Evaluating Federated continual learning for heart failure risk prediction: a simulation study

This study evaluates federated continual learning for heart failure risk prediction under changing clinical data conditions. Using a simulated 12-month, five-site dataset, it compares attention-based continual learning with FedAvg, FedProx, linear, and replay-based approaches. The findings show that apparent benefits of continual-learning components can largely disappear when baseline models receive comparable training budgets, highlighting the importance of fair baseline comparisons when evaluating federated clinical prediction methods.