Scholarlink

Adverse drug reaction (ADR) is a serious medical issue, so early ADR extraction from Electronic Medical Records (EMRs) is necessary. The majority of current researches on ADR extraction from EMRs are mainly oriented to sentence-level, non-real and single-source data, leading a gap in research and practice. To solve this problem, we propose a novel method LLMADR based on style aligned large language models (LLMs) fine-tuning for ADR extraction from document-level and real multi-source Chinese EMRs. We utilize the comprehension and generation capability of LLMs to accomplish ADR extraction from document-level EMRs where irrelevant information interference and long-distance ADR exist, and we craft prompts to guide LLMs in aligning multi-source EMRs with varying styles before training and reasoning, thereby enhancing the generalization capability of our model. Furthermore, We construct a document-level Chinese ADR dataset CADR from two medical organizations without simplification of EMRs to training and evaluating. Comparative experiments on CADR illustrate that from classification and extraction perspectives, LLMADR performs better than several mainstream models and has better generalization capability.

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

LLMADR: A Novel Method for Adverse Drug Reaction Extraction Based on Style Aligned Large Language Models Fine-Tuning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Huazi Yin,
  • Jintao Tang,
  • Shasha Li,
  • Ting Wang

摘要

Adverse drug reaction (ADR) is a serious medical issue, so early ADR extraction from Electronic Medical Records (EMRs) is necessary. The majority of current researches on ADR extraction from EMRs are mainly oriented to sentence-level, non-real and single-source data, leading a gap in research and practice. To solve this problem, we propose a novel method LLMADR based on style aligned large language models (LLMs) fine-tuning for ADR extraction from document-level and real multi-source Chinese EMRs. We utilize the comprehension and generation capability of LLMs to accomplish ADR extraction from document-level EMRs where irrelevant information interference and long-distance ADR exist, and we craft prompts to guide LLMs in aligning multi-source EMRs with varying styles before training and reasoning, thereby enhancing the generalization capability of our model. Furthermore, We construct a document-level Chinese ADR dataset CADR from two medical organizations without simplification of EMRs to training and evaluating. Comparative experiments on CADR illustrate that from classification and extraction perspectives, LLMADR performs better than several mainstream models and has better generalization capability.