A hierarchical federated multi-task transfer learning framework for paroxysmal arrhythmia detection, vital sign monitoring, and activity recognition

Abstract

An internet of medical things (IoMT) enabled remote patient monitoring system needs to accurately detect heterogeneous physiological and activity signals while maintaining patient privacy and data flow constraints on sensitive information. Most previous federated or multi-task learning methods focus on one of the following aspects: one-modality per site, one-task per site, or a distributed model optimization, but they do not consider both the signal processing specific to each modality and the knowledge transfer specific to each task. This study introduces a hierarchical federated multi-task transfer learning approach to solve this problem, which is used to detect paroxysmal arrhythmia, monitor heart rate from electrocardiogram (ECG) signals, and recognize human activities. The framework consists of three computational layers: Edge layer for signal preprocessing and learning local features, Federated layer for privacy preserving model aggregation, Cloud-level multi-task transfer learning layer for sharing transferable representations across related monitoring tasks. Wavelet-based denoising, R-peak detection, QR-interval delineation and convolutional neural network (CNN) dan bidirectional long short-term memory (BiLSTM)-attention based classification are performed on the processed ECG signals. The accelerometer and gyroscope data are not temporally synchronized with the ECG dataset, and all processing and activity recognition is done separately from the other data. With the use of the MIT-BIH Arrhythmia database and human activity recognition dataset, experimental results showed that the arrhythmia classification accuracy was 96.8% and the overall activity-recognition accuracy was 91.0%. The findings show that the proposed architecture is able to enable heterogeneous health-monitoring tasks under a shared decentralized learning architecture without moving raw data away from their source. While the framework offers a basis for scalable and privacy-compliant remote monitoring, there are key points that warrant further exploration such as communication overhead and the inclusion of client data diversity, computational complexity, and clinical validation.