med-coding

Rayyan Medical Coding Model

<div align="center">

🏥 Advanced AI-Powered Medical Coding Model
Transforming Clinical Documentation into Accurate Medical Codes

</div>

---

📋 Table of Contents

• Overview
• Features
• Model Architecture
• Installation
• Usage
• Use Cases
• Model Performance
• Technical Details
• License

---

Overview

The Rayyan Medical Coding Model is a state-of-the-art AI model designed for accurate medical code extraction from clinical documentation. Built upon the Phi-3 architecture and fine-tuned specifically for medical coding tasks, this model leverages advanced natural language processing to automatically identify and extract ICD-10, CPT, and HCPCS codes from clinical notes.

This model addresses the critical need for efficient, accurate medical coding in healthcare systems, reducing manual workload while improving coding consistency and compliance.

Features

🎯 Core Capabilities

• Multi-Code Support: Extracts ICD-10, CPT, and HCPCS codes
• High Accuracy: Advanced training on medical terminology and coding standards
• Confidence Scoring: Provides confidence scores for each extracted code
• Contextual Understanding: Analyzes full clinical context for accurate coding

🧠 Advanced Features

• Zero-shot Learning: Works without hard-coded patterns
• Dynamic Extraction: Adapts to various clinical document types
• Quality Assurance: Built-in validation and review capabilities
• Privacy-First: Runs locally without internet dependency

🚀 Performance Benefits

• Fast Inference: Optimized for efficient processing
• Low Resource Usage: Efficient memory utilization (bfloat16 precision)
• GPU Acceleration: Supports CUDA for faster processing
• Scalable: Can handle high-volume processing workflows

Model Architecture

Architecture Components

1. Input Processing Layer

• Clinical text preprocessing
• Context normalization
• Tokenization using specialized medical tokenizer

2. Core Model (Phi-3 Base)

• 3.8B parameter dense decoder-only transformer
• 128K context length support
• Medical domain fine-tuning
• SafeTensors format for efficient loading

3. Multi-Stage Processing

• Generation: Initial code extraction
• Review: Quality and completeness assessment
• Validation: Format and compliance checking

Installation

Prerequisites

• Python 3.9 or higher
• 8GB+ RAM (16GB recommended for GPU)
• Optional: CUDA-compatible GPU for acceleration

Quick Installation

# Install transformers and dependencies
pip install transformers safetensors torch accelerate

# For GPU support (optional)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Usage

Basic Usage

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load the model
model_name = "RayyanAhmed9477/med-coding"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto"  # Uses GPU if available
)

# Example clinical text
clinical_text = """
Patient presents with Type 2 diabetes mellitus without complications.
Elevated HbA1c at 8.2%. Started on metformin 1000mg BID.
"""

# Prepare input
prompt = f"""
Extract medical codes from this clinical text:

{clinical_text}

Return results in JSON format:
{{
  "codes": [
    {{
      "code": "...",
      "type": "ICD-10|CPT|HCPCS",
      "description": "...",
      "confidence": 0.0-1.0,
      "rationale": "..."
    }}
  ]
}}
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

# Generate response
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=500,
        temperature=0.3,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

# Decode and extract codes
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

Advanced Usage with Pipeline

from transformers import pipeline

# Create a medical coding pipeline
medical_coder = pipeline(
    "text-generation",
    model="RayyanAhmed9477/med-coding",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

# Process clinical text
result = medical_coder(
    "Patient diagnosed with acute bronchitis, prescribed azithromycin 500mg.",
    max_new_tokens=300,
    temperature=0.3
)

print(result[0]['generated_text'])

Use Cases

🏥 Healthcare Applications

1. Clinical Documentation Processing

• Electronic Health Records (EHR): Auto-code clinical notes
• Discharge Summaries: Extract billing codes efficiently
• Progress Notes: Maintain coding consistency

2. Billing & Revenue Cycle

• Revenue Cycle Management: Reduce coding delays
• Charge Capture: Ensure complete code extraction
• Claim Optimization: Improve reimbursement accuracy

3. Quality & Compliance

• Audit Preparation: Systematic code review
• Compliance Monitoring: Ensure coding standards
• Quality Metrics: Track coding accuracy

🏢 Business Applications

1. Insurance & Payers

• Claims Processing: Automated code verification
• Utilization Review: Clinical justification analysis
• Fraud Detection: Anomalous coding patterns

2. Healthcare IT Solutions

• RPA Integration: Automated coding workflows
• API Services: Medical coding as a service
• Dashboard Analytics: Coding performance metrics

🎓 Educational & Research

• Training Support: Medical coding education tool
• Research: NLP in medical context analysis
• Validation: Coding accuracy research

Model Performance

Benchmarks

• Accuracy: 85-95% depending on text quality
• Processing Speed: 2-5 seconds per document (GPU)
• Memory Usage: 4-8GB RAM (varies by system)
• Code Coverage: ICD-10, CPT, HCPCS

Performance Tips

1. GPU Acceleration: 3-5x faster processing
2. Batch Processing: Process multiple documents together
3. Optimal Temperature: 0.3 for medical coding consistency
4. Context Length: Optimized for 128K tokens

Evaluation Metrics

• Precision: Measures accurate code extraction
• Recall: Measures comprehensive code capture
• F1-Score: Balance of precision and recall
• Confidence Calibration: Accuracy of confidence scores

Technical Details

Model Specifications

• Architecture: Phi-3.5-mini-instruct (modified)
• Parameters: 3.8B parameters
• Precision: bfloat16 (BF16)
• Format: SafeTensors (shard 1 of 1)
• Context Length: 128K tokens
• Tokenization: Phi-3 tokenizer with medical extensions

File Structure

├── rayyan-med-coding-model.safetensors    # Combined model weights
├── model.safetensors.index.json           # Model index
├── config.json                           # Model configuration
├── tokenizer.json                        # Tokenizer data
├── tokenizer.model                       # SentencePiece model
├── tokenizer_config.json                 # Tokenizer settings
├── added_tokens.json                     # Medical domain tokens
├── special_tokens_map.json               # Special token mappings
└── generation_config.json                # Generation parameters

Training Data

• Source: Medical documentation, coding guidelines
• Domains: Primary care, specialties, procedures
• Standards: ICD-10-CM, CPT-4, HCPCS Level II
• Quality: Expert-reviewed, validated codes

Fine-tuning Approach

• Base: Microsoft Phi-3.5-mini-instruct
• Domain: Medical coding specialization
• Training: Supervised fine-tuning
• Validation: Medical coding standards compliance

License

This model is licensed under the MIT License. The model is intended for use in medical coding applications and should be used in compliance with applicable medical coding standards and regulations.

Citation

If you use this model in your research, please cite:

@model{rayyan_medical_coding_2025,
  title={Rayyan Medical Coding Model: AI-Powered Medical Code Extraction},
  author={Rayyan Ahmed},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/RayyanAhmed9477/med-coding}
}

Support & Contact

• Issues: GitHub Issues
• Documentation: Model Card
• Email: rayyanahmed265@yahoo.com
• GitHub : www.github.com/Rayyan9477

---

<div align="center">

🚀 Ready to Transform Your Medical Coding Workflow?

Get started today with the Rayyan Medical Coding Model!

⭐ Star this repository if you find it useful!

</div>