Generating synthetic time-to-event data is important for survival analysis, but censoring makes it statistically difficult. The authors propose censored flow matching (CFM), a flow-based generative method that does not rely on survival likelihoods. Instead, it exploits the monotonicity property of differential flows to address censoring through pullback sampling.
The method alternates between pullback sampling and CFM steps, forming a flow-based expectation-maximization algorithm tailored for time-to-event inference. Under mild distributional conditions, the authors prove that CFM converges to the true event-time distribution as the EM iterations proceed. This provides a flexible way to learn probabilistic densities while incorporating observed censoring times for truncated sampling or density estimation.
In simulation studies, CFM shows benefits for complex and large-scale data generation. Applied to the Organ Procurement and Transplantation Network (OPTN) and MIMIC-IV datasets, it demonstrates improvements in synthetic and predictive survival analyses, including data fidelity, predictive performance, and empirical privacy characteristics.