Abstract
Background/Objectives: Lung cancer remains a leading cause of cancer-related death, and prognostic assessment relies mainly on TNM staging, which incompletely captures patient heterogeneity. Machine learning (ML) methods may improve survival prediction, but their use in real-world national registries with rigorous validation remains limited. This study aimed to develop ML-derived phenotypes and 1-year mortality risk groups and to evaluate their performance and clinical utility in a national lung cancer cohort from Kazakhstan. Methods: We conducted a retrospective study using a national registry including 13,685 patients. Eight routinely collected predictors were analyzed. K-means clustering was used for exploratory phenotyping. A random survival forest (RSF) model estimated 1-year mortality risk and defined low, intermediate, and high risk groups. Performance was evaluated using temporal validation, cross-validation, and bootstrap correction. Discrimination was assessed using the concordance index, prediction accuracy using the Brier score, and calibration using risk group comparisons. Comparator models included penalized Cox and TNM-only models. Clinical utility was assessed using decision-curve analysis. Results: Two phenotypes showed distinct survival outcomes, although cluster separation was modest. The RSF model showed stable performance (C-index 0.679; corrected 0.663). Risk groups demonstrated strong survival separation (high vs. low: HR 5.66). The RSF model outperformed the penalized Cox (C-index 0.544) and TNM (0.606), with improved accuracy (Brier 0.169 vs. 0.212). Calibration was generally good. Decision-curve analysis showed greater net benefit. Conclusions: An RSF-based model using routine registry data provided robust internally validated risk stratification and improved predictive performance. Clustering results were exploratory. External validation is re-quired before clinical implementation.