A federated learning-based approach to recognize subjects at a high risk of hypertension in a non-stationary scenario

Background: Transferring data across nodes could raise concerns about data security and privacy. Federated learning is a tech- nological remedy for these problems. However, in a real federated scenario, there are two main issues: 1) it is difficult to collect locally a large and representative dataset, and 2) the data are non-stationary, where non-stationary means that the data increase in size over time, which may generate catastrophic forgetting events affecting the learning process. Objective: The aim of this paper is to investigate and assess the performance and behavior of a federated model during the occurrence of catastrophic forgetting events within the context of a non-stationary data scenario. Methods: To achieve this objective, the performance of a proposed federated learning approach has been evaluated in terms of the continuous flow of data in order to train a time series-based model. Results and Conclusion: The results demonstrate the goodness of the solution in terms of performance, with improvements ranging from 2% to 28%. This indicates the benefits that the proposed approach brings to a node in terms of recovering from a catastrophic forgetting event, also taking into account the fact that the probability of catastrophic forgetting is higher at the beginning of the learning process.