The task
Convert differently organized safety forms into a consistent structured representation. Separate document reading, normalization, field alignment and evaluation so failures can be inspected.
The experiment
The workflow includes filling synthetic PDF report forms before extraction. It uses 130 synthetic reports: 30 CIOMS and 100 MedWatch forms. Python workflows compare traditional machine learning and transformer-based extraction in an academic, synthetic-data study.
The boundary
Synthetic-data performance cannot establish readiness for real clinical or regulated use. This is a learning prototype with source-to-field traceability and a clear human-review requirement.