In my PhD I am exploring how to model a patient as a trajectory rather than a snapshot: a single coherent latent state that any modality can write into and that any modality can be sampled from. A patient's record is many time series on their own irregular clocks, where diagnoses, medications, labs and reports accumulate cheaply over years while an expensive channel such as a CT scan arrives only a handful of times, and each channel is recorded only when cost, protocol or clinical reason calls for it, which leaves the record sparse and unevenly spaced rather than complete. I want that state to update from whichever channels happen to be available and to decode back into any of them at a chosen timepoint within the modelled horizon, including timepoints at which that channel was never recorded. What it has to encode is how the channels inform one another across time: which changes in one announce something that only becomes visible later in another. A scan can be ordered for today, but not for a year from now, and not for the treatment that was not given. I am equally interested in what longitudinal medical datasets share, since intensive-care stays, cancer follow-up series and national registries differ in horizon and population yet all show irregular sampling, informative missingness and censoring, which is what lets me ask how generally such a state behaves across modalities and cohorts.