Delirium Identification from Nursing Reports Using Large Language Models

This study investigates large language models for delirium detection from nursing reports, comparing keyword matching, prompting, and finetuning. Using a manually labelled dataset from the University Hospital Freiburg, Germany, we tested Llama3 and Phi3 models. Both prompting and finetuning were effective, with finetuning Phi3 (3.8B) achieving the highest accuracy (90.24%) and AUROC (96.07%), significantly outperforming other methods.