[2F1-OS-20] OS-20

10:15 AM - 10:30 AM JST(1:15 AM - 1:30 AM UTC)

[2F1-OS-20-06] Evolutionary Maintenance: A Data-Driven Automated Maintenance Framework for Prescription Drug Document Inspection RulesTackling Combinatorial Explosion: Practical Test-Driven Maintenance of Inspection Rules

〇Junichi Yasumi¹, naoko koizumi¹, hiroharu yabu¹, takashi goto¹, Tatsuya Gotoda², Kazuya Tanaka^2,3 (1. HAKUHODO MEDICAL INC., 2. scheme verge, Inc., 3. GRIPS)

Keywords:

Evolutionary Maintenance,Data-Driven Diagnostics,Statistical Probability Estimation

In our previous work [1], we established auditability for pharmaceutical document inspection through a deterministic rule-based architecture. However, maintaining approximately 360,000 cross-check combinations per document proved unmanageable by human effort alone, as rule enhancements to improve recall structurally increased false positives. To address this maintenance bottleneck, we propose an Evolutionary Maintenance framework comprising three components: (1) an automated diagnostic engine that deterministically classifies false-positive causes into CONFIG, CODE_BUG, and STRUCTURE categories without relying on LLMs; (2) a multi-rater ground-truth labeling system that distinguishes evaluated from unevaluated rules; and (3) a unified diagnostic UI integrating inspection runs with cause analysis. Empirical evaluation on an internal demonstration document with approximately 320 inspection rules showed the diagnostic engine classified 99.1% of false positives as system-correctable, while recall improved from 50.0% to 85.4% over the development period. The labeling system further refined apparent precision from 27.7% to 93.2% by excluding unevaluated rules.