Bayesian nonparametric survival analysis via deep Dirichlet process

Time-to-event data analysis has attracted growing interest across various domains. A major challenge lies in the inherent incompleteness of such data, with right censoring being the most common form. Although Cox's proportional hazards model is effective for traditional datasets, it often fails in modern settings where high-dimensional covariates violate the proportional hazards assumption. Additionally, conventional parametric survival models, typically based on the exponential family, impose strong assumptions and struggle to capture long-tail behavior due to their exponential decay. To address the aforementioned challenges, we propose DDPSurv, a novel deep learning framework for survival analysis that leverages a deeply parameterized Dirichlet Process (DP) mixture model. Unlike prior deep parametric approaches that rely on strong statistical assumptions, DDPSurv offers greater flexibility in modeling survival distributions by incorporating a DP mixture, enabling improved representation of heavy-tailed behavior. We provide theoretical guarantees showing that the model can closely approximate the true survival distribution with a tight concentration rate. Empirical results on standard survival benchmarks demonstrate the effectiveness of our method, while large-scale experiments on MIMIC-III and MIMIC-IV datasets further validate its scalability and clinical utility. Codes are available at https://github.com/HKU-MedAI/DDPSurv-net.