Toward explainable precision oncology: SHAP-guided cox survival modeling from population-level prediction to patient-specific explanation in NSCLC

Accurate risk stratification and transparent prognostic modeling are essential for personalized management of non-small cell lung cancer (NSCLC). While radiomic features offer rich quantitative characterization of tumor phenotypes, their integration into survival models is often limited by multicollinearity, reduced interpretability, and challenges in clinical translation. We propose an interpretable analytical framework that integrates Cox Proportional Hazards (CPH) modeling with SHapley Additive exPlanations (SHAP) to provide both robust survival prediction and multi-level interpretability. A two-stage feature selection strategy was implemented to mitigate multicollinearity and optimize model parsimony. First, correlation-based pruning was applied to radiomic features to remove redundancy while preserving informational content. Subsequently, a stepwise selection procedure guided by the Akaike Information Criterion (AIC) identified a final subset of prognostically relevant radiomic, clinical, and demographic variables. Model performance was evaluated using the concordance index (C-index), and patients were stratified into low-, medium-, and high-risk groups based on predicted survival probabilities. Kaplan–Meier analyses were used to compare model-derived risk groups with traditional staging systems. SHAP analysis was conducted at global, risk-stratified, and individual levels to elucidate feature contributions and patient-specific risk profiles. The final CPH model incorporated 14 predictors (10 radiomic and 4 non-radiomic features) and demonstrated stable prognostic performance. Model-informed risk groups achieved superior separation of survival outcomes compared to conventional staging, with sustained discrimination observed over extended follow-up periods of up to twelve years. SHAP analysis revealed that clinical and demographic factors (e.g., stage, age, sex) along with radiomic features contributed substantially to mortality risk, with feature importance varying across risk strata. Risk-stratified and patient-specific SHAP visualizations highlighted heterogeneous patterns of feature influence, clarified cases of concordant and discordant predictions, and provided clinically interpretable explanations for individual risk assessments. This study demonstrates that combining traditional survival modeling with SHAP-based interpretability yields a powerful and transparent framework for personalized risk assessment in NSCLC. The proposed approach outperforms standard staging in long-term prognostic stratification while offering clinically meaningful explanations at both cohort and individual levels. These findings support the integration of interpretable radiomic-driven survival models into precision oncology workflows and provide a foundation for future extensions incorporating treatment-specific variables and nonlinear survival modeling techniques. By unifying robust survival prediction with explainable artificial intelligence, this framework advances interpretable, data-driven decision support for personalized lung cancer prognosis and management.