Automating Performance Status Annotation in Oncology Using Llama-3.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Oncology

Background:

• Automated information extraction from clinical notes is crucial for research.
• Dutch clinical notes present unique linguistic challenges.
• Accurate patient performance status is vital for palliative cancer care.

Purpose of the Study:

• To evaluate automated medical information extraction from Dutch clinical notes.
• To compare zero-, one-, and few-shot learning approaches using Llama-3.
• To identify the optimal method for extracting performance status in palliative esophagogastric cancer.

Main Methods:

• Utilized Llama-3 for automated information extraction.
• Employed zero-shot, one-shot, and few-shot learning paradigms.
• Used ACSESS-selected examples for one-shot and few-shot learning.
• Focused on extracting performance status from Dutch clinical notes of cancer patients.

Main Results:

• Few-shot learning, particularly with ACSESS-selected examples, demonstrated superior performance.
• One-shot learning with ACSESS-selected examples also yielded strong results.
• Zero-shot learning showed lower performance compared to supervised methods.

Conclusions:

• Few-shot learning is a promising approach for extracting performance status from Dutch clinical notes.
• The choice of examples significantly impacts the performance of few-shot learning models.
• Further refinement is needed to enhance the precision of automated extraction models.