Clinical-Anchored Incremental Survival Learning: A Stability-Constrained Framework for Molecularly Enhanced Survival Prediction in Lung Adenocarcinoma

Accurate survival prediction in lung adenocarcinoma (LUAD) remains challenging because high-dimensional molecular data can introduce overfitting and unstable feature selection. We developed Clinical-Anchored Incremental Survival Learning (CAISL), a framework that integrates a sparse molecular risk correction with a clinical baseline to improve prognostic discrimination while constraining model complexity and prediction variability. CAISL was evaluated using The Cancer Genome Atlas LUAD cohort comprising 470 patients and 119 observed deaths. Performance was compared with clinical Cox, molecular-only, and unanchored multimodal models using five independently seeded, event-stratified five-fold cross-validation runs. CAISL achieved a mean concordance index (C-index) of 0.7246, compared with 0.6970 for clinical Cox, yielding an absolute improvement of 0.0276 (95% bootstrap confidence interval [CI], 0.005–0.050). Time-dependent discrimination was higher at three years (AUC, 0.742 vs. 0.712) and five years (AUC, 0.721 vs. 0.691), while the integrated Brier score was lower for CAISL than for clinical Cox (0.164 vs. 0.181). Ablation analysis identified clinical anchoring as the component associated with the largest reduction in discrimination when removed. Stability analysis showed a mean pairwise Jaccard similarity of 0.46 between selected feature sets and a mean Spearman correlation of 0.86 between molecular correction scores. These findings support the potential of clinically anchored, stability-constrained molecular learning to improve internal survival-risk prediction in LUAD. Independent-cohort validation is needed to establish generalizability.