Automating Performance Status Annotation in Oncology Using Llama-3

Abstract

This work explores the automated extraction of medical information from Dutch clinical notes using Llama-3 and a limited amount of annotations. We compared zero-, one- and few-shot learning for the extraction of performance status of patients with palliative esophagogastric cancer few-shot learning and one-shot with ACSESS-selected examples, showed the best performance. Future work shall focus on improving the model's precision.

Keywords: Few-shot Learning; Generative Models; Llama-3; NLP; WHO Performance Status.