Prognostic biomarker discovery in pancreatic cancer through hybrid ensemble feature selection and multi-omics data

BACKGROUND: Accurate prediction of patient survival using high-dimensional multi-omics data requires effective feature selection methods that balance predictive performance, sparsity, and stability to enable reproducible and transparent prognostic biomarker discovery. Existing approaches often rely on ad hoc thresholds, limiting their reproducibility and clinical utility. METHODS: We developed a hybrid ensemble feature selection (hEFS) framework that integrates data subsampling with multiple prognostic models, combining embedded and wrapper-based strategies for survival prediction. Features are ranked using a voting-theory-inspired aggregation approach across models and subsamples, while the optimal feature subset is determined via Pareto front optimization, eliminating the need for user-defined thresholds. We benchmarked hEFS against conventional CoxLasso using 100 Monte Carlo cross-validation iterations on multi-omics datasets from three pancreatic ductal adenocarcinoma (PDAC) cohorts, assessing sparsity, stability, redundancy, discrimination, and computational cost. RESULTS: Across all cohorts, hEFS selected substantially fewer and more stable features than CoxLasso (~10 vs. ~60 per modality), with lower variance, while still maintaining comparable discrimination performance with clinical baselines (C-index ~0.54–0.60). Feature redundancy was low, and ensemble subsampling further improved stability. Prognostic performance depended mainly on modality choice, rather than integration model, with gene expression plus clinical variables achieving the highest discrimination power (C-index ~0.64). CONCLUSIONS: hEFS provides a fully automated, interpretable, and reproducible framework for multi-omics biomarker discovery in high-dimensional survival analysis. By producing sparse, yet stable biomarker panels without loss of predictive power, hEFS represents a practical tool for integrative multi-omics modeling and it is openly available in the mlr3fselect R package.