Diffusion models learn to predict when events will happen

Survival analysis predicts when events (like disease progression) will occur, but existing methods either assume specific hazard patterns or artificially discretize time into buckets. This work applies diffusion probabilistic models to generate realistic survival distributions directly in continuous time, then extracts survival curves using standard Kaplan-Meier estimation. On ten real datasets, SDPM matches or beats tree-based, boosting, and neural baselines on C-index and calibration metrics. Code is released.