Project 08 / 11 — AI PRIOR AUTHORIZATION
08
Prism
AI prior-authorization review, with the evidence attached.
A dashboard for nurses and claims reviewers: it reads the clinical PDF, extracts healthcare entities and asks GPT-4o to judge the case against a payer policy — returning APPROVED, DENIED or ACTION REQUIRED with the exact evidence quote, a criteria checklist and a drafted follow-up email.
“Conservative therapy for 8 weeks documented; imaging confirms…”
“Physical therapy notes not attached.”
“No documented trial of first-line treatment.”
Illustrative decisions — the three outcomes and evidence quote are how Prism answers.
Prior-authorization review means comparing clinical notes against dense payer policies by hand, and missing documentation causes long back-and-forth delays.
Intake
A case is created with patient, policy, provider and SLA; Gold Card providers are auto-approved without AI.
Read
Azure Document Intelligence OCRs the clinical PDF.
Extract
Azure Text Analytics for Health pulls out medical entities.
Judge
GPT-4o compares the note with the policy and returns strict JSON, validated and recovered if malformed.
Act
The workboard shows the decision, checklist, evidence quote, PDF and a request-for-information draft.
Three-way decision
Approved, denied or action required — with the quote that decided it.
Auto-drafted follow-up
Request-for-information emails generated from templates.
SLA queue
Case board with SLA countdowns.
Gold Card bypass
Trusted providers skip AI review entirely.
Split-screen review
Decision and source PDF side by side.
Three-stage pipeline
OCR → healthcare entities → policy reasoning, each stage feeding the next.
server/routes/analyze.py ↗Schema-validated LLM output
Status normalisation and recovery from fenced or nested JSON.
server/services/llm_service.py ↗