Scholarlink

Accurate and efficient extraction of clinical information for urinary incontinence (UI) severity assessment from diverse data sources is a significant challenge. This paper presents a novel framework leveraging large language models (LLMs) for robust named entity recognition (NER) from multisource, heterogeneous, semistructured clinical UI data, including clinical notes, patient-reported outcome measures (PROMs), and electromyography (EMG) signals. The study investigates the efficacy of various prompting strategies and combinations of leading LLMs (DeepSeek-V3, GPT-4o, GPT-4o-mini) for the NER task. Performance was evaluated by comparing UI severity assessments, generated by a unified LLM using the extracted entities, against manually annotated ground truth. Data from a cohort of 287 patients were utilized, with 50 for development and 237 for evaluation. Experimental results demonstrate the framework’s strong capability in accurately extracting key clinical information, offering a promising approach to enhance the efficiency and precision of physician-led UI severity evaluations.

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

LLM-Based Structured Information Extraction for Urinary Incontinence from Multi-modal Clinical Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tianyu Wu,
  • Mingxiang Luo,
  • Xueyan Shen,
  • Shengxiang Liang,
  • Xinyu Wu,
  • Wujing Cao

摘要

Accurate and efficient extraction of clinical information for urinary incontinence (UI) severity assessment from diverse data sources is a significant challenge. This paper presents a novel framework leveraging large language models (LLMs) for robust named entity recognition (NER) from multisource, heterogeneous, semistructured clinical UI data, including clinical notes, patient-reported outcome measures (PROMs), and electromyography (EMG) signals. The study investigates the efficacy of various prompting strategies and combinations of leading LLMs (DeepSeek-V3, GPT-4o, GPT-4o-mini) for the NER task. Performance was evaluated by comparing UI severity assessments, generated by a unified LLM using the extracted entities, against manually annotated ground truth. Data from a cohort of 287 patients were utilized, with 50 for development and 237 for evaluation. Experimental results demonstrate the framework’s strong capability in accurately extracting key clinical information, offering a promising approach to enhance the efficiency and precision of physician-led UI severity evaluations.