Improved estimation of hospital length of stay via Large Language Models (collaboration with Artefact and CHC Liège)

Details

Supervisors

Legat, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

enA recurring issue in hospitals is the high pressure on resources and patient man agement, particularly due to bed occupancy and care organization constraints. Improving the estimation of hospital length of stay could help address these challenges by enabling better planning and resource allocation. Rather than directly estimating length of stay, this thesis focused on extracting meaningful clinical factors influencing it as a first step toward predictive modeling. This work was motivated by the emergence of large language models (LLMs) and the richness of unstructured clinical data, which remained underused despite its high informational value. However, due to the black-box nature of LLMs, a Retrieval-Augmented Generation (RAG) pipeline was proposed. RAG made it possible to identify the exact sentences from patient records that were used by the LLM for factor detection, ensuring traceability and improving interpretability of the results. The system combined sentence-level chunking, medical-specific embeddings, vector retrieval and used Mistral Small. The approach was evaluated on a synthetic dataset. Using a simple top-K strategy with K = 3 achieved approximately 84% recall and 94% precision while remaining suitable for real-world use by avoiding excessive information for clinicians to review. Overall, this approach demonstrated the potential of RAG to transform clinical notes into structured and interpretable data.