<p>In the analysis of censored survival data, it is crucial to consider the heterogeneity of treatment effects in order to avoid biased inferences. The deep neural network-based heterogeneous partially linear Cox model offers a flexible and robust solution by incorporating both heterogeneous linear and homogeneous nonlinear components. The authors propose a novel approach to subgroup detection for this model under right-censoring, using deep neural networks to approximate nonlinear effects. To simultaneously estimate parameters and identify subgroups, the authors employ a concave pairwise penalty and the alternating direction method of multipliers (ADMM) algorithm. Furthermore, the authors demonstrate that the proposed estimator possesses oracle properties and achieves model selection consistency. Through simulation studies and empirical data analysis on breast cancer, the authors illustrate the effectiveness of the proposed method.</p>