FEDRAL.AI: Privacy-Preserving Federated Learning for ICU Biomarker Discovery

A privacy-preserving federated learning platform for robust and fair biomarker discovery in ICUs using MIMIC-IV data. Enables hospitals to collaborate on AI model training without sharing sensitive patient data.

📋 Overview

FEDRAL.AI is a state-of-the-art privacy-preserving federated learning platform designed for robust and fair biomarker discovery in Intensive Care Units (ICUs). Leveraging the MIMIC-IV dataset, this project enables multiple healthcare institutions to collaborate on training high-performance AI models without ever sharing sensitive patient data.

Key Research Focus

• Robustness: Resilient against non-IID (Independent and Identically Distributed) data and adversarial outliers.
• Fairness: Ensuring equitable model performance across diverse patient demographics.
• Privacy: Implementing Differential Privacy (DP) and secure aggregation to protect patient confidentiality.
• Biomarker Discovery: Extracting interpretable and clinical-ready insights from complex ICU electronic health records.

---

🏗️ System Architecture

High-Level Design

The system consists of a centralized Federated Server and multiple Hospital Agent nodes. Each node performs local training on its internal MIMIC-IV slice and shares only encrypted gradients or model weights.

graph TD
    subgraph Hospitals ["Hospital Clients (Local Nodes)"]
        H1[Hospital A]
        H2[Hospital B]
        H3[Hospital N]
    end

    subgraph FedServer ["Federated Training Server (Central Hub)"]
        Aggregator[Global Model Aggregator]
        Verifier[Robustness & Fairness Verifier]
        Encrypter[Privacy Layer / DP]
    end

    subgraph Data ["MIMIC-IV Dataset (Local)"]
        D1[(Patient Data A)]
        D2[(Patient Data B)]
        DN[(Patient Data N)]
    end

    H1 --- D1
    H2 --- D2
    H3 --- DN

    H1 -- "Local Gradients" --> FedServer
    H2 -- "Local Gradients" --> FedServer
    H3 -- "Local Gradients" --> FedServer

    FedServer -- "Aggregated Global Model" --> H1
    FedServer -- "Aggregated Global Model" --> H2
    FedServer -- "Aggregated Global Model" --> H3

    style FedServer fill:#2c3e50,stroke:#3498db,stroke-width:4px,color:#fff
    style Hospitals fill:#34495e,stroke:#bdc3c7,stroke-width:2px,color:#fff

Loading

Federated Training Workflow

The training follows a recursive round-based approach to minimize communication overhead while maintaining privacy.

flowchart LR
    Start([Start FL Round]) --> Distribute[Distribute Global Model]
    Distribute --> LocalTrain[Local Training on MIMIC-IV Data]
    LocalTrain --> Aggregation[Secure Aggregation]
    Aggregation --> PrivacyCheck[Privacy & DP Verification]
    PrivacyCheck --> UpdateGlobal[Update Global Model]
    UpdateGlobal --> NextRound{Next Round?}
    NextRound -- Yes --> Distribute
    NextRound -- No --> End([Final Biomarker Discovery])

Loading

---

📊 Experiment Results

Our experiments compare traditional centralized learning against our robust Federated Learning approach. The results show only a minimal (5%) loss in utility for a significant gain in privacy and robustness.

Scenario	Stability	Privacy Guarantee	Fairness Score	AUC Performance
Baseline (Centralized)	High	None	Medium	0.96
FedAvg (Standard FL)	Medium	None	Low	0.92
Fedral (Private - ε=1.0)	High	High	High	0.91
Fedral (Robust Median)	Highest	Medium	High	0.93

---

🛠️ Technology Stack

• Backend: FastAPI (Python 3.9+)
• Frontend: React + Vite (Tailwind CSS)
• Desktop: Electron (Hospital Agent)
• AI/ML: PyTorch / TensorFlow, Scikit-learn
• Database: SQLite (SQLModel)
• Deployment: Docker & Docker Compose

---

🚀 Getting Started

Prerequisites

• Node.js (v18+)
• Python (v3.9+)
• Docker (optional)

Setup & Installation

1. Clone the repository:

   git clone https://github.com/USER_NAME/privacy-preserving-fl-icu.git
   cd privacy-preserving-fl-icu

2. Run the setup script:

   chmod +x setup.sh
   ./setup.sh

Running the Application

• Web Dashboard: ./run-web
• Hospital Agent (Desktop): ./run-electron
• Run All Experiments: ./run_all_experiments.sh

---

🖥️ Fedral.AI Web Dashboard

The central management console for federated learning experiments, hospital management, and compliance monitoring.

---

🏥 Hospital Agent Desktop Application

The desktop application used by participating hospitals to securely contribute their data to the federated learning network.

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🤝 Acknowledgments

• MIMIC-IV: For providing the critical electronic health records.
• Privacy-FL Community: For inspiration on robust aggregation methods.