Scholarlink

In clinical research, patient recruitment requires a large amount of chart review, and the interpreted data must be converted into a common format (such as FHIR) to support circulation across multi-ple systems. This paper builds a system that covers the entire pipeline from “eligibility-criteria parsing to schema generation to NLP code and Bundle construction to automatic validation”. The method: (1) uses an intermediate schema to decouple information extraction from FHIR-structured data generation into two steps, reducing generation difficulty; (2) automatically samples representative case snippets to assist code generation and improve accuracy; (3) enforces at least two distinct code-generation strategies to improve the recall of information extraction; and (4) uses the Claude agent SDK to generate high-quality code and auto-matically complete testing. The overall system can effectively accomplish the information extraction and FHIR-format data storage required for patient recruitment, thereby supporting inclusion/exclusion screening. The proposed pipeline achieved first place in “CHIP2025 Task 3: Medi-cal NLP Code Generation Evaluation”[1], demonstrating its engineering and deployment value in medical scenarios.

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

A Large Language Model-based System for Automatic Medical NLP Code Generation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jiming Xu,
  • Simin Li,
  • Bo Kang,
  • Jun Yan

摘要

In clinical research, patient recruitment requires a large amount of chart review, and the interpreted data must be converted into a common format (such as FHIR) to support circulation across multi-ple systems. This paper builds a system that covers the entire pipeline from “eligibility-criteria parsing to schema generation to NLP code and Bundle construction to automatic validation”. The method: (1) uses an intermediate schema to decouple information extraction from FHIR-structured data generation into two steps, reducing generation difficulty; (2) automatically samples representative case snippets to assist code generation and improve accuracy; (3) enforces at least two distinct code-generation strategies to improve the recall of information extraction; and (4) uses the Claude agent SDK to generate high-quality code and auto-matically complete testing. The overall system can effectively accomplish the information extraction and FHIR-format data storage required for patient recruitment, thereby supporting inclusion/exclusion screening. The proposed pipeline achieved first place in “CHIP2025 Task 3: Medi-cal NLP Code Generation Evaluation”[1], demonstrating its engineering and deployment value in medical scenarios.