Scholarlink

The scarcity of high-quality annotated data hinders content quality control (QC) for clinical electronic medical records (EMRs). Addressing the CHIP-2025 shared task with only 300 annotated samples, we propose an integrated solution featuring a novel phased training strategy. Our approach combines systematic data preprocessing and multi-granularity prompt engineering, where the most effective “single-rule judgment” method decomposes the complex QC task into 19 independent binary classifications. The two-phase training strategy first learns from 700 pseudo-labeled samples generated via the DeepSeek API, then refines knowledge using the 300 human annotations, with 8 × oversampling to mitigate class imbalance. Our method achieved a score of 0.73 on the Test-B dataset, with phased training as the key contributor, providing a reproducible framework for LLM application in low-resource clinical QC.

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

Leveraging Phased Training and Multi-granularity Prompting with Large Language Models for Few-Shot Quality Control of Electronic Medical Records

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chengfei Li,
  • Chunyu Wang,
  • Bin Liu,
  • Yunjie Zhang,
  • Liwei Cai,
  • Bei Zhao

摘要

The scarcity of high-quality annotated data hinders content quality control (QC) for clinical electronic medical records (EMRs). Addressing the CHIP-2025 shared task with only 300 annotated samples, we propose an integrated solution featuring a novel phased training strategy. Our approach combines systematic data preprocessing and multi-granularity prompt engineering, where the most effective “single-rule judgment” method decomposes the complex QC task into 19 independent binary classifications. The two-phase training strategy first learns from 700 pseudo-labeled samples generated via the DeepSeek API, then refines knowledge using the 300 human annotations, with 8 × oversampling to mitigate class imbalance. Our method achieved a score of 0.73 on the Test-B dataset, with phased training as the key contributor, providing a reproducible framework for LLM application in low-resource clinical QC.