Sensor-Based Human Activity Recognition for Elderly In-patients with a Luong Self-Attention Network

The mobility status of older adults is directly con-nected to their health conditions and physical abilities. Decrease in mobility is a serious issue that leads to adverse health outcomes and declined functional abilities in older adults. Hence, it is highly desirable to assess their mobility status continuously. With the advent of wearable technology, Sensor-based Human Activity Recognition (HAR), which uses measurements from body-worn sensors to gain insights into activities undertaken by individuals, has become an active research field. However, HAR for older adults poses two major challenges: (i) Collecting data from older adults is inherently difficult due to their declining physical abilities, and (ii) Diversity in sensor data distributions is more pronounced across different older adults due to their different mobility statuses. In this paper, we propose an attention-based deep learning model to accurately estimate older adult's mobility through wearable sensor-based HAR. The proposed model achieves a 7.5 % and 7.25 % improvement over baseline methods in a pooling task learning and meta-learning settings on an in-house IMU dataset collected from in-hospital patients, the first of its kind.