ClinAgent: A Five-Layer Architecture for Autonomous Clinical Trial Statistical Programming

Abstract

Abstract Clinical trial statistical programming requires 12-24 FTE-months for a typical Phase 3 study, producing 100-500 tables, listings, and figures (TLFs) across 8-15 ADaM domains. Modern AI coding agents (Augment Code, Claude Code, Cline, Cursor) demonstrate remarkable reasoning capabilities but lack the domain-specific tools needed for clinical programming: they cannot read SAS datasets, parse ADaM specifications, analyze regulatory-relevant log issues, or generate CDISC-compliant code without extensive guidance. We present ClinAgent , a skill and tool layer that augments any AI coding agent with clinical programming capabilities. Unlike approaches that build custom LLM agents, ClinAgent provides domain expertise as MCP (Model Context Protocol) tools that any compatible agent can invoke. Our “Thin MCP, Thick Skills” design separates minimal data access operations from rich domain logic: Skills package embedded prompts, rule engines, and decision trees encoding expert knowledge; MCP tools provide stateless I/O for SAS datasets, Excel specifications, and log files. Users retain their preferred AI agent while gaining clinical programming tools. We validate ClinAgent’s nine skills (SK-001 Study Setup through SK-009 eSub Packaging) using artifacts from a production Phase 2 cardiovascular study: specifications, SAS programs, and synthetic datasets matching the study structure (11 ADaM domains, 93,239 synthetic observations). All skills pass functional validation. Deterministic components achieve high accuracy: log analysis identifies 1 error and 7 warnings across 10 logs with 100% precision; data validation matches all 56 ADSL variables. Prompt-based specification generation achieves 72.1% derivation accuracy overall, with simple domains exceeding 96%. Our contributions include: (1) an agent-augmentation architecture enabling any AI coding agent to perform clinical programming; (2) nine production-ready skills with embedded prompts and deterministic rule engines; (3) MCP tool implementations for clinical data formats; (4) a validation methodology distinguishing tool correctness from LLM performance; and (5) quantitative results from production study evaluation. Availability: Architecture specifications at https://github.com/yanmingyu92/ClinAgent .

Ähnliche Arbeiten