The latter, not the former. The former is extraction, not prediction.

A lot of factors go into predicting the post-acute care route that results in the best outcome. You are correct that it would be possible to create features using an LLM, but that is a very difficult problem compared to simply treating the problem as a sequence classification.