2601 - Clinical Deployment of a Supervised Fax Triage and EMR Integration System for Workflow Augmentation in Radiation Oncology

Purpose/Objective(s):

Incoming faxed and emailed records for new radiation oncology consultations require manual review, printing, patient identification, redundancy checking, sorting, and EMR upload. This process is labor-intensive, introduces filing and duplication risk, and can delay consultation readiness. We developed and deployed a locally integrated fax triage and EMR ingestion system within our clinical production environment to augment workflow through supervised decision support.

Materials/Methods:

Baseline manual operator processing time per document was recorded across 10 timed events encompassing the full operational cycle: fax/email retrieval, printing, sorting, and EMR scanning. Documents uploaded per first consultation (N_doc) were recorded in 20 randomly selected cases. The deployed system performs asynchronous backend preprocessing including document ingestion, validation, and normalization; structured patient identifier extraction using multi-method OCR with constrained, locally hosted large language model (LLM) adjudication; EMR cross-verification against ARIA via the Varian FHIR API; suppression of referring provider identifiers; removal of non-clinical pages; separation of multi-patient faxes; and LLM-assisted document-type classification mapped to ARIA filing categories. All inference and processing occur on internal servers without external data transmission. Transactions are logged to an internal database for audit traceability. Automated decisions are presented through a desktop triage interface for human verification prior to upload.

Results:

In 20 randomly selected consultation cases, mean N_doc was 21.3 ± 5.0 documents per case. Processing metrics are summarized in Table 1.

To date, the production system has processed 40 consultation cases (821 documents). Per-document operator time decreased from 4.0 ± 0.9 min (manual, N = 213) to 29.7 ± 8.3 s (supervised triage, N = 821), an 88% reduction (Welch’s t-test, p < 0.001). Backend preprocessing was executed asynchronously and did not contribute to operator interaction time.

Conclusion:

A supervised, production-deployed fax triage and ingestion system was associated with an 88% reduction in measured operator processing time while preserving human verification at each decision point. Future work includes retrieval-augmented model refinement using verified triage outcomes and expanded evaluation of error rates and downstream effects on consultation readiness.

Metric	Manual	Automated	p-value
Primary Workflow Endpoint
Operator time per document	4.0 ± 0.9 min	29.7 ± 8.3 s	< 0.001
Asynchronous Backend Latency
Patient identification	—	3.7 ± 1.2 s	—
Document classification	—	125.3 ± 30.1 s	—