Heterogeneous sensor configurations and the need to protect sensitive user data present ongoing challenges for Human Activity Recognition (HAR) in wearable healthcare systems. Due to label restrictions and privacy concerns, traditional machine learning techniques, which depend on centralized data aggregation and labeled target data, are frequently impractical in healthcare settings. We suggest a novel framework that combines Unsupervised Domain Transfer (UDT) and Federated Learning (FL) to enable privacy-preserving activity recognition across a variety of devices without the need for labeled target data to address these problems. The FedProx algorithm ensures stability in the federated setting while the UCI HAR dataset serves as the source domain and PAMAP2 as the target domain for the framework’s evaluation. To evaluate the generalizability of the approach, a variety of architectures are investigated, such as Transformers, Bidirectional LSTMs, and Convolutional Neural Networks (CNNs). According to experimental findings, the suggested framework successfully bridges domain gaps under federated constraints while achieving competitive performance across all architectures. These results demonstrate how crucial domain alignment and signal preprocessing are to the development of scalable, privacy-sensitive HAR for wearable medical devices.