LightGBM–ElasticNet-Cox: a gradient boosting regularization hybrid model for survival analysis

Survival analysis is fundamental to modeling time-to-event outcomes in biomedical research. However, the conventional Cox proportional hazards (PH) model is limited by its linearity assumption and incapability of capture complex interactions and nonlinear effects, particularly in high-dimensional settings. This paper proposes Light Gradient Boosting Machine (LightGBM)–ElasticNet–Cox (LEC), a hybrid framework that integrates Cox partial likelihood with gradient-boosted decision trees and Elastic-Net regularization. By employing a second-order Taylor expansion, the model learns flexible nonlinear risk functions, enabling automated discovery of nonlinear effects, hierarchical interactions, and localized hazard dynamics. The Elastic-Net penalty promotes sparsity and enhances numerical stability under censoring. Extensive evaluations show that LEC outperforms the Cox PH model, penalized Cox models, Random Survival Forests, and conventional gradient boosting baselines in most cases on discrimination and calibration. Furthermore, Shapley-Additive-Explanation (SHAP)-based interpretability analysis reveals feature-level contributions and clinically meaningful interaction patterns, ensuring model transparency.