Decentralized Machine Learning for Healthcare: Survey on Disease Prognosis and Secure Communication Frameworks

Abstract

One of the biggest challenges in implementing AI in healthcare is the fragmentation of data across hospitals, as well as privacy regulations. Federated learning tackles these challenges by allowing models to be collaboratively built, while patient data is kept at its source. The survey focuses on two primary medical fields that can benefit from federated learning, cardiovascular risk assessment and confidential systems for exchanging information. We systemically categorize and discuss existing methods on training strategies, cryptographic protection, and adaptation techniques, as well as evaluate model performance across multiple clinical settings involving different data distributions and workflows. Adversarial-resistant secure communication schemes, such as encrypted computation and statistical privacy protection, are considered. The survey ends with proposals for research on customizing the model, implementing the model with less resources and ethical governance structures appropriate for real clinical use.