Towards normalized clinical information extraction in Chinese radiology report with large language models

Radiology reports serve as a fundamental component within electronic medical records. Converting unstructured free-text reports into structured formats holds paramount importance for the management and utilization of radiology reports. In this paper, we propose a novel information extraction paradigm named normalized clinical information extraction (NCIE) for Chinese radiology reports. Specifically, NCIE operates in an end-to-end fashion to extract normalized and structured clinical information without decomposing the information extraction process into multiple intermediate tasks. Motivated by recent progress in Large Language Models (LLMs), we address the NCIE problem based on the instruction tuning of LLMs. The proposed approach, termed R adiological I nformation E nd-to-end E xtraction with L LM (RIEEL), excels at extracting structural information comprising radiological observations alongside their corresponding anatomical locations and status. To ensure the model in learning the normalized medical concepts correctly, we establish a radiology knowledge base with expert knowledge and further curate a high-quality instruction tuning dataset. Moreover, we incorporate two data-efficient learning strategies based on data augmentation and self-training to enhance the model’s NCIE capabilities during instruction tuning. Through extensive experiments, we demonstrate that the proposed RIEEL achieves superior performances with different state-of-the-arts backbone LLMs, including Qwen1.5, Baichuan2 and LLaMA3. Remarkably, the best version of RIEEL surpasses GPT-4 in NCIE by a substantial margin of 30.61% in terms of the F1 score. • A new paradigm is proposed to end-to-end extract and normalize clinical information. • A new LLM-based method is proposed to extract well-structured clinical information. • Two data-efficient learning strategies are introduced for enhancing the LLMs. • Results show our method excels SOTAs in extracting normalized clinical information.