ClearPath — Clinical Prior Authorization Agent

Prior authorization costs US hospitals $13B/year in administrative overhead and delays patient care by an average of 14 days. ClearPath is an AI agent that processes authorization requests in under 30 seconds — from diagnosis extraction to approval prediction — using Azure OpenAI GPT-4.1.

Built by Senay Yakut at Musa Labs Hackathon SF — Enterprise Agents (Feb 27, 2026)

---

The Problem

Every time a doctor orders a procedure, insurance companies require prior authorization — a formal request proving medical necessity. Today this process is:

• Slow: 3–14 day average turnaround
• Expensive: $13B/year in US hospital admin costs
• Error-prone: 34% of requests initially denied due to incomplete documentation
• Manual: 45 minutes of staff time per request reading policy PDFs and filling forms

What ClearPath Does

ClearPath is a multi-step AI agent that automates the entire prior authorization workflow:

Patient Case Input
       │
       ▼
┌──────────────────────────────────────────────────┐
│           Azure OpenAI GPT-4.1 Agent             │
│                                                  │
│  ┌─────────────────┐  ┌──────────────────────┐   │
│  │ 1. Extract       │  │ 2. Policy Lookup     │   │
│  │    Diagnosis     │  │    (RAG Search)       │   │
│  │    ICD-10 / CPT  │  │    8 policy docs      │   │
│  └────────┬────────┘  └──────────┬───────────┘   │
│           │                      │               │
│  ┌────────▼────────┐  ┌──────────▼───────────┐   │
│  │ 3. Draft Auth   │  │ 4. Predict Approval  │   │
│  │    Request      │  │    Confidence Score   │   │
│  │    Letter       │  │    Gap Analysis       │   │
│  └────────┬────────┘  └──────────┬───────────┘   │
│           │                      │               │
│           └──────────┬───────────┘               │
│                      │                           │
│           ┌──────────▼───────────┐               │
│           │ 5. Route Decision    │               │
│           │    ≥70% → Submit     │               │
│           │    <70% → Human HITL │               │
│           └──────────────────────┘               │
└──────────────────────────────────────────────────┘
       │
       ▼
Output: Auth letter + confidence score + audit trail

Key Features

Feature	Description
Multi-step Agent	5 tools called sequentially with full context passing via GPT-4.1 function calling
RAG Pipeline	TF-IDF vector search over 8 payer policy documents across 3 insurance companies (UHC, Aetna, BCBS)
Confidence Scoring	Quantified approval prediction with strengths, risks, and missing documentation analysis
Human-in-the-Loop	Automatic routing to clinical reviewers when confidence < 70% — no autonomous decisions on borderline cases
Authorization Letters	Auto-generated formal request letters with clinical justification mapped to payer requirements
Full Audit Trail	Every tool call logged with inputs and outputs for HIPAA compliance and transparency

Sample Cases

The app ships with 5 cases spanning the full approval spectrum:

Patient	Procedure	Payer	Expected	Why
Maria Rodriguez, 67F	Total Knee Replacement	United Healthcare	Approved ~90%	Complete documentation, Grade IV OA, 18 months conservative treatment
Robert Kim, 71M	MRI Brain	Blue Cross Blue Shield	Approved ~85%	Red flag symptoms, elderly patient, CT completed
James Thompson, 52M	MRI Lumbar Spine	Aetna	Borderline ~65%	Only 8 weeks conservative treatment, no neurological findings
David Park, 43M	Total Knee Replacement	United Healthcare	Denied ~25%	Grade II only, BMI 46, no PT/injections/clearance
Susan Chen, 38F	Humira (Adalimumab)	Aetna	Denied ~20%	No DMARDs tried, no TB/Hep screening, no disease activity score

Tech Stack

Component	Technology	Role
AI Engine	Azure OpenAI GPT-4.1	Multi-step clinical reasoning with function calling
Backend	FastAPI (Python)	API server + agent orchestration loop
RAG	TF-IDF + Cosine Similarity	Vector search over payer policy documents
Frontend	React (CDN)	Clinical dashboard with case submission + review
Policies	8 documents, 3 payers	UHC, Aetna, BCBS across 4 procedure categories

Production upgrade path: Azure AI Search (replace TF-IDF), Cosmos DB (persistence), Azure Key Vault (PHI encryption), Azure Functions (serverless scaling).

Quick Start

# Clone the repo
git clone https://github.com/SenayYakut/ClinicalPriorAuthAgent.git
cd ClinicalPriorAuthAgent

# Install dependencies
pip install -r requirements.txt

# Configure environment
cp .env.example .env
# Edit .env with your Azure OpenAI credentials

# Run
python main.py

Open http://localhost:8080 in your browser.

Demo Flow

1. Open http://localhost:8080
2. Click a sample case — try "Maria Rodriguez" (strong) or "David Park" (weak)
3. Click Submit for Prior Authorization
4. Watch the agent process in ~10-30 seconds
5. Review: confidence score, auth letter, agent trace, gap analysis
6. Weak cases route to the Review Queue tab for human approval/denial

Project Structure

ClinicalPriorAuthAgent/
├── main.py              # FastAPI backend + GPT-4.1 agent loop + API endpoints
├── rag_engine.py        # TF-IDF RAG pipeline over payer policy documents
├── payer_policies.py    # Policy database, ICD-10/CPT codes, 5 sample cases
├── static/index.html    # React frontend dashboard
├── presentation.html    # Hackathon presentation slides (arrow keys to navigate)
├── requirements.txt     # Python dependencies
├── .env.example         # Environment variable template
└── .gitignore           # Excludes .env, __pycache__, venv

Responsible AI

• Human-in-the-loop: Agent never auto-approves uncertain cases (< 70% confidence)
• Grounded retrieval: RAG ensures agent references actual policy documents, not hallucinated requirements
• Transparency: Full tool execution trace visible for every case
• Gap analysis: Proactively flags missing documentation before submission
• Audit compliance: All inputs, outputs, and decisions logged for regulatory review

API Endpoints

Method	Endpoint	Description
GET	/	Frontend dashboard
GET	/api/sample-cases	List sample cases
POST	/api/cases	Submit a case for processing
GET	/api/cases	List all processed cases
GET	/api/cases/{id}	Get case details
GET	/api/review-queue	Cases pending human review
POST	/api/review	Submit human review decision
GET	/api/rag-search?query=...	Direct RAG search over policies
GET	/api/stats	Dashboard statistics

---

Built with Azure OpenAI GPT-4.1 at Musa Labs Hackathon SF 2026