Lung adenocarcinoma survival models stop letting molecular data dominate: clinical anchoring plus sparse correction lifts C-index from 0.697 to 0.7246
Synopsis
Molecular data in lung adenocarcinoma are high-dimensional relative to the number of observed events, so concatenating them into a survival model invites overfitting and unstable feature selection. CAISL writes predicted risk as a clinical Cox risk plus a sparse molecular correction, and penalizes how much that correction varies across resampled training cohorts. On 470 TCGA-LUAD patients with 119 observed deaths, mean C-index across five independent cross-validation runs was 0.7246 versus 0.6970 for clinical Cox, a paired gain of 0.0276 (95% CI 0.005–0.050); this is internal validation and still needs an independent cohort.
Interpretation
Survival-risk ranking in lung adenocarcinoma can improve through a small molecular correction that keeps the clinical baseline in charge. A prior multimodal gradient-boosting model on this cohort did not beat the clinical Cox baseline, and gene-selection stability was low. On 470 TCGA-LUAD patients with 119 observed deaths, across five independently seeded, event-stratified five-fold cross-validation runs, CAISL had the highest C-index in every run, mean 0.7246, against 0.6970 for clinical Cox, 0.681 for the molecular-only model and 0.692 for the unanchored multimodal model.
The molecular correction is built as an increment that can shrink to zero, spending its effort on patient pairs the clinical model ranks wrongly or uncertainly. The usual route concatenates molecular features with clinical variables and lets an unrestricted predictor set the whole risk ordering. Training weights censoring-aware pairwise ranking by the cross-fitted clinical risk margin, then adds sparse regularization and a penalty on disagreement between corrections across resampled subsamples; time-dependent AUC at 1, 3 and 5 years was 0.704, 0.742 and 0.721 for CAISL versus 0.688, 0.712 and 0.691 for clinical Cox.
Removing clinical anchoring costs more discrimination than any other component ablation, marking the clinical reference as the load-bearing part of the framework. Dropping the sparsity or stability penalty produced smaller drops whose confidence intervals included zero. Full CAISL reached a C-index of 0.725; without clinical anchoring it fell to 0.692, a paired difference of −0.033 (95% CI −0.053 to −0.013), while the clinical baseline alone differed by −0.028 (−0.048 to −0.008).
Perspective
The results come from internal cross-validation of a single TCGA-LUAD cohort of 470 patients with 119 observed deaths, and the evaluated target is prognostic risk ranking, not treatment response or clinical benefit. The authors state that independent external validation, followed by prospective trials assessing impact on clinical workflows and patient outcomes, is required before clinical implementation.
Calibration drifts with horizon: at 5 years mean predicted risk was 0.53 against a censoring-adjusted observed risk of 0.56, and the calibration slope point estimate fell from 0.97 at 1 year to 0.88 at 5 years, though every slope confidence interval covered 1. Mean pairwise Jaccard similarity of selected feature sets was only 0.46 while mean Spearman correlation of molecular correction scores was 0.86, and the authors stress that prediction-score consistency does not establish biological validity. Discrimination varied non-monotonically with anchoring strength, falling to 0.712 at four times the primary setting.