Enhancing the interpretability of non-linear proportional hazard models introducing ghost variables

Interpreting complex non-linear survival models while maintaining proportional hazards semantics remains a significant challenge in statistical learning. This work introduces GhostCox, a novel method designed to enhance the interpretability of non-linear proportional hazard models by quantifying covariate contributions to the risk function via ghost variables. Our approach evaluates the unique conditional relevance of each covariate by substituting its values with their estimated conditional expectations, given the remaining features. Numerical experiments and real-world applications demonstrate that GhostCox effectively identifies true predictors and is significantly more computationally efficient than established benchmarks, such as the Holdout Randomization Test (HRT). Furthermore, GhostCox yields compact feature sets with robust predictive performance. Overall, GhostCox provides a transparent, statistically grounded, and scalable framework for interpreting complex survival models, offering actionable insights into covariate effects.