Smart Triage

System Design

Epic Integration Architecture

End-to-end FHIR R4 pipeline — from referral intake to patient notification

Click any node above to explore the technical details

Inbound FHIR R4 Resources

Every GI referral arriving from Epic carries six FHIR R4 resources. Together, they give Smart Triage everything it needs to auto-classify acuity, flag alarm symptoms, and recommend a sedation pathway — without a coordinator touching a keyboard.

ServiceRequest — The referral itself

{
  "resourceType": "ServiceRequest",
  "id": "gi-referral-1042",
  "status": "active",
  "intent": "order",
  "category": [{
    "coding": [{
      "system": "http://snomed.info/sct",
      "code": "73761001",
      "display": "Colonoscopy"
    }]
  }],
  "code": {
    "coding": [{
      "system": "http://www.ama-assn.org/go/cpt",
      "code": "45378",
      "display": "Diagnostic colonoscopy"
    }]
  },
  "subject": {
    "reference": "Patient/pt-8829"
  },
  "reasonCode": [{
    "text": "Positive FIT test, iron deficiency anemia, Hgb 9.1"
  }],
  "note": [{
    "text": "62M with positive FIT on screening. Hgb 9.1, ferritin 12. Has been putting off colonoscopy for 2 years."
  }]
}

Patient — Demographics

{
  "resourceType": "Patient",
  "id": "pt-8829",
  "name": [{"family": "Johnson", "given": ["Michael"]}],
  "gender": "male",
  "birthDate": "1963-08-15",
  "address": [{"city": "Portland", "state": "OR"}]
}

Condition — Active diagnoses

{
  "resourceType": "Condition",
  "code": {
    "coding": [{
      "system": "http://snomed.info/sct",
      "code": "271737000",
      "display": "Iron deficiency anemia"
    }]
  },
  "subject": {"reference": "Patient/pt-8829"},
  "clinicalStatus": {
    "coding": [{"code": "active"}]
  }
}

Observation — Lab results (CBC, iron studies)

{
  "resourceType": "Observation",
  "code": {
    "coding": [{
      "system": "http://loinc.org",
      "code": "718-7",
      "display": "Hemoglobin"
    }]
  },
  "valueQuantity": {
    "value": 9.1,
    "unit": "g/dL",
    "system": "http://unitsofmeasure.org"
  },
  "interpretation": [{
    "coding": [{"code": "L", "display": "Low"}]
  }]
}

MedicationRequest — Current medications

{
  "resourceType": "MedicationRequest",
  "medicationCodeableConcept": {
    "coding": [{
      "system": "http://www.nlm.nih.gov/research/umls/rxnorm",
      "code": "855332",
      "display": "Omeprazole 20mg"
    }]
  },
  "dosageInstruction": [{
    "text": "Take 20mg by mouth daily"
  }]
}

DiagnosticReport — Imaging/pathology

{
  "resourceType": "DiagnosticReport",
  "code": {
    "coding": [{
      "system": "http://loinc.org",
      "code": "24604-1",
      "display": "CT Abdomen"
    }]
  },
  "conclusion": "No focal lesion. No obstruction.",
  "status": "final"
}

FHIR Data Extraction & Mapping

Before we can parse a referral, we authenticate with Epic using OAuth 2.0 and the SMART on FHIR authorization framework. Here is the step-by-step auth flow, followed by the field mapping table that translates raw FHIR paths into our internal schema.

Authorization Flow

App Registration

Smart Triage is registered in Epic's App Orchard with scopes: patient/*.read system/ServiceRequest.write

Backend Service Auth (Client Credentials)

POST to /oauth2/token with signed JWT assertion. No user interaction required — this is a system-to-system flow.

Access Token Received

Bearer token with 5-minute TTL. Automatically refreshed by the extraction service before expiry.

FHIR Bundle Retrieval

GET /ServiceRequest?_include=*&_revinclude=Observation,Condition,MedicationRequest,DiagnosticReport

Validation & Normalization

Each resource is validated against the FHIR R4 schema. Missing optional fields are flagged, not rejected. Dates normalized to UTC.

Field Mapping: FHIR → Smart Triage Schema

FHIR Resource	Field	Our Schema
`ServiceRequest`	`.reasonCode`	`indication`
`ServiceRequest`	`.note`	`referral_free_text`
`Patient`	`.birthDate`	`age` (calculated)
`Observation` (Hgb)	`.valueQuantity.value`	`hemoglobin`
`Condition`	`.code`	`active_diagnoses[]`
`MedicationRequest`	`.medicationCodeableConcept`	`current_medications[]`

AI-Powered Referral Parsing

The LLM Parser reads unstructured referral notes and extracts structured, clinically actionable data. Below is a real-world example showing the raw text a referring provider might write, and the structured output our model produces.

Raw Input

62 yo male with positive FIT test on routine screening. Has been putting off colonoscopy for 2 years. PMH significant for COPD on home O2 2L, CHF with EF 30%, BMI 38, OSA on BiPAP. On warfarin for DVT history. INR last 2.8. Hgb 10.2.

Parsed Output

Procedure Diagnostic colonoscopy

Indication Positive FIT test

Category Diagnostic

Alarm Symptoms POSITIVE_FIT ANEMIA

BMI 38

ASA Class III-IV

Medications Warfarin (requires bridge)

Comorbidities COPD (home O2), CHF (EF 30%), OSA (BiPAP)

Confidence 0.94

Confidence Scoring Methodology

The confidence score (0.00 to 1.00) is a composite of three factors:

1. Extraction completeness — Were all expected fields successfully parsed? Missing critical fields (indication, procedure) reduce confidence heavily.

2. Code mapping certainty — Did the parsed terms map to known SNOMED/CPT codes, or was fuzzy matching required?

3. Internal consistency — Do the parsed values agree with each other? (e.g., if alarm symptoms are flagged but category says "screening," confidence drops.)

Referrals with confidence below 0.70 are automatically routed to a coordinator for manual review rather than being auto-triaged.

Acuity Tier Assignment & Sedation Logic

Once the LLM Parser produces structured data, the Rules Engine applies two deterministic decision trees: one for acuity tier assignment (how urgently this patient needs to be seen) and one for sedation pathway recommendation (what level of anesthesia support is needed).

Acuity Tier Decision Tree

Has alarm symptoms?

Alarm symptoms: GI bleeding, positive FIT, unexplained weight loss, dysphagia, iron deficiency anemia, abnormal imaging, family history of GI cancer

YES

Active bleeding, acute obstruction, or Hgb < 7?

YES

Tier 1 — Emergent

Schedule within 24-48 hours

Tier 2 — Priority

Schedule within 2 weeks

Surveillance with prior polyp/cancer history?

YES

Tier 3 — Routine

Schedule within 4-6 weeks

Tier 4 — Screening

Average-risk screening. Schedule 6-8 weeks.

Sedation Decision Matrix

Four sedation tiers, evaluated top-down. The first matching rule wins.

MAC Required

BMI ≥ 40
ASA IV+
Home O2
Severe OSA
EF < 35%
ERCP / EUS

MAC Preferred

BMI 35-39
ASA III+
Moderate OSA
Age > 75

NAPS Eligible

Age > 70
BMI 30-34
Mild OSA
Longer procedures

Nurse Sedation

ASA I-II
BMI < 30
No OSA
Routine procedures

Coordinator Workflow Dashboard

The dashboard is the coordinator's command center. It surfaces the right referrals at the right time so no high-risk patient falls through the cracks. Here is what they see when they open Smart Triage each morning.

Key Metrics (Top Bar)

127

Total Referrals

This week

Pending Triage

Needs review

High-Risk Alerts

Immediate action

2.3d

Avg Wait Time

Down from 4.1d

Alert & Routing Workflow

Every referral follows this path. The system handles the grunt work; coordinators focus on clinical judgment.

New Referral Arrives

Inbound FHIR ServiceRequest triggers the pipeline. Extraction begins immediately.

Auto-Triage

LLM Parser + Rules Engine assign acuity tier and sedation pathway. High-confidence results (score ≥ 0.85) proceed automatically.

Human Review

Low-confidence referrals, Tier 1 cases, and edge cases surface in the "Needs Review" queue. Coordinator confirms or overrides the AI recommendation.

Schedule or Escalate

Approved referrals push back to Epic for scheduling. Tier 1 cases alert the on-call GI physician directly via in-basket message.

Outbound FHIR Integration & CDS Hooks

After triage is complete, Smart Triage writes three FHIR resources back to Epic: an updated ServiceRequest with acuity and sedation metadata, a Communication resource that triggers a MyChart message to the patient, and a Flag resource for high-risk cases that surfaces an alert banner in the patient's chart.

ServiceRequest — Updated with triage metadata

{
  "resourceType": "ServiceRequest",
  "id": "gi-referral-1042",
  "status": "active",
  "extension": [{
    "url": "http://smart-triage.io/acuity-tier",
    "valueString": "Tier 2 — Priority"
  }, {
    "url": "http://smart-triage.io/sedation-rec",
    "valueString": "MAC Required"
  }]
}

Communication — MyChart notification to patient

{
  "resourceType": "Communication",
  "status": "completed",
  "category": [{
    "coding": [{
      "code": "notification",
      "display": "MyChart Notification"
    }]
  }],
  "payload": [{
    "contentString": "Your GI procedure has been scheduled. Please review prep instructions in MyChart."
  }],
  "recipient": [{"reference": "Patient/pt-8829"}]
}

Flag — High-risk alert in patient chart

{
  "resourceType": "Flag",
  "status": "active",
  "category": [{
    "coding": [{
      "code": "clinical",
      "display": "Clinical"
    }]
  }],
  "code": {
    "text": "HIGH RISK — Positive FIT + Anemia. Expedite GI evaluation."
  },
  "subject": {"reference": "Patient/pt-8829"}
}

CDS Hooks Integration

CDS Hook

Hook Name smart-triage-alert

Hook Type patient-view

Fires When Provider opens chart of patient with active high-risk Flag

Card Type Warning

Card Message "This patient has a pending high-priority GI referral. Positive FIT + anemia. Recommend expedited colonoscopy."

Patient	Age/Sex	Referred	Indication	Acuity Tier	Sedation	Risk	Status
R.M.	55/M	Feb 10	Screening colonoscopy	Tier 4	MAC	5	Pending
S.P.	48/F	Feb 12	Surveillance (prior adenomas + FHx)	Tier 3	Nurse Sedation	8	Reviewed
J.K.	62/M	Feb 5	Positive FIT, diagnostic	Tier 1	MAC	92	High Risk
M.L.	71/F	Feb 1	IDA + weight loss, diagnostic	Tier 1	NAPS	88	High Risk
A.B.	34/F	Feb 15	Chronic abdominal pain	Tier 3	Nurse Sedation	3	Pending
D.W.	58/M	Feb 8	Surveillance, prior HGD polyp	Tier 2	Nurse/NAPS	45	Pending
L.H.	45/F	Feb 18	Rectal bleeding workup	Tier 3	Nurse Sedation	25	Reviewed
T.G.	67/M	Feb 3	Dysphagia + weight loss, EGD	Tier 1	MAC	95	High Risk

Smart Triage

Epic Integration Architecture

Inbound FHIR R4 Resources

FHIR Data Extraction & Mapping

Authorization Flow

Field Mapping: FHIR → Smart Triage Schema

AI-Powered Referral Parsing

Confidence Scoring Methodology

Acuity Tier Assignment & Sedation Logic

Acuity Tier Decision Tree

Sedation Decision Matrix

Coordinator Workflow Dashboard

Key Metrics (Top Bar)

Alert & Routing Workflow

Outbound FHIR Integration & CDS Hooks

CDS Hooks Integration

Triage Dashboard

Acuity Tier Breakdown

Sedation Distribution

Recent High-Risk Alerts

Clinical Evidence

Clinical Evidence

Clinical Evidence

Sedation Decision Reference

MAC Required

MAC Preferred

NAPS Eligible

Nurse Sedation

Synthetic Data Sandbox

Average-Risk Screening

Surveillance (Polyp History)

Family History CRC

GERD Evaluation

Positive FIT + Anemia

Iron Deficiency Anemia

Abnormal CT Finding

Dysphagia + Weight Loss

Acute Lower GI Bleed

Suspected Obstruction

Complex Sedation Case

IBD Flare

Analysis Pipeline