🧠 Federated Learning on the MIMIC-III Dataset

This project explores the use of Federated Learning for predicting patient mortality in Intensive Care Units (ICUs) using the MIMIC-III clinical dataset.

🎯 Objectives

• Perform supervised classification to predict mortality outcomes.
• Implement non-IID data distribution across multiple simulated clients (e.g., hospitals).
• Compare performance between a federated model and a centralized model.

---

📁 Project Structure

• data/ : Contains raw CSV files extracted from MIMIC-III (e.g., CHARTEVENTS.csv, LABEVENTS.csv, etc.)
• client_{i}_data.csv : Partitioned non-IID data for each client.
• final_dataset_with_mortality.csv : Final dataset including mortality labels.
• non_iid_distribution.png : Visualization of data distribution across clients.
• federated_vs_centralized_performance.png : Side-by-side performance comparison.
• bdm_federated_learning_mimic_iii.py : Main script containing the full pipeline.

---

⚙️ Technologies Used

• Python (NumPy, Pandas, Matplotlib)
• PyTorch
• Scikit-learn
• Google Colab
• MIMIC-III Dataset (via PhysioNet)

---

🧪 Project Pipeline

1. Data Preprocessing

   • Extraction of key clinical features (e.g., GCS, vital signs, PaO₂/FiO₂, urine output, lab results)
   • Cleaning and handling of missing/outlier values

2. Label Engineering

   • In-hospital mortality
   • 48-hour ICU mortality
   • 30-day post-discharge mortality

3. Federated Learning Implementation

   • Non-IID partitioning across 10 clients
   • Local neural networks trained independently
   • Federated Averaging (FedAvg) for global model updates
   • Performance monitoring + early stopping

4. Evaluation & Visualization

   • Local and global model accuracy
   • Loss and accuracy curves per round
   • Comparison against a centralized model

---

📊 Results

• Federated Model:

  • Aggregates local client updates using FedAvg
  • Achieves competitive test accuracy across clients

• Centralized Model:

  • Trained on combined data for baseline comparison

Visualizations showcase class distributions, training dynamics, and final performance comparisons.

---

▶️ Demonstration Video

Watch a quick walkthrough of the project here:
👉 Project Demo (Google Drive)

---

🚀 How to Run

1. Download the MIMIC-III dataset and upload it to Google Drive
2. Run the bdm_federated_learning_mimic_iii.py script on Google Colab
3. Outputs and plots will be saved automatically

---

👩‍💻 Author

• Developed by [DJABRI Maroua ,BOUYAHIAOUI Meriem , DINARI Yasmine ,REZZOUG Aicha] as part of the Big Data Mining course project
• Dataset used: MIMIC-III Clinical Database

---

---