New preprint – we use LLMs to streamline building powerful patient representations in complex datasets to enable sample-efficient downstream learning.