Paradigm Health builds clinical trial management software, and its study conduct module ran on people retyping the trial record out of the EHR. We built the pipeline that carries that data into the EDC on its own, character for character, holding back any field it cannot place with certainty and attaching the route back to the source to every field it delivers.
moved from manual data entry onto AI verification and process improvement
character-level precision on every field the pipeline delivers
traceability, every field back to its document, page and passage
THE PROBLEM
The trial record, moved one character at a time
Study conduct produces the bulk of a trial’s data, and all of it has to reach the EDC exactly. Coordinators opened source documents in a hospital EHR, read the values, and keyed them into electronic data capture field by field. A second person re-checked the first, because in a regulated trial one mistyped character is a finding.
Thirty people did that work. It put a hard ceiling on the business, since taking on more studies meant hiring more coordinators, and it rested the integrity of the trial record on human attention holding steady across thousands of fields a day.
THE PIPELINE
EHR to EDC, character for character
We designed and implemented a precision RAG architecture for a job most retrieval systems are the wrong shape for. Ordinary retrieval is satisfied by a passage that is roughly relevant. Here the requirement is the exact value from the exact document, reproduced without drift, so the architecture is built for exactness at the character level and holds back any field it cannot place with certainty.
On top of it the study conduct pipeline runs end to end: source records leave the EHR, are abstracted into structured fields, and land in the EDC. Every field carries the route back to the document, page and passage that produced it, so any value in the trial record can be followed to its origin by anyone who asks.
THE DOCUMENTS
Reading the pages OCR gives up on
A large share of source material arrives as scans, faxes and photographs. Conventional OCR handles clean type well and degrades fast on everything else: a stamp across the text, handwriting in the margin, a table that loses its columns, a page photographed at an angle.
We implemented an LLM-driven OCR layer over the process already in place, with custom local models augmenting it, so the pipeline reads a page with an understanding of the structure it is looking at. That is what makes document abstraction accurate enough to feed precise data abstraction into the EDC, and it runs locally, which keeps trial source documents inside the environment they belong in.
THE TRANSFORMATION
Thirty people, moved up the stack
Once a pipeline is exact enough to be trusted with the whole task, the open question becomes what the people should be doing. Paradigm answered it by keeping every one of them. Thirty staff whose days went into data entry now verify what the pipeline produces, work the fields it holds back, and improve the process generating them.
That is the shift underneath the figures. A product whose study conduct module depended on manual transcription runs AI-native from the source document forward, and the team that used to be its throughput limit is now its quality layer. Study volume can grow without headcount growing behind it.
Also delivered at Paradigm Health
AI engineering across the study conduct module, architecture through to operation.
- Precision RAG architecture designed and implemented for exact-value retrieval
- End-to-end study conduct pipeline, EHR source records through to EDC entry
- Automated data transfer at character-level precision, field by field
- Traceability from every EDC field back to its source document, page and passage
- LLM-driven OCR implementation over the existing document process
- Custom local model augmentation for abstraction on degraded pages
- Verification workflows for the team now reviewing pipeline output
- AI engineering support across the clinical trial management platform
Discuss your next system.
Start with a free, one-hour engineering review. Leave with a practical next step.