Secure Automatic Framework for blind ETL using sYnthetic datasets

Short description of the innovation study

Onboarding hospital databases into federated analytics platforms is slow, manual and blocked by privacy approvals, often taking months. SAFETLY develops a metadata-driven Generative AI framework for "blind ETL": designing and validating ETL pipelines without ever accessing real patient data. The work combines (1) a synthetic database generator that reconstructs realistic, privacy-preserving replicas of hospital databases from schema metadata alone, and (2) an agentic ETL assistant (LangGraph-orchestrated) that maps source schemas to the OMOP Common Data Model and generates transformation code using only synthetic data. Models are trained and fine-tuned on MareNostrum5, using Synthea and MIMIC-IV for training and benchmarking. Expected results: a TRL 6 framework, validated against real hospital metadata, that cuts onboarding from months to 4–5 weeks, plus a blind ETL demo.

Organisations involved: 

Main SME: Promptly – Software Solutions for Health Measures S.A.

Technology expert: Associação Fraunhofer Portugal Research (Fraunhofer Portugal)