Decentralised AI for Infectious Diseases in LMICs: A Scoping Review of Clinical Decision Support and Public Health Applications

Résumé fourni par la source

This systematic scoping review will examine the existing literature on decentralised artificial intelligence for infectious disease-related clinical decision support and public health applications in low- and middle-income countries (LMICs). Decentralized Artificial intelligence (dAI) systems such as swarm – or federated learning architectures allow for training on distributed datasets of multiple institutions without exposing raw patient data, hence keeping data privately while instituions maintain data sovreignty. This is particularly relevant in LMIC settings, where data sharing often raises privacy concerns, questions of data sovereignty and unequal research partnerships. The review will map which decentralised AI methods have been applied or proposed, which infectious diseases and use cases have been addressed, and which LMIC settings are represented in the literature. It aims to find deployed and tested infrastructures and it will summarise reported outcomes and considerations related to model performance, privacy, data sovereignty, equity and feasibility. The review will follow a scoping review approach and will be reported in line with PRISMA-ScR guidelines. The expected output is a structured overview of the available evidence, including a categorisation of applications and methods and an identification of important research gaps. The review is intended to inform future research on privacy-preserving and context-sensitive AI approaches for infectious disease decision-making in LMICs. Definition of centralized and decentralized AI: Centralized AI: Raw data are transferred from their place of origin to a central data hub or cloud server for model training. Model aggregation and updates are typically controlled by a central entity. Decentralized AI (dAI): Raw data remain local to the respective data owner and are not directly shared for collaborative model training. Instead, learned model parameters or other derived model-related information are exchanged. Coordination and aggregation may be performed by a central coordinator, through a hierarchical architecture, or through a fully decentralized peer-to-peer protocol. The key difference is that centralized AI requires giving up data ownership to a central aggregator, while decentralized AI protects data sovereignty by moving the algorithm to the data rather than the data to the algorithm.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.