Case study · Registry Abstraction
Emory made reading the record someone else's job — without losing data quality.
Emory
Dr. Madhu Behera
PhD Chief Research Informatics Officer, Emory University
“Health Universe has approached registry automation in a new way. Their agentic platform improves efficiency while keeping clinical quality at the center.”
Accuracy
F1 accuracy vs expert ground truth, with semantic-equivalence matching
Speed
Faster — ~7.5 seconds per document vs 15–30 minutes
Depth
Schema-validated fields per patient, each traceable to source
The challenge
Hundreds of pages read by hand, for a handful of fields.
The same case abstracted by two registrars can come out two different ways, and the queue never shrinks because reading time scales directly with volume.
What they needed
01
To extract registry fields directly from source documents instead of reading every page by hand
02
Deterministic, schema-conforming outputs — a registry field can't be written more than one way
03
Recognition of clinical equivalence (e.g., “Grade 1” = “well differentiated”)
04
Full registrar control and audit trails over every AI-assisted field
The difference
Why they chose Health Universe
A summary can be written more than one way, but a registry field cannot — it has to conform to the schema every time. Only Health Universe produces structured, schema-validated fields that match how a registrar thinks.
Extracts straight from the source
Pathology, notes, and testing read directly, with 150–250+ structured fields pulled per patient.
Conforms to the schema, every time
Outputs are schema-validated and deterministic, with no free-text drift or guesswork.
Keeps registrars in control
Every field is inspectable and routed for review, with full audit trails and clinical-equivalence recognition built in.
In practice
The Approach
Health Universe reads the source documents — pathology, clinical notes, and prior testing — directly, and extracts the registry fields as structured, schema-validated outputs that are deterministic: the same input maps to the same field every time, with no drift. Every field is inspectable and routed to a registrar for review, with full audit trails on the AI-assisted work.
Before Health Universe
After Health Universe
Field entry
Read hundreds of pages, key in by hand
Extracted straight from source documents
Time per document
15–30 min
~7.5 seconds — 120× faster
133-document cohort
50–65 hrs of manual work
Completed in minutes
Consistency
5–15% inter-annotator error
Schema-validated, deterministic outputs
Fields per patient
Limited by reading time
150–250+ schema-validated, traceable fields
Registrar role
Reading + entering
Reviewing inspectable, routed fields
Field entry
Before
Read hundreds of pages, key in by hand
After
Extracted straight from source documents
Time per document
Before
15–30 min
After
~7.5 seconds — 120× faster
133-document cohort
Before
50–65 hrs of manual work
After
Completed in minutes
Consistency
Before
5–15% inter-annotator error
After
Schema-validated, deterministic outputs
Fields per patient
Before
Limited by reading time
After
150–250+ schema-validated, traceable fields
Registrar role
Before
Reading + entering
After
Reviewing inspectable, routed fields
Proof
The Results
Reading the record is no longer the job — reviewing the fields is.
Against expert-curated ground truth across six cancer types, Health Universe reached 93.2% F1 accuracy with semantic-equivalence matching — abstraction quality that holds up against trained registrars — at roughly 7.5 seconds per document versus a 15–30-minute manual baseline, capturing 150–250+ schema-validated fields per patient, each traceable to source.
Accuracy
F1 accuracy vs expert ground truth, with semantic-equivalence matching
Speed
Faster — ~7.5 seconds per document vs 15–30 minutes
Depth
Schema-validated fields per patient, each traceable to source
See these outcomes reproduced on your own charts, live with our team.