Traditional manual MedDRA coding in clinical data management (CDM) faces persistent challenges, including suboptimal site data quality, terminology complexity, low efficiency, inconsistent outcomes, frequent dictionary updates, and regulatory timeliness pressures—all of which hinder trial progress and compliance.

This study aimed to identify applicable artificial intelligence (AI) technologies for MedDRA coding, validate their performance in core CDM scenarios, and propose solutions for technical and regulatory hurdles.

Conducted from January 2025 to April 2025, the study utilized real-world adverse event (AE) data from Hengrui Pharmaceutical’s trials: 200 English/200 Chinese AE records (spelling error detection) and 2712 unique AE terms from a Phase II Chinese oncology trial (automated coding). A RAG-AI-Agent integrated framework was developed, incorporating integrated data processing (EDC-MedCoding integration), Large Language Models (LLMs; DeepSeek-R1/V3, Gemini 2.5 Pro, Grok3), Retrieval-Augmented Generation (RAG), and AI-Agent technologies. Performance was evaluated via Precision, Recall, coding time, and workload metrics.

DeepSeek-V3 (task-optimized non-reasoning LLM) achieved 100% Precision in both languages for spelling detection; real-world validation yielded 85.1% Precision and ~ 70% manual review workload reduction. The automated coding system reduced average coding time by 48.7% (8.0 → 3.9 min/record) with 95.8% coverage of AE terms, while maintaining 70%precision against manual gold standards. Compliance was ensured via local deployment and 21 CFR Part 11-aligned audit trails.

The AI-enhanced framework significantly improves CDM efficiency, consistency, and compliance. Despite limitations (oncology-only validation, narrow language scope, synonym library dependence), it provides a replicable model for pharmaceutical digital transformation, accelerating drug development and enhancing patient safety through high-quality clinical data.