Clinical data-driven modeling of disease-specific survival in lung cancer: Insights from the national lung screening trial dataset

Lung cancer remains the leading cause of cancer-related mortality worldwide, highlighting the importance of accurate disease-specific survival (DSS) prediction to support clinical decision making. DSS, which isolates outcomes directly attributable to lung cancer, provides a precise and clinically relevant endpoint for prognostic evaluation and treatment planning. In this study, we developed and evaluated a clinical modeling framework for DSS prediction using data from the National Lung Screening Trial (NLST), incorporating detailed clinical information from 2,058 patients. Clinical features were categorized into three temporal sets: pre-diagnosis, diagnosis-informed, and full clinical information, and subjected to rigorous preprocessing, including LASSO-based feature selection and Variance Inflation Factor filtering to reduce redundancy and multicollinearity. Machine learning models, including Random Forest, Extreme Gradient Boosting, Support Vector Machine, and Logistic Regression, were trained using these features and were evaluated for predictive performance. Models trained on full clinical information achieved high discriminative performance (AUC = 0.919), while diagnosis-informed models still demonstrated strong predictive capability (AUC = 0.873). Explainability analysis consistently identifies the TNM (tumor, node, metastasis) staging as the most predictive variable in all clinical models. Kaplan–Meier survival analysis further confirmed that clinical models effectively stratify patients into statistically distinct survival groups. These findings highlight the dominant role of clinical data in lung cancer DSS prediction and demonstrate the potential of clinically grounded, interpretable models to inform personalized prognosis and treatment planning.