ZERO-SHOT INFORMATION EXTRACTION FROM RADIOLOGICAL REPORTS USING CHATGPT

Electronic health records contain an enormous amount of valuable information, but much of it is recorded in free text. Information extraction (IE) transforms unstructured text into structured data for secondary analysis, yet traditional IE components (e.g., named entity recognition and relation extraction) require annotated data, creating a major bottleneck. With large language models (LLMs) performing well on downstream NLP tasks without parameter tuning, zero-shot IE becomes feasible. This study explores whether ChatGPT can extract useful information from radiological (CT) reports. The authors design a prompt template targeting lung cancer–related information, generate prompts by combining the template with CT reports, and obtain ChatGPT responses. A rule-based post-processing module converts responses into structured outputs. Experiments on 847 CT reports from Peking University Cancer Hospital show ChatGPT achieves competitive performance on some extraction tasks compared with a baseline IE system, though notable limitations remain.

Danqing Hu, Bing Liu, Xiaofeng Zhu, Xudong Lu, Nan Wu