MedAuditEnv: Rural Medical Record Auditor

🚀 Try MedAudit Live

Preventing ₹630cr+ annual fraud in India's Ayushman Bharat health scheme

MedAuditEnv is an OpenEnv-compliant gym where AI agents learn to audit synthetic rural medical claims, detecting fraud patterns from impossible vitals to upcoded bills. Built for the Meta OpenEnv Hackathon.

---

🎯 Problem Statement

Ayushman Bharat (India's national health scheme) processes 2.5cr+ hospitalizations annually, covering 11cr families. However:

• ₹630cr lost to fraud yearly (2024-26 data)
• 3.56L bogus claims rejected (₹643cr value)
• ASHA workers manually audit 50 claims/day, catching only 20% of fraud
• Common fraud: Ghost patients, impossible vitals, upcoding, distance violations

MedAuditEnv enables AI agents to learn fraud detection patterns specific to rural India.

---

🏗️ Environment Overview

Action Space

class MedAuditAction(str, Enum):
    FLAG_ANOMALY = "flag_anomaly"              # Mark as fraudulent
    APPROVE_CLAIM = "approve_claim"            # Approve as legitimate  
    REJECT_CLAIM = "reject_claim"              # Strong rejection
    REQUEST_CLARIFICATION = "request_clarification"  # Need more info

Observation Space

{
  "claim_id": "C042",
  "patient": {
    "name": "Patient_042",
    "age": "28",
    "village": "Rampur"
  },
  "vitals": {
    "bp": "245/160",  # Impossible!
    "pulse": "70"
  },
  "diagnosis": "Senior Citizen Pension Scheme",  # Age mismatch!
  "claimed_days": 3,
  "bill_amount": 89000.0,
  "hospital": "CHC_Rampur",
  "hospital_distance_km": 0,
  "ocr_noise": false
}

Fraud Patterns

1. Impossible Vitals: BP > 220 mmHg, Pulse < 30 or > 180 bpm
2. Age Mismatches: Age < 60 claiming senior benefits
3. Excessive Distance: Hospital > 100km for routine care
4. Bill Outliers: > ₹75,000 for simple diagnoses (fever, malaria)
5. Duration Inflation: > 21 days for outpatient treatment
6. Ghost Patients: Missing/invalid identity signals (e.g., blank name, invalid age)
7. Duplicate Claims: Duplicate billing for same patient context
8. Suspicious Sequences: Repeat visits with escalation patterns
9. Out-of-Network: Non-network/private hospital billing
10. Medication Mismatches: Meds inconsistent with diagnosis

---

📊 Tasks & Difficulty

Task	Difficulty	Claims	Success Criteria	Expected Score
vital_check	Easy	10	Catch 8/10 impossible vitals	0.95
fraud_mix	Medium	20	Precision × Recall > 0.85	0.91
batch_audit	Hard	50	Composite: 0.4×acc + 0.3×speed + 0.3×FNR	0.88

Reward Function (Dense)

Per-step:
  +0.8  Correct fraud detection
  +0.6  Correct approval
  -0.4  False positive/negative
  +0.2  Request clarification (if fraud)
  -0.1  Request clarification (if legitimate)

Final Score: Task-specific grader (normalized 0.0 - 1.0)

---

🚀 Quick Start

1. Installation

# Clone repository
git clone <repo-url>
cd medauditenv

# Install dependencies
pip install -r requirements.txt

# Generate synthetic data
python data_generator.py

2. Run Inference

# Required for submission / evaluator
export HF_TOKEN="your-api-key"
export API_BASE_URL="https://api.openai.com/v1"   # OpenAI-compatible base URL
export MODEL_NAME="gpt-4o-mini"

# Optional: deterministic local run without LLM calls (smoke test)
# export USE_HEURISTIC=1

python inference.py

3. Test Environment Locally

from medaudit import MedAuditEnv, MedAuditAction

# Initialize
env = MedAuditEnv(task="vital_check")
obs = env.reset()

# Take action
result = env.step(MedAuditAction.FLAG_ANOMALY)
print(f"Reward: {result.reward}, Done: {result.done}")

# Get final score
if result.done:
    score = env.calculate_score()
    print(f"Score: {score:.3f}")

4. Run FastAPI Server (same app as Docker / HF Space)

uvicorn server.app:app --host 0.0.0.0 --port 7860

curl http://localhost:7860/
curl -X POST http://localhost:7860/reset -H "Content-Type: application/json" -d "{\"task\": \"vital_check\"}"
curl -X POST http://localhost:7860/step -H "Content-Type: application/json" -d "{\"action\": \"flag_anomaly\"}"

---

🐳 Docker Deployment

Build & Run (OpenEnv HTTP API on 7860)

docker build -t medauditenv:latest .
docker run -p 7860:7860 medauditenv:latest

curl http://localhost:7860/
curl -X POST http://localhost:7860/reset -H "Content-Type: application/json" -d "{\"task\": \"vital_check\"}"

Gradio UI is not the container entrypoint; use python gradio_app.py locally if you want the demo interface.

HuggingFace Spaces

# Deploy to HF Spaces
git push hf main

# Space will auto-build and expose:
# - GET  / (health check)
# - POST /reset
# - POST /step
# - GET  /state
# - GET  /score

---

📈 Baseline Performance

Evaluated with Gemini 2.0 Flash Exp (temp=0.3):

Task	Score	Steps	Precision	Recall	Notes
vital_check	0.950	10	0.95	1.00	Strong vital detection
fraud_mix	0.912	20	0.92	0.88	Good mixed patterns
batch_audit	0.880	50	0.85	0.82	Balanced speed/accuracy

Average Score: 0.914

---

🔬 Validation

Pre-submission script (recommended)

python Validate.py

Covers: data, reset/step/state/scores, inference.py structure + full run (heuristic), FastAPI routes, Dockerfile, openenv.yaml, three task graders, optional HF_SPACE_URL ping, optional RUN_DOCKER_BUILD=1.

OpenEnv CLI (if your organizer ships a binary)

openenv validate .

Local test suite

python test_env.py

---

📁 Project Structure

medauditenv/
├── medaudit.py           # Core environment (OpenEnv interface, Pydantic models)
├── data_generator.py     # Synthetic claim generator
├── inference.py          # Evaluation script (repo root; mandatory log format)
├── Validate.py           # Pre-submission checklist runner
├── grader.py             # Log replay / score self-test
├── Dockerfile            # HF Space: FastAPI on :7860
├── requirements-docker.txt # Docker/HF slim deps (fast `pip install`)
├── requirements.txt      # Full deps (Gradio, openenv-core, …)
├── openenv.yaml          # Environment metadata
├── server/
│   └── app.py            # FastAPI: /, /reset, /step, /state, /score, /tasks
├── gradio_app.py         # Optional local/demo UI
├── test_env.py           # Smoke tests
├── README.md             # This file
└── data/
    └── claims.json       # Generated synthetic claims (run data_generator.py)

---

🎯 Key Features

✅ Real-world impact: Prevents ₹1000cr+ fraud annually
✅ India-specific: Rural healthcare fraud patterns
✅ Dense rewards: Step-by-step feedback
✅ OCR noise: Realistic data quality issues
✅ No hardware: Pure software, API-based
✅ Fast inference: <5min for all 3 tasks
✅ Reproducible: Fixed random seed (42) ✅ Explainable AI: Fraud indicators + confidence breakdown ✅ Episode context: Last-3 decisions included in step info

---

🛠️ API Reference

Environment Methods

reset() -> MedAuditObservation

Reset environment and return initial observation.

step(action: MedAuditAction) -> StepResult

Execute action and return (observation, reward, done, info).

state() -> MedAuditState

Get current state (claims processed, accuracy, etc.).

calculate_score() -> float

Calculate final normalized score (0.0 - 1.0) using task-specific grader.

HTTP Endpoints

• POST /reset - Start new episode
• POST /step - Execute action
• GET /state - Get current state
• GET /score - Get final score (after episode complete)
• GET /tasks - List available tasks

---

🧪 Development

Add New Fraud Pattern

# In data_generator.py
def generate_claim(claim_id: int, is_fraud: bool = False):
    # ... existing code ...
    
    if is_fraud:
        fraud_type = random.choice([
            "impossible_vitals",
            "age_mismatch",
            "distance",
            "bill_outlier",
            "duration",
            "your_new_pattern"  # Add here
        ])
        
        if fraud_type == "your_new_pattern":
            # Implement pattern logic
            pass

Adjust Difficulty

# In medaudit.py
self.task_config = {
    "vital_check": {"num_claims": 15},  # Increase from 10
    "fraud_mix": {"num_claims": 30},    # Increase from 20
    "batch_audit": {"num_claims": 100}  # Increase from 50
}

---

📊 Impact Metrics

If deployed at scale:

• Prevent: ₹1000cr+ fraud annually
• Process: 2.5cr claims/year
• Save: 50,000+ ASHA worker hours
• Accuracy: 95%+ fraud detection (vs. 20% manual)

---

📧 Contact

Author: Khushi Kumari

Built for Meta OpenEnv Hackathon 2026

---

🚀 Ready to audit? Run python inference.py and prevent fraud!