MediClaim — Intelligent Prior Authorization System
An end-to-end AI workflow that automates healthcare prior authorization. It involves extracting clinical data from PDFs, images, and text with Claude, scoring each request against payer policies, and routing edge cases through agentic + human review. Auto-approvable decisions drop from 2–5 days to under 5 minutes, with a complete HIPAA/CMS audit trail.
MediClaim — Intelligent Prior Authorization SystemOverview
MediClaim is an end-to-end automated prior authorization (PA) workflow built on the Opus AI workflow engine. It takes a raw medical record — a discharge summary PDF, a scanned lab report, clinical notes, or a CSV batch — and produces a submission-ready authorization decision with a complete audit trail, in minutes instead of days.
The problem
Healthcare prior authorization is slow and expensive: 20–30% of PA requests need resubmission for incomplete documentation, processing averages 2–5 business days, and providers spend ~$35B/year on PA administrative overhead — all while patients wait for care.
The solution — a 5-stage pipeline
Input → Understand → Decide → Review → Deliver
- Ingest multi-format records (PDF, image, text, CSV/Sheets, EHR API).
- Extract patient demographics, diagnosis (ICD-10) and procedure (CPT) codes, and clinical justification using OCR + Claude, with per-field confidence scores.
- Decide via a hybrid engine: deterministic rules (required fields, eligibility, formulary) combined with AI confidence into a composite score (60% rules / 40% AI).
- Route by score — ≥80% auto-approve, 50–80% agentic policy review → human review if needed, <50% human review required.
- Deliver the decision to providers via email (PDF), Google Sheets tracking, and webhooks for EHR integration.
Agentic policy review
For borderline cases, an AI agent evaluates the request against payer guidelines, compares it to similar approved cases, flags missing documentation, and returns a structured verdict (approve / request_clarification / deny) with a safety score, rationale, and specific policy references — keeping a human clinician in the loop for the final call.
Key features
- Multi-format input — PDFs, images, plain text, CSV/Google Sheets, and HL7/FHIR EHR APIs.
- AI extraction — 95%+ accuracy on structured fields (ICD-10, CPT, LOINC) with confidence scoring.
- Hybrid decision engine — deterministic rules + AI judgment combined into an explainable composite score.
- Human-in-the-loop — clinical review UI with accept / reject / request-clarification / override and rationale capture.
- Complete audit trail — input integrity hash, every extracted field, every rule fired, AI reasoning at each stage, and the human decision — for HIPAA/CMS compliance.
- Multi-channel delivery — email, Sheets dashboard, and webhook/EHR integration.
Architecture
A React operator UI drives an Opus workflow (top-level orchestration + intake-normalizer and parallel-processor sub-workflows) whose agent nodes call Claude for extraction and policy reasoning. A Flask mock-API layer stands in for payer/EHR services (patients, policies, formulary, similar cases) in the demo.
Tech stack
Workflow: Opus (opus.ai) · AI: Claude 3.5 Sonnet via Opus agent nodes Frontend: React 18, Vite, Tailwind CSS, React Router, Axios Backend: Flask (Python 3.11), Flask-CORS · Storage: JSON (demo) / PostgreSQL (prod) Deploy: Replit, Docker, Nix · Testing: pytest · Test data: Synthea synthetic patients
Results
| Metric | Value |
|---|---|
| Auto decision time | <5 min (vs. 2–5 days manual) |
| Extraction accuracy | 95–98% structured / 85–92% unstructured |
| Admin burden | ~80% reduction |
| Batch (100 cases) | 5–15 min via parallel processing |
- Auto-approval in <5 min vs. 2–5 days manual
- 95%+ extraction accuracy across ICD-10, CPT & LOINC codes
- Composite 60/40 rules-plus-AI scoring with confidence-based routing
- Agentic payer-policy review with human-in-the-loop for edge cases
- Complete HIPAA/CMS audit trail with full input-to-decision provenance
