A new arXiv preprint introduces ClinSDT, a method that adds LLM-encoded clinical semantic guidance to Decision Transformer-based models for offline sepsis treatment. The work targets the challenge of learning treatment policies from retrospective ICU data, where existing Decision Transformer approaches rely primarily on raw trajectory information.
The authors argue that incorporating clinical semantics encoded by a language model can provide additional structure beyond raw sequences. While the abstract is truncated, it clearly frames ClinSDT as a response to the limitations of current sequence-modeling approaches in this domain.
As a preprint, the full details of the architecture and evaluation are not yet available from the abstract alone. The paper is listed on arXiv under announcement type "new," indicating a recent release.