Page 84 - AIH-2-4

Artificial Intelligence in Health Federated learning health stack against pandemics

Figure 1. Schematic representation of the hierarchical learning framework implemented in the proposed FL scheme. Here, d represents the depth of the
hierarchy, and root may represent the central organization that controls all subordinate levels (e.g., World Health Organization).

Our framework can be extended to train multiple specialized Hospitals), while data units include their affiliated
models at the central unit, depending on data characteristics hospitals, clinics, and pharmacies. These data owners
and application requirements. In addition, given the interact with designated local servers rather than directly
assumption that future pandemics will generate massive with the central server. Each data owner conducts private
volumes of data, we anticipate that the hierarchical structure training on its localized data and communicates model
of our framework can be scaled accordingly by incorporating parameters to the respective local server. The local servers
additional aggregation layers or nodes to efficiently manage aggregate these parameters and relay them to the central
the increased load and maintain performance. server.
2. Proposed FL framework Our proposed hierarchical FL framework effectively
addresses security risks during communication phases
In this section, we described our proposed secure and and at the server level. The hierarchical structure enables
scalable FL framework, capable of incorporating dynamic scalability by accommodating a wide range of healthcare
data during model training. In collaborative medical data, irrespective of its volume.
research, to enhance predictive diagnostic accuracy,
multiple healthcare data owners aim to train a unified model 2.1. Secure model training procedure
(or a set of unified models, depending on the nature of a The proposed framework consists of three primary entities:
future pandemic). The trained model should be available the central server, local servers, and local healthcare data
locally to all collaborators. In scenarios where a single data holders acting as clients. The central server is responsible
owner operates multiple healthcare units, privacy concerns for global aggregation, i.e., the final updated parameters
are minimized, as such organizations can consolidate their will be computed at the central server. The local or regional
data internally without external exposure. For example, servers are responsible for aggregation at the cluster or
in the United States (US), Kaiser Permanente operates a regional level, while the client organizations train their
network of hospitals and clinics as a unified entity. Similarly, local models on the datasets they manage locally.
in India and internationally, Apollo Hospitals manages an
extensive network of hospitals, clinics, and pharmacies. Let S and L represent the central server and a regional
i
The proposed framework allows these data owners to server, respectively, and n denote the number of regional
centralize patient data for analysis and predictive modeling servers, where 1 ≤ i ≤ n. Each regional cluster may contain
while ensuring strict internal control and compliance with a different number of clients. Let p be the number of
privacy regulations. Figure 2 illustrates the architectural clients under L, with each client denoted as C , where i
i
ij
structure of our proposed framework. Here, data owners corresponds to the local server and j represents the client,
refer to entities (such as Kaiser Permanente or Apollo where 1 ≤ j ≤ p. Every client C holds a private dataset D .
ij ij
Volume 2 Issue 4 (2025) 78 doi: 10.36922/AIH025080013