Transition-Aware Housekeeping Task Monitoring Using Single Wrist-Worn Sensor

• Published in: IEEE sensors journal 2018-11, Vol.18 (21), p.8950-8962
• Main Authors: Liu, Kai-Chun, Hsieh, Chia-Yeh, Chan, Chia-Tai
• Format: Article
• Language: English
• Subjects: Activity recognition; awareness of transition; Data collection; Demographics; Health care; housekeeping task monitoring; Medical services; Monitoring; Monitoring systems; Older people; Public health; Target detection; Target recognition; Task analysis; Wrist; wrist-worn sensor
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2018.2868278

Population aging is one of the general issues of public health over the world. Such demographic shifts pose challenges to healthcare system. Several wearable-based activity monitoring systems have been developed to improve the quality of healthcare and provide monitoring information for health professionals, such as the information of kitchen task, dressing task, food and fluid intake, and medication intake. However, few works pay attention to the housekeeping task, as the housekeeping performance correlates to the cognitive and functional health status in elderly people. Since typical clinical approaches of measuring and assessing housekeeping task performance suffer issues in long-term observation and manual error, a transition-aware household task monitoring system is proposed to support clinical professionals to gather fine-grained housekeeping task information for clinical assessment in this paper. Novel algorithms and models are proposed and designed based on knowledge and hierarchical approaches, including preliminary target activity recognition, transition detection, transition point identification, activity model, and activity inference. In addition, the typical activity classification and transition detection approach are implemented to compare with the proposed system. In the experiment, five healthy elderly participants are invited to take part in performing a set of four housekeeping activities, and there are 948 collected instances. The monitoring system is validated by using leave-one-subject-out cross-validation approach. The experimental results show that the best overall accuracy, recall, and precision of the proposed system can achieve 81.63%, 78.40%, and 78.58% when the window size is 2.0 s.