Overview
The Survival Diffusion Probabilistic Model (SDPM) is a generative approach designed for continuous-time survival analysis. It addresses the challenges of estimating time-to-event distributions from data containing censored observations by avoiding both structural assumptions on the hazard function and discretization of the time axis. SDPM models the conditional distribution of the survival outcome, represented as a pair of observed time and a censoring indicator $\mathbb{P}(T,\delta \mid \mathbf{x})$, utilizing a denoising diffusion model.
Research Context
Survival analysis aims to estimate a time-to-event distribution from observed data. A common issue in this field is the presence of censored observations, where the event of interest has not yet occurred or is not fully observed for some subjects. Existing methodologies often rely on imposing parametric assumptions on the hazard function or discretizing the time domain. These approaches can potentially limit model flexibility and introduce approximation errors. SDPM was developed as an alternative to mitigate these limitations.
Approach
SDPM employs a denoising diffusion model to represent the conditional distribution of the survival outcome $(T, \delta \mid \mathbf{x})$. The model operates under the assumption of conditionally independent censoring. Conditional samples generated by this model can be converted into survival function estimates through the application of the Kaplan-Meier estimator. This formulation enables SDPM to avoid making parametric assumptions about the event-time distribution and eliminates the need for discretizing the output time space.
The model processes data in a transformed target space. This transformation involves the use of standardized log-times. Additionally, it incorporates a continuous Gaussian-mixture representation for the censoring indicator. The methodology explicitly aims to account for censored data within a continuous-time framework.
Findings
- SDPM was evaluated on ten real survival datasets and compared against five baseline methods. The baselines included tree-based, boosting-based, and neural survival models.
- The model demonstrated competitive predictive performance across multiple metrics: the C-index, integrated time-dependent AUC, and integrated Brier score.
- A study involving synthetic Cox-Weibull data indicated that SDPM could recover the shape of an underlying continuous survival distribution more accurately than a strong nonparametric baseline, provided that a sufficient number of samples were generated.
- An ablation study confirmed the importance of the proposed target-space transformations. These transformations were found to improve event-rate calibration, reduce the generation of invalid times, and consistently enhance predictive discrimination.
Why This Matters
SDPM's ability to estimate continuous-time survival distributions without enforcing parametric assumptions or time discretization offers a flexible approach for modeling time-to-event data, including data with censored observations. Its competitive performance across standard metrics and ability to recover underlying distribution shapes suggest its utility in settings where precise time-to-event predictions are valuable.