From Study Intent to USDM-Compliant Protocol: What We Learned Building a Multi-Agent AI Digital Protocol Authoring System Under 21 CFR Part 11

About this session

Part of Session 5, Track B: Digital Data Flow (Chair: Erin Muhlbradt, NIH/NCI).

Clinical protocol authoring remains slow, manual, and inconsistently integrated with downstream data systems. This abstract reveals the results of our experiments in building an AI-assisted digital-protocol authoring system aligned with CDISC USDM 4.0 and the ICH M11 Technical Specification (2025).
We implemented multi-agent architecture covering study planning, section authoring, data architecture, and quality validation using the Anthropic Python SDK. The system generated all ICH M11 sections from CT.gov and structured study-input, producing USDM-compliant JSON through an end-to-end automated workflow. A 21 CFR Part 11-compliant audit trail with SHA-256 hash chaining was validated across all generated sections. Practical challenges included mapping free-text clinical intent to USDM entity classes, containing agent hallucinations in regulatory contexts, and maintaining ALCOA+ data integrity throughout AI-assisted generation. Our experiments revealed that agentic-AI has real potential to support DDF adoption at a pharma organization; however human review of all AI-generated content remains a firm requirement.

Speaker

• ST

  Shekhar Thumake

  Ephicacy

  Speaker