Page 82 - AIH-2-4

Artificial Intelligence in Health Federated learning health stack against pandemics

1. Introduction model under the coordination of a central server without
exchanging raw data.
1.1. General overview: The big picture
The FL process typically begins with the central server
The emergence of the COVID-19 pandemic in 2019 shook initializing a global model and distributing its parameters
the foundations of human civilization. While it was not the to selected clients. Each client independently trains the
deadliest pandemic in history, it ranks fifth in terms of death model using its local dataset, ensuring data privacy by
toll, with an official World Health Organization (WHO) keeping raw data on the local device. After local training,
1
estimate of 7 million deaths. When ranked by death toll, clients send model updates, such as model gradients, to the
the following pandemics are the deadliest, in descending central server. The server then aggregates these updates,
order: (i) the Spanish flu (17 – 100 million deaths, 1918 – commonly using methods such as federated averaging
1920), (ii) the Plague of Justinian (15 – 100 million deaths, (FedAvg), to update the global model. The updated model
2
11
541 – 549), (iii) human immunodeficiency virus/acquired is then redistributed to the clients for the next round of
3
immunodeficiency syndrome (approximately 43 million training. This cycle continues iteratively until the model
deaths, 1981–present), and (iv) the Black Death (7 – 35 converges or reaches a predefined accuracy threshold.
4
million deaths, 1346 – 1353). 4
Finally, the trained global model is deployed for use.
What sets COVID-19 apart from previous pandemics As discussed, FL eliminates the need to transfer raw
is that it occurred in a highly globalized, post-internet data to a central server; instead, only model parameters
world, where information could spread instantaneously. are exchanged. This approach significantly reduces the risk
5
Combined with significant advancements in modern of data breaches, as sensitive information remains on the
sciences, particularly in medical and computational client’s side. FL thus upholds data privacy and security,
sciences, it was shocking to see how the pandemic exposed making it one of the highly accepted protocols across
the limitations of human technological capabilities. countries, communities, and agencies. For this reason, it
As argued in previous studies, although we had the has been chosen in this article as the foundational model
5
components of a modern technological infrastructure to for addressing future pandemics.
create a formidable defense, we failed to synergize them
effectively to prevent the pandemic’s escalation. We envision that combating future pandemics will
require the holistic deployment of the entire computing
Therefore, the central question addressed in infrastructure, including, but not limited to, the Internet-
this article is: What would the next-generation of-Things (IoT) devices, sensors, edge computing, and
computational infrastructure look like to effectively cloud infrastructure. While FL offers privacy-preserving
combat future pandemics? Such infrastructure should model training, its integration with emerging technologies,
be deployable within current medical record-keeping such as edge intelligence and IoT, in healthcare introduces
systems, which involve heterogeneous data types and new challenges that must be addressed. Recent studies
regulations (mostly privacy-related). To stimulate this have highlighted the growing need for edge intelligence
discussion, we propose a federated learning (FL) - and distributed learning in health care. 12,13 For example,
6,7
based, global machine learning (ML) architecture with research presents an overview of mobile health systems
the potential to address this key issue. While we do not and IoT technologies, focusing on system architecture
claim this is the ultimate solution (though we believe it and data integration strategies. However, centralized
14
may be), the main purpose of this article is to motivate or cloud-based models often require raw or partially
the journey towards that goal. For brevity, we will processed data to be transmitted, raising concerns about
highlight one or a few methods for each step, although data security in sensitive applications, such as health
multiple viable solutions may exist. A comprehensive care. In another study, the authors discussed the role
13
review of FL applications in smart healthcare is of artificial intelligence in IoT-enabled smart healthcare
presented elsewhere; 8-10 hence, we will only summarize applications, with a primary focus on lightweight
key technical preliminaries of FL here. neural network architectures deployed directly on edge
In the absence of a global enforcing agency for devices. Yet, the study does not address the security risks
administering critical pandemic/medical issues (as the associated with transmitting trained model parameters
WHO lacks such authority), a decentralized approach from edge devices. This omission overlooks potential risks
offers the most promising pathway for global acceptance such as adversaries intercepting and reverse-engineering
and implementation. FL is a decentralized ML approach the model parameters to extract sensitive information,
in which clients (e.g., organizations, healthcare units, or manipulating them to induce bias or abnormal
mobile devices, and sensors) collaboratively train a shared model behavior. In our proposed framework, we use a

Volume 2 Issue 4 (2025) 76 doi: 10.36922/AIH025080013