Characterizing pituitary adenomas in clinical notes: Corpus construction and its application in LLMs

Area of Science:

• Medical Informatics
• Natural Language Processing
• Oncology

Background:

• Pituitary adenomas present complex diagnostic challenges due to varied clinical features and pathological diversity.
• Accurate characterization of pituitary adenomas in clinical notes is crucial for effective diagnosis and treatment.

Purpose of the Study:

• To construct a high-quality annotated corpus for characterizing pituitary adenomas in clinical text.
• To explore the application of this corpus in large language models (LLMs) for clinical information extraction.

Main Methods:

• Retrospective collection of 2000 clinical documents from a pituitary adenoma neurosurgery center.
• Development of a semi-automatic framework for corpus construction and annotation.
• Analysis of the Text Corpus of Pituitary Adenomas (TCPA) and LLM fine-tuning/prompting experiments.

Main Results:

• The TCPA contains 4782 medical entities and 28,998 tokens, with high inter-annotator agreement (0.862-0.986).
• LLM experiments demonstrated TCPA's effectiveness in automatically identifying clinical information from free text.
• Incorporating clinical characteristics reduced the need for training data, lowering labor costs.

Conclusions:

• The developed TCPA effectively characterizes pituitary adenomas in clinical notes.
• The proposed methodology offers a valuable reference for medical natural language processing research in specialized domains.