Abstract

Abstract
Machine learning–based survival models are increasingly recognized for their ability to handle complex, high-dimensional survival data, particularly when traditional model assumptions are violated. Yet, their empirical value in competing risks frameworks remains insufficiently understood. This study provides a rigorous comparative evaluation of Random Survival Forests (RSF) and the Fine–Gray subdistribution hazard model under two challenging conditions: (i) violation of the proportional hazards assumption and (ii) presence of intricate nonlinear interactions. Controlled simulation experiments comprising 1,000 replicates were conducted, followed by external validation using two large-scale clinical datasets-the Rotterdam breast cancer cohort (n = 2,982; recurrence risk = 51%) and the FLchain lymphoid malignancy cohort (n = 7,874; cancer-specific mortality = 7%). Model performance was assessed via integrated Brier scores and time-dependent prediction errors. In simulations, RSF substantially outperformed Fine–Gray models, yielding 30.7% and 32.3% improvements in predictive accuracy under non-proportional hazards and nonlinear interaction settings, respectively. However, real-world validation revealed nuanced context-specific patterns. Within the FLchain cohort, RSF achieved a 28.9% gain over Fine–Gray despite only minor violations of proportionality (global p = 0.016), while in the Rotterdam data, Fine–Gray retained a slight edge (4.2% difference) even under pronounced assumption breaches (global p < 2 × 10⁻¹⁶). These findings underscore that the relative performance of machine learning versus classical survival approaches is contingent upon the data structure and the nature of assumption violations, rather than their mere presence. RSF exhibits clear advantages in scenarios dominated by biomarker-driven, nonlinear risk mechanisms, whereas traditional competing-risk models maintain remarkable robustness in complex clinical decision settings. This study provides empirical evidence and methodological guidance for researchers and practitioners selecting appropriate tools for competing-risk analysis in heterogeneous biomedical data.
Keywords: Competing Risks, Random Survival Forest, Fine-Gray Model, Proportional Hazards Violation, Simulation Study, Clinical Prediction.