(185) Federated learning as a smart tool for data analysis in infectious disease research: a scoping review

Background: The use of real-world data has become increasingly popular, also in the field of infectious disease (ID), particularly since the emergence of the COVID-19 pandemic. While much useful data for ID research is being collected, data storage is often fragmented across different sources. Privacy issues, legislation, and intellectual property concerns limit the possibility of storing data centrally, thereby also limiting the possibility of fully leveraging the potential power of combined data. Federated learning (FL) has been suggested to overcome privacy issues by making it possible to perform research on data from a variety of sources without those data leaving local servers.

Objectives: To summarize the existing applications of FL in ID research and to discuss the most relevant opportunities and challenges of FL in the field.

Methods: This is a scoping review of the literature. References were identified through searches of MEDLINE/PubMed, Google Scholar, Embase and Scopus using search terms related to FL methods and ID. References identified through those searches, as well as those identified through screening the reference lists of retrieved manuscripts, were included until July 2023.

Results: 30 studies were included in this review and these could be divided into 4 sub-topics: disease screening (n=21), prediction of clinical outcomes (n=3), infection epidemiology (n=4), and vaccine research (n=2). Most articles applied FL in the context of COVID-19. Data used in the identified studies comprise, amongst others, imaging data, electronic health records and mobile device data. In all studies, FL achieved good accuracy when predicting diseases and outcomes, and also in comparison to non-federated methods. However, most studies did not make use of real-world federated data, but rather showed the potential of FL by using a large dataset and partitioning it manually.

FL is shown to be a promising method for ID research, especially due to the possibilities of privacy preservation, early detection of disease, more equitable research, and local calibration of models. Challenges in the application of the method also exist, for instance due to data heterogeneity, lack of clinical interpretability, privacy-accuracy trade-offs, infrastructure issues, and the lack of external validation.

Conclusions: FL is a promising methodology, especially in cases where the data size at individual institutions is small, for instance when the disease of interest is rare. Examples of real-world applications of FL in ID research are limited and some challenges remain, hence further research on the application of FL in the field is warranted.