[2F1-OS-20] OS-20

10:00 AM - 10:15 AM JST(1:00 AM - 1:15 AM UTC)

[2F1-OS-20-05] A Twin-Track Architecture for Deterministic Verification of Medical Regulatory Documents

〇Naoko Koizumi¹, Jynichi Yasumi¹, Hiroharu Yabu¹, Takashi Goto¹, Tatsuya Gotoda², Kazuya Tanaka^2,3 (1. Hakuhodo-medical Inc., 2. scheme verge, Inc., 3. GRIPS)

Keywords:

Pharmaceutical Promotional Materials,Formalization of Expert Tacit Knowledge,Human-in-the-Loop,Twin-Track Hybrid architecture,Hallucination Elimination

he rapid advancement of generative AI has accelerated and scaled the production of documents, including pharmaceutical promotional materials. However, medical information is life-critical and demands high accuracy and auditability; LLM-only approaches cannot sufficiently guarantee reproducibility of regulatory rules or eliminate hallucinations. This study proposes a verification support architecture based on a Twin-Track Strategy that separates document structure extraction from logic-based verification. Deterministic document structuring and rule verification using logic trees ensure reproducibility and explainability. In practical evaluation, the primary review process was reduced from 10–25 hours to 10–30 minutes. Furthermore, analysis of inter-reviewer variability suggests that delegating explicit rule verification to AI enables human reviewers to focus on higher-order judgments requiring contextual understanding.