Extract data from referral letters

A referral letter moves a patient from one clinician to another, and it carries the clinical context the receiving clinic needs before the first appointment is booked. A Nephrology service at Boston Medical Center reading a new referral pulls the same handful of things off every letter regardless of which practice sent it: who the patient is, who is referring, which specialty is asked to take the case, why, and how soon. Referral coordinators and specialist intake staff read the reason for referral, the urgency level, and any requested appointment date to place the patient in the right slot, while the receiving physician reads the clinical summary, the current medications, the known allergies, and whatever results the referring practice attached. Layouts drift because every EHR exports its own, and a scanned PDF faxed from a two-doctor practice looks nothing like a structured export from a hospital EMR. Urgency is the field that carries the most risk: a routine referral and an urgent two-week referral read differently and route differently. Attached results arrive as LOINC-coded observations with values, interpretations, and reference ranges, medications carry dose instructions, and diagnoses carry ICD-10 codes. In FHIR terms a referral bundles DiagnosticReport, Observation, Patient, and MedicationRequest content into one document, and the receiving system needs each of those pulled into its own field rather than left as prose. Because the letter is PHI, it cannot be pasted into a general chatbot to be parsed. Hand the referral to Talonic and the patient identity, the referring and receiving providers with their NPI where shown, the requested specialty, the reason, the urgency, and the appointment request come back as structured JSON, with attached medications, allergies, and LOINC-coded results kept as their own arrays. A letter dated 2026-05-04 from Dr. Alan Prakash at Mercy General Hospital to Dr. Maria Ochoa, marked urgent for a patient born 1976-03-12 with a creatinine flagged high, lands in the specialist intake queue as data the scheduling desk can act on for a 2026-05-20 slot. Talonic structures what the letter states and links every value to its source region; it does not triage the patient, set the urgency, or reach any clinical decision.

What gets extracted from referral letters

How extraction works for referral letters

Referrals land in a specialist office as EHR exports, secure-message attachments, and scanned faxes, so the same letter shows up under a different layout every time. Talonic classifies the document and maps it to the clinical-document model in the Field Registry, which separates the patient demographics, the referring and receiving providers, the reason and urgency, and the attached medications, allergies, and results into distinct fields rather than one narrative block. LOINC-coded observations keep their code, value, interpretation, and reference range, ICD-coded diagnoses stay tied to their descriptions, and the urgency is read exactly as marked. Each value returns with a confidence score and a pixel-region pointer conforming to DIN SPEC 91491, so a coordinator at a US clinic can check a lab result or an NPI against the source before the patient is booked. The extraction structures the letter; it does not assign urgency or reach a clinical judgment.

Sample extraction

A one-page urgent nephrology referral with an attached basic metabolic panel

{
  "document_date": "2026-05-04",
  "patient.name": "Priya Raman",
  "patient_identifier": "4471902",
  "birth_date": "1976-03-12",
  "referring_provider.name": "Dr. Alan Prakash",
  "referred_provider.name": "Dr. Maria Ochoa",
  "referred_provider.specialty": "Nephrology",
  "reason_for_referral": "Suspected acute kidney injury",
  "urgency_level": "urgent",
  "appointment_requested": true,
  "requested_appointment_date": "2026-05-20",
  "allergies": "Penicillin (rash)",
  "test_results": [
    {
      "observation_code": "2160-0",
      "value": "2.8",
      "unit": "mg/dL",
      "interpretation": "H",
      "reference_range_low": "0.6",
      "reference_range_high": "1.3"
    }
  ]
}

Frequently asked