LLM4ODM: Synthetic Clinical Study Data Generation for CDISC ODM

Reliable study databases require realistic data for validation and testing. Manual creation is slow, and rule-based generators often yield clinically implausible cases. We present LLM4ODM, a dockerized web system that uses large language models to generate synthetic, context-aware clinical study data directly from CDISC ODM v1.3 metadata. ODM definitions (e.g., ItemDefs) are transformed into structured prompts, processed by Gemini 1.5 Flash, and returned as JSON subject records. A validation layer ensures schema compliance, data types, branching logic, and temporal coherence. Across ten ODM datasets, including three from Heidelberg University Hospital, LLM4ODM achieved 100% schema adherence, >78% reduction in manual effort, and better clinical plausibility than rule-based methods based on structured expert review. All code, Docker images, prompts, and ODM files will be openly available.