Scholarlink

Abstract <p>Processing clinical unstructured data is crucial in disease research. However, existing approaches often rely on domain-specific annotated data and model fine-tuning, facing challenges such as complex workflows, lack of systematic prompt design, and data security risks. To address these issues, this study proposes a zero-shot data processing paradigm based on large language models (LLMs) and the CMQOE (Concept–Method–Quality–Output–Example) prompt framework. This framework systematically integrates concept definitions, task modeling, quality constraints, output specifications, and comparative examples to guide general-purpose LLMs in semantic parsing and structured transformation of medical texts—without requiring task-specific training. Using outpatient medical records of hypertension patients as a case study, experiments demonstrate that the CMQOE framework significantly outperforms conventional prompt methods across multiple LLMs. For instance, with Qwen2.5-14B, it achieved an accuracy of 99.70%, a precision of 98.85%, a specificity of 99.89%, and an F1-score of 98.29%. Moreover, the framework exhibited strong generalization capability on a cross-dataset entity recognition task (CCKS 2019), performing comparably to fully-supervised models and even surpassing them in certain models such as the Qwen series. This study provides a lightweight, secure, and standardized approach for processing clinical unstructured data.</p> Graphical abstract <p>CMQOE Framework for Clinical Unstructured Data Processing. (Left) Concept diagram of the paradigm for processing unstructured data. (Right) CMQOE (Concept-Method-Quality-Output-Example) prompt framework design. Applied to hypertension patient records, the framework enabled Qwen2.5-14B to achieve 99.70% accuracy and 98.29% F1-score, outperforming conventional prompts and showing strong generalization on cross-dataset tasks (CCKS 2019). This provides a lightweight, secure, and standardized approach for clinical text processing without task-specific training</p> <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Processing unstructured clinical notes with LLMs: applying the CMQOE framework for hypertension

  • 期刊论文

  • 发表日期：

    2026年2月11日

查看全文
  • An Gong,
  • Bitian Fan,
  • Yong Liu,
  • Duan Wang,
  • Zhifei Tong,
  • Xingtong Wei,
  • Anxuan Jia,
  • Zhouqi Zhang,
  • Shuhui Wu

摘要

Abstract

Processing clinical unstructured data is crucial in disease research. However, existing approaches often rely on domain-specific annotated data and model fine-tuning, facing challenges such as complex workflows, lack of systematic prompt design, and data security risks. To address these issues, this study proposes a zero-shot data processing paradigm based on large language models (LLMs) and the CMQOE (Concept–Method–Quality–Output–Example) prompt framework. This framework systematically integrates concept definitions, task modeling, quality constraints, output specifications, and comparative examples to guide general-purpose LLMs in semantic parsing and structured transformation of medical texts—without requiring task-specific training. Using outpatient medical records of hypertension patients as a case study, experiments demonstrate that the CMQOE framework significantly outperforms conventional prompt methods across multiple LLMs. For instance, with Qwen2.5-14B, it achieved an accuracy of 99.70%, a precision of 98.85%, a specificity of 99.89%, and an F1-score of 98.29%. Moreover, the framework exhibited strong generalization capability on a cross-dataset entity recognition task (CCKS 2019), performing comparably to fully-supervised models and even surpassing them in certain models such as the Qwen series. This study provides a lightweight, secure, and standardized approach for processing clinical unstructured data.

Graphical abstract

CMQOE Framework for Clinical Unstructured Data Processing. (Left) Concept diagram of the paradigm for processing unstructured data. (Right) CMQOE (Concept-Method-Quality-Output-Example) prompt framework design. Applied to hypertension patient records, the framework enabled Qwen2.5-14B to achieve 99.70% accuracy and 98.29% F1-score, outperforming conventional prompts and showing strong generalization on cross-dataset tasks (CCKS 2019). This provides a lightweight, secure, and standardized approach for clinical text processing without task-specific training