Abstract / Summary
Lung adenocarcinoma (LUAD), one of the most challenging cancer subtypes, is the leading cause of cancer-related mortality worldwide. LUAD is characterized by profound molecular heterogeneity, leading to inconsistent outcomes among clinically similar patients. Conventional tumor–node–metastasis (TNM) staging fails to capture LUAD. We developed an integrated prognostic framework that combines transcriptomic and post-transcriptional regulatory layers. We analyzed 443 TCGA-LUAD patients with matched clinical, miRNA (1881 features), and gene expression (60,660 genes) data, including 155 deaths (35.0% event rate). Genome-wide univariate Cox screening was performed separately for miRNA and gene features with Benjamini–Hochberg correction and embedded within a 5 × 5 nested cross-validation framework with elastic-net Cox hyperparameter selection. Features selected in ≥ 2 of 5 outer folds formed a consensus panel, which was refit on the development cohort (n=376), evaluated on a held-out test set (n=67), and externally validated across three GEO cohorts (n=913). Censoring-aware IPCW concordance and time-dependent AUC were used throughout. Nested cross-validation yielded modest performance (Harrell’s C=0.550 ± 0.083; IPCW C=0.552 ± 0.042; mean AUC=0.543 ± 0.113). A four-feature consensus panel comprising LDLRAD3, ANLN, hsa-mir-4435-2, and ERO1A achieved Harrell’s C=0.645, IPCW C=0.573, and mean AUC=0.700 on the internal holdout. Because miRNA was unavailable on the external microarray platforms, a separately locked three-gene model (ANLN, LDLRAD3, ERO1A) was evaluated externally, achieving Harrell’s C=0.637–0.648 in GSE31210, GSE30219, and GSE72094. Rigorous nested validation and censoring-aware evaluation indicate modest but externally reproducible prognostic discrimination in three independent cohorts, while avoiding feature-selection leakage and overoptimistic performance estimates.