Interoperability · Document Workflow
FaxBridge
A fax can arrive digitally and still leave its information trapped.
FaxBridge is an AI-assisted clinical document intake prototype that explores how information inside faxed PDFs could be extracted, structured as FHIR resources and reviewed before reaching a patient record.
01 — The problem
Healthcare still depends heavily on faxed clinical information.
Receiving a PDF electronically solves transmission.
It does not make the information inside that PDF structured, searchable or interoperable.
Every step below still depends on a person reading the document and re-entering what it contains.
Fax / PDF
Human reads
Human identifies patient
Human enters information
Document filed
Could more of that work become structured without allowing AI-generated information to enter the record unchecked?
02 — The workflow
What FaxBridge does.
PDF received
Document classification
Patient information extraction
Patient matching
FHIR resource generation
Human verification
Approved information ready for downstream use
Each step is a place the prototype can stop. Classification and extraction run on their own; patient matching and verification are the two points where the system is designed to hand back to a person.
The two screenshots below show those two safeguards: first whether the patient can be identified at all, then whether the extracted values can be trusted.
03 — Patient identity
The system needs to be able to say: “I don’t know.”
Patient matching is one of the highest-risk steps in any document intake workflow. An incorrect match attaches clinical information to the wrong patient record.
Clear match
Proceed to review
Ambiguous match
Surface candidate records for human selection
No match
Hold for review and flag possible misdirected document
FaxBridge does not automatically create a patient record when identity is uncertain. That is deliberate.
04 — Human review
Extraction and verification are different jobs.
The prototype keeps the original source document visible while presenting extracted information for review.
The goal is to let someone verify ordinary clinical fields against their source without requiring them to read raw FHIR resources.
The machine can speak FHIR.
The reviewer should not have to.
05 — Consequence matters
Confidence alone is not enough.
A system may be equally confident about two extracted values that carry very different consequences.
A phone number and a medication dosage should not automatically be treated as equally safe.
High-consequence information needs stronger verification.
Requires stronger verification
- Medication / dosage
- Allergies
- Abnormal findings
- Patient identity
- Information affecting treatment
06 — What I learned
The difficult part was not getting information out of the document. It was deciding when the system should refuse to act.
FaxBridge started as a FHIR problem.
It became a trust problem.
- What should the software automate?
- What should it escalate?
- How should uncertainty be represented?
- And how can a reviewer always trace structured information back to its source?
Limitations
- Independent prototype
- Uses synthetic / demo information
- Not connected to live clinical systems
- Not clinically validated
- FHIR generation does not equal universal EMR integration
- Requires privacy, security and workflow validation before any real clinical deployment