Federated Learning for Privacy-Reversing Healthcare Challenges, Applications and Future Directions

In the today’s digital age, healthcare organizations are collecting vast amounts of highly sensitive data, such as patient records, diagnostic reports, and medical images. Traditionally, machine learning models required all of the data to be collected and stored in one centralized location. This presented several challenges concerning the confidentiality and security of the data. Federated Learning (FL) provides another way to use ML. Under FL, each organization keeps its own data on its premises. Rather than sharing the actual patient data, each organization shares only the updated version of the model. In this paper, we review several studies concerning FL applications within the field of healthcare. The studies we selected focus on medical imaging, smart healthcare, data protection, security, data heterogeneity, communication efficiency, scalability and interoperability. The studies demonstrate that FL enables organizations to collaborate and train models using their collective knowledge base without having to share the underlying data. However, there are several challenges and problems preventing wide-scale implementation of FL in real-world scenarios. Some of these include: heterogeneity among different datasets used for training models, communication requirements for FL, potential security risks associated with FL systems, limits to how large an FL system can grow, and interoperability issues between various healthcare systems. Based upon our review of the relevant studies, we identify a major gap in the current body of literature regarding FL applications in healthcare. The conceptual model we developed is also intended to provide a logical structure for incorporating all of the previously discussed limitations. While this conceptual model is based on prior research (and therefore, not yet empirically proven), experimental studies will need to validate its effectiveness in order to make it usable in future research.