@Sam: For a RNN model, it is less about the number of samples, and more about repetition (same feature repeated at each time step means same thing) or having a changeable sequence length.

As usual, the only way to be sure is to give a few different models a try

You might find some time-distributed layers work for you - where some of the layers near the input effectively share weights, so would learn the same feature representation of the raw clinical data.

For clarity, my clinical measurements will be of the same type in that they are numeric (not sure if this is what you mean). In terms of the sequence being the same length, are you referring to the number of time points for a given data point? i.e. 5 time-varying readings for each patient. I was under the assumption keeping the number of time points the same was a prereq for these type of models.

Thanks again

@Sam: I didn't mean numeric vs non-numeric, but for instance, are the same measurements taken repeatedly (e.g. blood pressure, heart rate etc), or different clinical measures taken each time. If they are the same (or mostly the same) then this is more sequence-like, so a better fit for RNN or time-distributed.

You can always take a simple flat input into a model such as xgboost, and use it as a basedline to compare more complex models against.

So you can still do what your employer wants, and you will have a way to assess whether it was worth doing

@NeilSlater: The data is structured for measurements to be taken for each covariate at fixed time points. So yes, it is the same measurements being taken at different points. When you say flat input, are you referring to transforming subsequent readings as a separate covariate? I.e. bmi_t; bmi_t+1; bmi_t+2 for instance? Up until now, this is how I've been treating time-varying covariates. Thanks again.