Federated Learning Performance Depends on Site Variation in Global HIV Data Consortia.

Area of Science:

• Infectious disease epidemiology
• Machine learning in healthcare
• Public health informatics

Background:

• Machine learning (ML) models for HIV care are hindered by data sharing restrictions, limiting multi-site collaboration.
• Federated Learning (FL) offers a privacy-preserving method for training ML models across multiple sites without sharing patient-level data.

Purpose of the Study:

• To evaluate the effectiveness of Federated Learning (FL) for developing clinical prediction models in HIV care.
• To assess FL performance across diverse international sites and compare it to centralized and site-specific models.

Main Methods:

• Utilized data from 22,234 people living with HIV (PLWH) across six sites in five countries (CCASAnet).
• Applied FL algorithms for four prediction tasks: 1-year mortality, 3-year mortality, tuberculosis incidence, and AIDS-defining cancer incidence.
• Compared FL performance to centralized training and individual site-specific models, exploring the impact of site size and heterogeneity.

Main Results:

• FL algorithms achieved performance comparable to centralized models across all prediction tasks.
• FL substantially outperformed traditional site-specific models, demonstrating significant gains in predictive accuracy.
• Performance improvements varied by site, influenced by site size and data heterogeneity; local fine-tuning offered task-dependent benefits.

Conclusions:

• Federated Learning (FL) provides a scalable and privacy-preserving infrastructure for multi-site machine learning in international HIV research.
• FL overcomes data sharing limitations, facilitating robust model development and improving HIV care predictions globally.
• The findings support the broader adoption of FL for collaborative research in infectious diseases.