSBoost, NN) were optimized via Bayesian hyperparameter tuning and evaluated using stratified cross-validation and an independent test set. Comprehensive metrics (AUC-ROC, F1, MCC, Accuracy, Precision, Recall, Kappa) identified MRMR as the most effective feature selection method, with detailed comparative results presented. Ensemble classifiers, particularly RF, XGBoost, and AdaBoost, outperformed others. The optimal pipeline combined RF with MRMR feature selection. We conclude that integrating mutual information-based feature selection with ensemble learning yields a high-performance, generalizable transcriptomic classifier, forming a robust foundation for a cost-effective molecular diagnostic tool for early gastric cancer detection.