Reading 4,000 referral letters a week, and knowing when not to

1 The problem

Cadence Care received around 4,000 referral letters a week as scanned PDFs, faxes and emails. Three staff spent their days reading each one, deciding the specialty and urgency, and typing it into the patient system. Backlogs of five days were normal, and an urgent referral could sit unread over a weekend.

2 Our approach

The important design decision was not the model — it was the confidence threshold. We classify specialty and urgency, but anything the model is not confident about goes straight to a human queue rather than being guessed at. Clinicians see the extracted fields alongside the original document and can correct them in one click, and those corrections feed the next evaluation round.

3 What we built

• OCR and layout extraction for scans, faxes and native PDFs
• Specialty and urgency classification with calibrated confidence
• Low-confidence cases routed to a human review queue by design
• Side-by-side review interface with one-click correction
• Full audit trail of model output and every human override
• Weekly evaluation against a clinician-labelled holdout set

4 Technology