Enhancing Medication Recommendation with LLM Text Representation

Abstract

Most existing medication recommendation models rely heavily on structured data such as diagnosis and procedure codes, often overlooking the wealth of clinical information embedded in unstructured text. To better exploit this data, we propose enhancing medication recommendation using text representations derived from Large Language Models (LLMs). These representations capture semantic and contextual information from clinical notes and doctors’ narratives, enabling deeper integration of free-text and coded EMR data. We incorporate LLM-derived embeddings into several baseline models and evaluate performance on two datasets: the public MIMIC-III and a real-world inpatient dataset from Ditmanson Medical Foundation Chia-Yi Christian Hospital in Taiwan. Experimental results show consistent performance gains across models when LLM-based text representations are used, with some models achieving results comparable to or better than using medical codes alone. Our findings demonstrate that LLM-enhanced representations offer a scalable and model-agnostic approach for leveraging unstructured data in clinical decision support.

Ähnliche Arbeiten