Secure and Efficient Federated Learning for Predictive Modeling in Resource-Constrained Healthcare Systems

Abstract

Predictive modeling in healthcare holds promise for improving clinical outcomes, but in many low-resource settings, data fragmentation, privacy concerns, and infrastructural limitations hinder centralized machine learning approaches. These barriers are especially critical in human immunodeficiency virus (HIV) care, where privacy concerns are heightened, and any use case involving patient-level data raises significant ethical, regulatory, and confidentiality concerns. We developed a privacy-preserving federated learning (FL) framework to predict HIV viral load suppression using retrospective data from 50,000 patients and over one million visits across 30 health facilities in Uganda. The framework utilizes federated averaging for distributed training, secure multiparty aggregation, and differential privacy to ensure data confidentiality. To address cross-site heterogeneity, we integrated domain-adversarial neural networks to promote domain-invariant feature learning. A multilayer perceptron model was trained collaboratively across facilities using only local data. The federated model achieved an area under the ROC curve (AUC) of 0.874, nearly matching a centralized baseline (AUC 0.881) and substantially outperforming site-specific models (average AUC 0.758). Sensitivity (89.6%) and specificity (66.8%) demonstrate strong capability in identifying both suppressed and unsuppressed cases. Domain adaptation reduced inter-facility performance variability, and differential privacy imposed minimal degradation in accuracy. Training was completed within one hour using modest hardware, which supported feasibility in low-resource settings. Our study demonstrates that FL can deliver robust, privacy-preserving predictive performance in HIV care without requiring the centralization of sensitive patient data. The proposed architecture is adaptable to other clinical prediction tasks and represents a practical pathway for scaling ethical AI across decentralized healthcare systems in low- and middle-income countries.

Received: 9 September 2025 | Revised: 3 November 2025 | Accepted: 11 November 2025

Conflicts of Interest

The authors declare that they have no conflicts of interest to this work.

Data Availability Statement

The data that support this work are available upon reasonable request to the corresponding author.

Author Contribution Statement

Alex Mirugwe: Conceptualization, Methodology, Software, Formal analysis, Resources, Data curation, Writing – original draft, Visualization. Juwa Nyirenda: Methodology, Validation, Resources, Writing – review & editing.