Development and internal validation of a postoperative prognostic model integrating low-dose computed tomography radiomics with TRAP1 and COL3A1 expression in resected non-small cell lung cancer

Accurate postoperative risk stratification in non-small cell lung cancer (NSCLC) requires information beyond anatomical staging. We aimed to develop and internally validate a model integrating clinical variables, preoperative low-dose computed tomography (LDCT) radiomics, and resection-tissue TRAP1 and COL3A1 expression. This retrospective single-center study included 300 patients with completely resected stage I–IIIA NSCLC, randomly allocated to a training set ( n = 210) and frozen internal validation set ( n = 90). Prediction was defined after resection and definitive pathological assessment but before adjuvant therapy. A seven-feature Rad-score was derived from 1,316 LDCT radiomic features using reproducibility filtering and least absolute shrinkage and selection operator Cox regression. Tumor-cell TRAP1 and stromal COL3A1 H-scores were analyzed continuously. A prespecified clinical reference model was compared with models adding the Rad-score, both H-scores, or all three predictors. Performance was assessed using full-pipeline bootstrap correction, C-index, 3-year time-dependent area under the curve (AUC), calibration, Brier score, and decision-curve analysis. The training and validation sets contained 65 and 26 deaths, respectively. Higher Rad-score, TRAP1 H-score, and COL3A1 H-score were associated with poorer overall survival in both cohorts. The full model achieved an optimism-corrected C-index of 0.810 and 3-year AUC of 0.826 in training; corresponding validation estimates were 0.798 and 0.814 (95% CI, 0.737–0.891). Compared with the clinical reference model, validation improvements were 0.082 for C-index, 0.090 for AUC, and − 0.029 for Brier score. Calibration was acceptable, but validation estimates were imprecise because of the limited event count. The full model showed the highest estimated net benefit across most thresholds. Integrating LDCT radiomics with continuous TRAP1 and COL3A1 expression improved postoperative 3-year prognostic estimation beyond a prespecified clinical reference model. This postoperative, tissue-dependent model requires temporal and multicenter external validation before clinical use.