• 

  Login Page

MedPredict: Hospital Readmission Prediction System with LLM

Inspiration

MedPredict was inspired by the pressing need to reduce hospital readmissions and improve patient outcomes. With rising healthcare costs and operational pressures, a system that could early identify at-risk patients was needed, enabling timely interventions and more efficient resource management.

What it does

MedPredict is a hospital readmission prediction system that leverages advanced Large Language Models (LLMs) to analyze medical data and predict patients at risk of readmission. Key features include:

• 92% accurate readmission prediction using PubMedBert, ClinicalXLNet, and BigBird models
• HIPAA-compliant data security for handling sensitive clinical information
• Interactive RAG (Retrieval-Augmented Generation) prototype using GPT-4 API for improved clinical decision-making
• Full-stack application with Angular frontend and FastAPI backend
• Integration with AWS S3 for managing large-scale MIMIC III clinical data
• Firebase authentication for secure user access

How I built it

MedPredict was built using a comprehensive tech stack:

• Frontend: Angular for a responsive and interactive user interface
• Backend: FastAPI for efficient API development
• Containerization: Docker for consistent deployment across environments
• Cloud Storage: AWS S3 for managing 100GB of MIMIC III clinical data
• Authentication: Firebase for secure user management
• AI Models: PubMedBert, ClinicalXLNet, BigBird, and GPT-4 API for advanced natural language processing and prediction
• Data Processing: Extensive preprocessing and feature engineering on the MIMIC III dataset

Challenges I ran into

• Handling and processing the large-scale MIMIC III dataset (100GB) efficiently
• Ensuring HIPAA compliance while maintaining system performance
• Integrating multiple LLMs and optimizing their performance for readmission prediction
• Developing an effective RAG system that could improve clinical decision-making
• Balancing model accuracy with interpretability for healthcare professionals

Accomplishments that I'm proud of

• Achieving 92% accuracy in readmission prediction using advanced LLMs
• Successfully managing and processing 100GB of MIMIC III clinical data
• Developing a HIPAA-compliant system for handling sensitive medical information
• Creating an interactive RAG prototype that improved clinical decision-making efficiency by 80%
• Building a full-stack, Dockerized application that integrates complex AI models with a user-friendly interface

What I learned

• Advanced techniques in processing and analyzing large-scale clinical datasets
• Integration of multiple LLMs for improved prediction accuracy
• Implementation of HIPAA-compliant data security measures in a cloud-based environment
• Development of RAG systems for enhancing AI-assisted decision-making in healthcare
• Balancing technical complexity with user experience in healthcare applications

What's next for MedPredict

• Expand the system to include more diverse datasets for improved generalization
• Enhance the RAG system to provide more detailed and actionable insights for healthcare providers
• Implement real-time data integration for more up-to-date predictions
• Develop a mobile application for easier access by healthcare professionals
• Conduct clinical trials to validate the system's effectiveness in real-world healthcare settings
• Explore integration with electronic health record systems for seamless adoption in hospitals

Built With

• docker
• fastapi
• firebase
• llm
• openai
• postman
• python
• react