--- title: 'Stanford Health Care: FastFax urgent referral fax queue routing' slug: stanford-fastfax-referral-triage stable_id: ba3aaddbae2a7271 company: Stanford Health Care function_code: clinical_ops pattern_codes: - queue_elimination - intake_redesign - step_before_decision evidence_strength: mixed publication_tier: showcase freshness: current reviewed_at: '2026-09-06' updated_at: '2026-09-06' source_quality_summary: >- 1 primary operational source and 2 peer-reviewed publications; workflow metrics are scoped to supported forms and urgent-fax processing. caveat_summary: >- The 78% recall applies only to marked-urgent referrals on the supported standard form. It is not recall across all 2,332 marked-urgent referrals and is not a measure of clinical triage. Stanford's page labels the 231 nonurgent faxes as an 18% false positive rate, but 231 divided by 1,290 is the false-discovery share, complementary to the reported 82% precision; this record preserves the reported counts and precision without generalizing to clinical accuracy. This is a contemporaneous comparison between differently prioritized queues, not a randomized causal estimate, and it measures processing at this workflow step rather than referral-to-appointment time or patient outcomes. The published summary does not establish referral-to-appointment reduction, improved clinical outcomes, or improved collections: - human-still-decides - queue-eliminated - regulated-autonomy bundle_version: 1.0.0 bundle_fingerprint: sha256:8bbbaa076bde20fdad72e1c25c656436643f25efe84d187d0ec482f60acddc98 canonical_url: https://brianletort.ai/transformations/stanford-fastfax-referral-triage --- # Stanford Health Care: FastFax urgent referral fax queue routing Referral staff do not have to open every incoming fax before provider-declared urgent referrals can enter an expedited work queue. On supported standardized cover sheets, a classifier can detect the referring provider's checked urgent box and select the queue without making a clinical urgency judgment. Function: Clinical operations. Patterns: Queue elimination; Intake redesign; AI prepares, human decides. Evidence: mixed. Freshness: current. Reviewed: 2026-09-06. Updated: 2026-09-06. Source quality: 1 primary operational source and 2 peer-reviewed publications; workflow metrics are scoped to supported forms and urgent-fax processing. ## Before 1. **Referring provider** — Faxed an external referral to Stanford Health Care and, when applicable, marked the referral urgent on the cover sheet. (control: The referring provider, not Stanford's classifier, determined whether the referral was urgent.) 2. **RightFax and OnBase** — Moved incoming faxes directly into one shared OnBase intake queue containing both urgent and routine referrals. (control: No automated urgent-versus-routine queue split occurred.) 3. **Enterprise Contact Center referral management staff** — Opened and read each fax cover sheet to find referrals that the referring provider had marked urgent. (control: Human review was the first mechanism for identifying the provider's urgency mark.) 4. **Referral staff and clinical schedulers** — Processed the referral, handled missing or nonstandard information, and made downstream scheduling and clinical workflow decisions. (control: Operational and clinical authority remained with Stanford staff.) ## After 1. **Referring provider** — Faxes an external referral and determines urgency by checking the urgent box on the Stanford standardized cover sheet or by using another form or notation. (control: Urgency remains a provider declaration; FastFax does not assess the patient's clinical condition.) 2. **MuleSoft connector** — Monitors the RightFax queue and sends each new fax to the Stanford FastFax API running on Google Cloud. (control: The connector inserts the classification service between RightFax and OnBase.) 3. **FastFax image preprocessor and machine-learning checkbox classifier** — Processes the cover-sheet image and detects whether the urgent checkbox is marked at a known location on a supported standardized form. (control: The model detects a form mark; it does not infer clinical urgency, diagnose, prioritize from clinical content, or override the referring provider.) 4. **MuleSoft connector and OnBase** — Use the API result to place a recognized provider-marked urgent fax in the urgent OnBase queue and other or unrecognized faxes in the routine OnBase queue. (control: Queue selection accelerates human attention but does not authorize downstream clinical action.) 5. **Enterprise Contact Center referral management staff** — Work the urgent and routine queues, verify and process referrals, and handle false positives, unsupported forms, handwritten urgency, and other exceptions. (control: Humans retain referral-processing responsibility; every fax remains in a human-worked queue.) 6. **Clinical schedulers and clinical teams** — Make scheduling, escalation, and any clinical decisions after referral processing. (control: FastFax has no clinical decision rights.) ## Decision rights The referring provider determines and marks urgency. FastFax detects the mark and selects an OnBase queue. Enterprise Contact Center referral staff own referral review and processing, while clinical schedulers and clinical teams retain downstream scheduling, escalation, and clinical decisions. ## Exception path Unsupported or old forms, handwritten urgent notations, and differently positioned checkboxes may not be recognized and remain in the routine human-worked queue. This preserves eventual human review but does not prevent delay for a missed provider-marked urgent referral. Referral staff handle false positives and other intake exceptions; public sources do not document a separate automated failover or downtime protocol. ## Outcomes - **Recall among marked-urgent referrals using the supported standard cover sheet** (verified): The early operational snapshot contained 1,352 provider-marked urgent referrals using the standard cover sheet. → FastFax routed 1,059 of the 1,352 to the urgent queue, reported as 78% recall.. The 78% recall applies only to marked-urgent referrals on the supported standard form. It is not recall across all 2,332 marked-urgent referrals and is not a measure of clinical triage. - **Precision of the urgent queue in the early operational snapshot** (verified): Before FastFax, urgent and routine referrals entered one shared queue, so there was no classifier-created urgent queue. → FastFax routed 1,290 faxes to the urgent queue; 1,059 were provider-marked urgent and 231 were nonurgent, reported as 82% precision.. Stanford's page labels the 231 nonurgent faxes as an 18% false positive rate, but 231 divided by 1,290 is the false-discovery share, complementary to the reported 82% precision; this record preserves the reported counts and precision without generalizing to clinical accuracy. - **Mean queue processing time in the early operational snapshot** (verified): Contemporaneous routine-queue faxes averaged 152 minutes. → Urgent-queue faxes averaged 38 minutes.. This is a contemporaneous comparison between differently prioritized queues, not a randomized causal estimate, and it measures processing at this workflow step rather than referral-to-appointment time or patient outcomes. - **Average urgent-referral processing time in the peer-reviewed case** (reported): About 33 hours before FastFax. → About one hour after FastFax; the accompanying editorial describes the post-intervention average as about 63 minutes.. The published summary does not establish referral-to-appointment reduction, improved clinical outcomes, or improved safety. The approximately one-hour measure should not be conflated with the earlier 38-minute urgent-queue snapshot or applied to FastFax 2.0. - **Effective coverage of all provider-marked urgent referrals in the early snapshot** (inferred): The snapshot reported 2,332 total provider-marked urgent referrals, including unsupported forms. → The disclosed counts imply that 1,059 divided by 2,332, or approximately 45%, of all provider-marked urgent referrals reached the urgent queue.. This is a catalog inference from disclosed counts, not a metric stated by Stanford. It assumes the reported 1,059 true positives are the urgent referrals reaching the urgent queue and highlights the material coverage loss from unsupported forms. ## Executive lesson A narrow model can materially compress a queue without making the consequential decision. FastFax worked by detecting a provider's explicit priority mark and moving matching work forward, while referral staff and clinicians retained action and authority. ## Anti-pattern Calling checkbox detection clinical triage, reporting 78% recall as if it covered all marked-urgent referrals, or using original FastFax metrics to imply referral-to-appointment, safety, patient-outcome, or FastFax 2.0 performance. ## Questions for leaders - Is the model detecting an authorized human's explicit decision, or making a new consequential judgment? - What percentage of the full priority population is covered after unsupported forms and other out-of-scope inputs are included? - Which form versions, layouts, and handwritten conventions are outside the validated boundary? - What happens operationally when a priority item is missed, and how quickly can a human recover it? - Are classifier recall, queue precision, queue-processing time, and end-to-end patient access being measured and communicated as distinct outcomes? - Who retains authority for downstream scheduling, escalation, and clinical action? ## Sources - [FastFax Uses Machine Learning to Triage Urgent Patient Referrals](https://med.stanford.edu/seal/explore-seal-app/fast-fax.html) — Stanford Emerging Applications Lab, Stanford Medicine - [Optimizing Enterprise Referral Processing through Automated Fax Triage](https://doi.org/10.1056/CAT.25.0273) — NEJM Catalyst Innovations in Care Delivery - [Culture and Leadership Are Essentials for Innovation](https://catalyst.nejm.org/doi/full/10.1056/CAT.26.0177) — NEJM Catalyst Innovations in Care Delivery