Stacked elastic net penalized logistic regression for ultra-high-dimensional cancer genomic classification

In high-dimensional genomic data analysis, tackling multicollinearity while preserving model interpretability remains a critical challenge. This study utilizes the Stacked Elastic Net (SEN) ensemble model to predict cancer classification based on gene expression and proteomics data. SEN integrates multiple elastic net models with varying regularization parameters through a two-layer framework: base learners generate diverse regularization paths, while a meta-learner with Lasso penalty optimally combines their predictions. Evaluation on four cancer datasets–thyroid carcinoma (THCA), head and neck squamous cell carcinoma (HNSC), pancreatic neuroendocrine tumor (PNET), and colorectal cancer (Colon)–demonstrated that the SEN method achieved a maximum prediction accuracy of 99.41% with AUC values ranging from 0.9412 to 0.9979, outperforming traditional machine learning methods. Systematic hyperparameter tuning identified optimal model configurations, and feature selection combined with SHAP analysis provided biological interpretation by identifying key biomarkers. This method offers an effective solution for high-dimensional cancer classification while maintaining interpretability.