Risk scores are fundamental to clinical decision-making, such as determining when to initiate statins for cardiovascular disease prevention. However, existing risk scores face two major limitations: they operate under a fixed, unspecified future policy, and their clinical utility depends on how the wrapping policy impacts patient outcomes. Prediction under different policies requires estimating interventional distributions that include timing, which is challenging because standard next-token, autoregressive transformers cannot naturally simulate altered timings for specific future events. To address this, we introduce Competing Event Models (CEMs), a generative autoregressive framework that simplifies effect estimation for both intervention type and timing.
