Enhancing Interpretability of Survival Models With High Predictive Performance: Integrating Cox Proportional Hazards With Machine Learning Methods

Recent advances in machine learning (ML) have substantially improved the performance of survival prediction, yet this progress has been accompanied by a critical decline in model interpretability. In clinical decision-making, such loss of transparency undermines trust and practical applicability. To address this issue, we developed an interpretable high-performance survival prediction framework by integrating multiple ML survival models with the Cox proportional hazards model (Cox model) as the meta-learner.The proposed framework preserves the nonlinear modeling ability of ML while embedding interpretability directly within the model through the Cox meta-learner. This structure enables feature-level interpretation, single-factor analysis, and construction of clinically intuitive nomograms—capabilities that conventional ML models such as Gradient Boosted Survival Trees (GBST) cannot provide. The Cox meta-learner yields statistically grounded hazard ratios (HRs) and confidence intervals, bridging rigorous survival analysis and explainable AI. Validation on the SEER database demonstrated that the proposed model achieved a C-index of 0.8313, comparable to GBST (0.8364), and superior to the classical Cox model (0.7958). Robustness was further confirmed through consistent generalization across multiple external datasets, particularly in small-sample datasets and a non-oncological heart failure dataset. These findings demonstrate that the proposed method restores interpretability while maintaining high predictive capacity, achieving clinically meaningful transparency and reliability across diverse survival prediction tasks.