Federated Learning Simulation

A privacy-preserving machine learning simulation demonstrating how multiple hospitals can collaboratively train a model without sharing sensitive patient data.

This project is published on AI Advances (Medium): https://medium.com/ai-advances/federated-learning-simulation-ff71e68ab1b5

🎯 Project Overview

This project simulates a federated learning scenario where three hospitals with different dataset sizes train individual logistic regression models for Parkinson's disease prediction. The models are then aggregated into a single federated model while preserving data privacy.

🏗️ Project Structure

├── data/                          # Original Parkinson's dataset
├── fake_hospitals_data/           # Simulated hospital datasets
├── shared_folder/                 # the only folder that is shared between the notebooks
├── Hospital_1.ipynb             
├── Hospital_2.ipynb              
├── Hospital_3.ipynb             
├── aggregation_and_testing.ipynb # Federated aggregation & evaluation 
├── federated_learning.py         # Core federated learning classes
├── hospital_model_trainer.py     # Individual hospital training logic
└── Notebook 0 - Data Preparation.ipynb # Dataset splitting simulation

🚀 How to Run

This is a simulation where each hospital notebook represents a different medical institution. They don't communicate directly - they only share model parameters through the shared_folder.

Step-by-Step Execution

1. Data Preparation (Optional - datasets already provided)

   jupyter notebook "Notebook 0 - Data Preparation.ipynb"

2. Train Individual Hospital Models

   jupyter notebook Hospital_1.ipynb
   jupyter notebook Hospital_2.ipynb  
   jupyter notebook Hospital_3.ipynb

3. Federated Aggregation & Testing

   jupyter notebook aggregation_and_testing.ipynb

🔬 Dataset Information

• Source: Parkinson's Disease Dataset by gargmanas
• License: GNU Free Documentation License 1.3
• Features: 22 voice measurement features (jitter, shimmer, fundamental frequency, etc.)
• Total Samples: 195 (split into 3 hospitals + test set)
• Task: Binary classification (Healthy vs. Parkinson's)

Hospital Data Distribution

• Hospital 1: 40 patients (7 healthy, 33 Parkinson's)
• Hospital 2: 42 patients (14 healthy, 28 Parkinson's)
• Hospital 3: 35 patients (6 healthy, 29 Parkinson's)
• Test Set: 78 patients (21 healthy, 57 Parkinson's)

🛡️ Privacy-Preserving Features

✅ No raw patient data sharing - Each hospital keeps their data locally ✅ Only model parameters exchanged - Weights and bias shared via shared_folder ✅ Simulates real regulatory compliance - Mimics GDPR, HIPAA restrictions ✅ Collaborative without centralization - Hospitals work together while maintaining independence

🎓 Educational Value

This project demonstrates:

• Federated learning concepts and implementation
• Healthcare data privacy preservation
• Model aggregation techniques
• Performance comparison methodologies
• Real-world medical ML applications

📄 License

This project is open source. The original Parkinson's dataset is licensed under GNU Free Documentation License 1.3.

🤝 Contributing

Feel free to fork this project and submit pull requests for improvements!

📚 References

• Original Dataset
• Federated Learning: Collaborative Machine Learning without Centralized Training Data