NanaAgemangKNUST/Thesis_Project

"Federated Learning Framework for Privacy-Preserving Malaria Risk Prediction Across West African Health Districts" # Federated Learning for Cross-Border Malaria Surveillance > Privacy-preserving collaborative machine learning for malaria prevalence prediction across West African health systems This repository contains the implementation and experimental code for our research on **federated learning for cross-border malaria surveillance** in West Africa, addressing the critical challenge of data heterogeneity while maintaining data sovereignty compliance with the Ghana Data Protection Act, Nigeria Data Protection Act 2023, and the African Union Malabo Convention. ## 📋 Table of Contents - Overview - Key Features - Research Questions - Installation - Quick Start - Dataset - Methods Compared - Experimental Protocol - Repository Structure - Configuration - Results - Citation - License - Acknowledgments --- ## 🔬 Overview Malaria surveillance in West Africa faces two conflicting imperatives: 1. **Regional collaboration** for effective cross-border disease tracking 2. **Data sovereignty** mandated by national data protection regulations This work demonstrates that **federated learning (FL)** can enable privacy-preserving collaboration while handling extreme data heterogeneity across countries. We compare **8 methods** — including centralized and local-only baselines and 6 federated learning algorithms — on real malaria surveillance data from Ghana, Mali, Nigeria, and Burkina Faso. ### The Challenge: Data Heterogeneity West African malaria data exhibits **22-fold variation** in prevalence: - **Coastal regions (Ghana):** 2.3% prevalence (hypoendemic) - **Sahelian regions (Mali):** >50% prevalence (hyperendemic) This extreme heterogeneity violates the i.i.d. assumption of standard federated learning, making algorithm selection critical. ### Our Contribution - **Comprehensive comparison** of 8 methods (2 baselines + 6 FL algorithms) - **Rigorous statistical validation** with 15 random seeds (~70% statistical power for effect size d=0.5 at α=0.05) - **Privacy-utility tradeof …