Sequential Trajectory Generation

Background: Generative foundation models, such as the next-visit prediction framework presented, currently focus on one-step-ahead predictions for clinical events.

Question / Future Work: The optimal methodology for leveraging the generative nature of these models to sequentially generate entire future patient trajectories, rather than just the next visit’s events, remains an open research question. This involves adapting existing inference-time algorithms from fields like large language models (e.g., those using reinforcement learning for long-form generation) to the structured, sequential nature of clinical data.