We built a generalist clinical foundation model across both time series and vision modalities with a novel RL training algorithm named DRPO.