Scholarlink

Clinical trial patient enrollment and exclusion criteria are typically expressed in unstructured natural language, containing complex nested logical structures, ambiguous temporal and numerical expressions, and synonymous medical concepts. This complexity makes manual conversion to executable screening rules inefficient. Additionally, healthcare data must conform to HL7 FHIR standards to enable cross-system interoperability. This paper presents a Prompt-Driven Program Synthesis approach for the CHIP2025 Task 3: Medical NLP Code Generation Evaluation. Using 51 screening criteria and corresponding patient case data, we design a structured prompt engineering framework with explicit constraints for class structure, function signatures, and FHIR Bundle format requirements. This approach guides GPT-5 to generate executable Python screening code and FHIR Bundles containing Condition, Observation, and ResearchSubject resources in transaction format. Through systematic experimentation, we identified that structured prompt design with independent session management yields optimal performance. Our method achieved a score of 0.3344 on the Test-B leaderboard.

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

Prompt-Driven Program Synthesis for Clinical Trial Screening Criteria: From Natural Language to Executable FHIR Code Generation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chengfei Li,
  • Chunyu Wang,
  • Bin Liu,
  • Yunjie Zhang,
  • Bei Zhao,
  • Wenkun Chen,
  • Shengyu Shao

摘要

Clinical trial patient enrollment and exclusion criteria are typically expressed in unstructured natural language, containing complex nested logical structures, ambiguous temporal and numerical expressions, and synonymous medical concepts. This complexity makes manual conversion to executable screening rules inefficient. Additionally, healthcare data must conform to HL7 FHIR standards to enable cross-system interoperability. This paper presents a Prompt-Driven Program Synthesis approach for the CHIP2025 Task 3: Medical NLP Code Generation Evaluation. Using 51 screening criteria and corresponding patient case data, we design a structured prompt engineering framework with explicit constraints for class structure, function signatures, and FHIR Bundle format requirements. This approach guides GPT-5 to generate executable Python screening code and FHIR Bundles containing Condition, Observation, and ResearchSubject resources in transaction format. Through systematic experimentation, we identified that structured prompt design with independent session management yields optimal performance. Our method achieved a score of 0.3344 on the Test-B leaderboard.