Zero-Interface Medical Biller

Autonomous medical billing automation using Amazon Nova AI

An intelligent system that automatically detects and corrects medical billing errors (upcoding and unbundling), then submits corrected claims to insurance portals without human intervention.

Hackathon Criteria

Multimodal AI

• Nova Pro for multimodal document parsing (scanned images → structured data)
• Processes scanned medical documents in JPG and PNG formats
• Extracts structured billing data from unstructured visual sources
• Medical terminology recognition (ICD-10, CPT codes)
• 100% success rate on test images (4/4 parsed correctly)

Agentic AI

• Nova Act for autonomous browser automation
• Navigates insurance portals without human guidance
• Fills forms, submits claims, extracts confirmation numbers
• Intelligent retry logic with exponential backoff
• Uses AWS IAM authentication via @workflow decorator

Features

• 100% Autonomous Operation — No human intervention for standard claims
• Multimodal Parsing — Nova Pro extracts data from scanned prescriptions
• NCCI Rules Engine — 40+ bundling rules for upcoding and unbundling detection
• Browser Automation — Nova Act fills and submits insurance portal forms
• Escalation Queue — Complex cases automatically routed to human review via S3
• End-to-End Pipeline — Parse → Detect → Correct → Submit in one flow

System Architecture

Scanned Image / JSON Record
        │
        ▼
  Parse (Nova Pro)  →  Structured claim data
        │
        ▼
  Detect Errors (NCCI Rules)  →  Upcoding / Unbundling violations
        │
        ▼
  Apply Corrections  →  Corrected claim
        │
        ▼
  Submit (Nova Act)  →  Confirmation number
        │
        ▼
  Results / Escalation

Nova Model Usage

Model	Component	Purpose
Nova Pro	Document Parser	Multimodal parsing of scanned medical images
Nova Lite	Error Detector	Intelligent error detection with NCCI rules
Nova Act	Browser Automation	Autonomous portal form submission

Demo Results

Tested on 20 synthetic medical claims + 4 scanned images:

• ✅ 100% Claim Success Rate (20/20 claims processed)
• ✅ 100% Image Parsing Rate (4/4 images parsed)
• ✅ 10 Errors Detected (5 upcoding + 5 unbundling)
• ✅ 10 Errors Corrected (100% correction rate)
• ✅ Escalation Queue functional for complex cases

Prerequisites

• Python 3.11+
• AWS Account with Bedrock access (Nova Pro, Nova Lite, Nova Act)
• AWS CLI configured with appropriate credentials

Quick Start

1. Install Dependencies

pip install -r requirements.txt

2. Configure Environment

Create .env file:

# AWS Configuration
AWS_REGION=us-east-1
AWS_ACCESS_KEY_ID=your_key
AWS_SECRET_ACCESS_KEY=your_secret

# Nova Model IDs
MODEL_NOVA_2_OMNI=us.amazon.nova-pro-v1:0
MODEL_NOVA_2_LITE=us.amazon.nova-lite-v1:0

# S3 Configuration
S3_BUCKET=nova-zero-medical-biller

# Mirror Portal (test portal)
PORTAL_URL=http://localhost:3000
PORTAL_USERNAME=admin
PORTAL_PASSWORD=admin123

3. Start Mirror Portal

python mirror-portal/server.py

Portal runs at http://localhost:3000

4. Run End-to-End Demo

python test_end_to_end.py

This processes claims through the full pipeline with real Nova Act submission.

5. Run Full Verification (20 claims)

python test_local_verification.py

Processes all 20 synthetic claims and generates detailed reports in data/results/.

6. Test Multimodal Image Parsing

python test_image_parsing.py

Demonstrates Nova Pro parsing 4 scanned medical images with 100% success rate.

Project Structure

.
├── src/
│   ├── parser/              # Document parsing (Nova Pro multimodal)
│   ├── validator/           # Error detection (NCCI rules engine)
│   ├── rag/                 # RAG engine with fallback rules
│   ├── escalation/          # Escalation queue (S3-based)
│   └── orchestrator/        # Main pipeline coordinator
├── data/
│   ├── synthetic/           # 20 synthetic test claims
│   ├── scanned/             # 4 scanned medical images (JPG)
│   ├── ncci-rules/          # 40+ NCCI bundling rules (JSON)
│   └── results/             # Processing results
├── mirror-portal/           # Test insurance portal (localhost:3000)
├── demo-dashboard/          # Live demo UI
├── docs/                    # Architecture documentation
├── nova_act_workflow.py     # Nova Act workflow definition
├── test_end_to_end.py       # End-to-end demo script
├── test_local_verification.py  # Full pipeline verification
├── test_image_parsing.py    # Multimodal image parsing demo
└── demo-visual.py           # HTML report generator

Error Detection

Upcoding Detection

• Compares billed CPT codes against visit documentation
• Detects when high-complexity codes are billed for simple visits
• Example: 99215 (high complexity) billed for a brief follow-up → corrected to 99213

Unbundling Detection

• Checks procedure code pairs against 40+ NCCI bundling rules
• Detects when procedures should be bundled together
• Supports modifier exceptions (-25, -59, -76, -77, -91)
• Example: 99214 + 36415 billed separately → flagged per NCCI rules

Escalation Queue

Claims are escalated to human review when:

• High-severity errors with low confidence
• Claims exceeding value thresholds ($1000+ = HIGH priority)
• Complex cases requiring expert judgment

Escalated claims are persisted to S3 with priority level, human-readable summary, and full error history.

Configuration

Required Environment Variables

• AWS_REGION — AWS region for Bedrock (default: us-east-1)
• MODEL_NOVA_2_OMNI — Nova Pro model ID for multimodal parsing
• MODEL_NOVA_2_LITE — Nova Lite model ID for error detection
• S3_BUCKET — S3 bucket for escalation queue

Optional Environment Variables

• PORTAL_URL — Mirror Portal URL (default: http://localhost:3000)
• PORTAL_USERNAME / PORTAL_PASSWORD — Portal credentials

Troubleshooting

Bedrock Access Denied

Ensure your AWS credentials have Bedrock permissions:

{
  "Effect": "Allow",
  "Action": ["bedrock:InvokeModel"],
  "Resource": "*"
}

Mirror Portal Not Starting

Check if port 3000 is available. Kill any existing process on that port.

Nova Act Playwright Error

If you see "Playwright Sync API inside asyncio loop", run the demo scripts directly (not from async contexts):

python test_end_to_end.py

Nova Act Falls Back to Simulation

Ensure NOVA_ACT_API_KEY is NOT set in .env — the system uses AWS IAM authentication via the @workflow decorator instead.

Links

• Architecture Documentation
• Demo Dashboard

License

MIT License

Acknowledgments

Built for the Amazon Nova AI Hackathon demonstrating:

• Multimodal AI: Nova Pro for scanned document parsing (4/4 images, 100% success)
• Agentic AI: Nova Act for autonomous browser automation
• Real-world healthcare billing automation use case
• NCCI compliance checking with 40+ bundling rules