Adaptive biosignal data gathering for distributed and continual remote monitoring

Long-term biosignal monitoring should adopt continual learning models considering past and current temporal traits. Continual monitoring requires both high quality data that capture each user's temporal traits and high performance models. Thanks to the latest sensing and device technology, user-local continual data acquisition has become easier. However, the data of many users distributed remotely need to be gathered to label data, train models, and analyze them in a real system. The biggest problems are the communication volume and the cloud capacity with uploading and storing raw biosignal data of all users. We, therefore, propose a distributed active sampling method for continual learning. Our method adaptively enables both a model to be efficiently trained and temporal traits to be extracted for each user at each time adaptively. We also verify our method in a use case of arrhythmia observation with ECG. The experimental results indicate its effectiveness in terms of efficiency for model training, graspability for temporal traits, and adaptability.