would embedding layers be preferable to one-hot encodings for conditions?

[cheind]

drawback: all conditions would have to be embeddable (i.e integers) - except the embedding layer is not part of the model, but lives outside.