Enhanced clinical cases dataset on hepato-pancreato-biliary with refined diagnoses and supporting evidence

Real and comprehensive clinical cases play a critical role in the clinical applicability of large language models (LLMs). The ClinBench-HPB dataset collects 337 real-world clinical cases of hepato-pancreato-biliary (HPB) diseases, sourced from journals, websites, and hospitals. These cases include patients’ chief complaints, present illness, past medical history, signs, laboratory tests, imaging findings, imaging impressions, and clinical diagnoses. However, we identified labeling errors, omissions, and redundancies in clinical diagnoses. To address this, we performed a case-by-case review, correcting the clinical diagnoses and providing supporting evidence. The improved dataset can be used for training and evaluation of LLMs in HPB-specific tasks. Importantly, the added supporting evidence can be structured as chains of thought (CoT), enabling the assessment and improvement of the reasoning ability and interpretability of LLMs, while mitigating hallucination generation and enhancing medical safety. Furthermore, as a collection of real-world clinical cases annotated with expert-level clinical diagnoses and supporting evidence, the enhanced dataset serves as a paradigm for advancing specialized disease datasets and medical LLMs in other clinical domains.