DrugGPT💊: Drug Guidance and Personalized Treatment Recommendations Agent

DrugGPT is an AI-powered system designed to assist clinicians and patients by providing personalized treatment recommendations using Large Language Models (LLM) and retrieval-augmented generation (RAG) architecture. The system emphasizes privacy protection, evidence-based recommendations, and efficient deployment.

Features

• Privacy-First Architecture: Utilizes RAG to protect patient data by converting sensitive clinical information into embeddings before LLM interaction
• Evidence-Based Recommendations: Integrates similar historical cases to provide treatment suggestions with transparent reasoning
• Efficient Deployment: Uses small and open-source LLaMA-3.2 1B model for fast response times and privacy assurance [Update: We changed to llama-7B]
• User-Friendly Interface: Built with Streamlit for easy interaction and clear presentation of recommendations

Project Structure

.
├── .devcontainer/     # Development container configuration
├── app/               # Application demo files
├── data/              # The MIMIC dataset is not public accessible
├── scripts/           # Utility scripts
└── src/               # Source codepoint, modules

Technical Architecture

Model Components

• Base Model: LLaMA-3.2-1B/7B for reasoning and medical domain tasks
• Embeddings: ClinicalBERT for medical text understanding
• Vector Database: FAISS for efficient similarity retrieval
• Framework: Streamlit for web interface

Data Processing

The system uses the MIMIC-III dataset (46,520 ICU patients) with the following preprocessing steps:

1. Text standardization
2. Key element extraction (diagnoses, medications)
3. Data structuring and unification
4. Train-test split (97% train, 3% test)

Getting Started

Prerequisites

• Permission of MIMIC dataset
• Python 3.x
• FAISS
• Streamlit
• LLaMA model
• ClinicalBERT

Installation

1. Clone the repository:

git clone https://github.com/yourusername/DrugGPT.git
cd DrugGPT

2. Install dependencies:

pip install -r requirements.txt

3. Run the application:

python app.py

Usage

The system accepts the following patient information:

• Age Group
• Gender
• Diagnosis
• Current Medications
• Clinical History

After submission, the system provides:

1. Medication suggestions with rationale
2. Monitoring recommendations
3. Similar cases analysis for reference

Future Development

Planned Improvements

• Fine-tuning LLaMA-3.2 1B
• Migration to LLaMA-2 7B
• Enhanced similarity scoring
• System validation with clinician feedback

Evaluation Metrics

• Baseline: Zero-shot prompt comparison with other LLMs
• Test Dataset: 3% of MIMIC-III data (approximately 100 patients)
• Performance Metrics:
  • Correctness against established medical guidelines
  • Computational cost analysis
  • Token usage efficiency

License

This project is licensed under the LICENSE file in the repository.

Authors

• Yuning Zheng @Igemirecle
• Mercury Liu @Europix

Acknowledgments

Carnegie Mellon University - School of Computer Science