Conceptual framework and algorithms to classify wandering travel patterns of elderly with dementia
Vuong, Nhu Khue
Date of Issue2017-11-23
School of Computer Science and Engineering
Wandering is one of the most common behavioral disturbances among people with dementia (PWD). Sensing and localization technologies have been used in wandering management, especially elopement prevention. However, little research has been focused on measuring wandering behavior and not many applications in wandering management are widely used in practice due to two reasons. First, technologists’ understanding and perception of wandering do not align with those of gerontologists. This consequently lowers the chance of proposed solutions to be accepted for clinical research and applications. Second, most solutions do not address all the dimensions of dementia wandering nor do they cater to the needs of other stakeholders involved including caregivers, physicians and researchers. The contributions of this thesis are two-fold. We first present a conceptual map of wandering science from the perspectives of gerontologists. Then we provide a framework which identifies main threads of technologies that can be further developed to manage dementia wandering. We further discuss research and design issues, human factors, ethical concerns, security and privacy that need to be considered when implementing solutions for wandering management. Second, we develop pattern recognition methods to identify and classify travel patterns automatically from sensor data. According to gerontologists, this is the first key step in any specific investigation of dementia wandering and subsequently measuring wandering behavior. In this thesis, we design and develop two discriminative algorithms to classify dementia-related travel patterns using different sensor modalities. The first algorithm uses spatial and temporal information from location sensors and the second one uses inertial information from inertial sensors. We have evaluated the performance of our developed algorithms on real world datasets of both dementia and non-dementia subjects. A comparison of our algorithms’ performance with one of classical machine learning classifiers, Markov models, and time series classification algorithms such as Symbolic Aggregation Approximation (SAX) and Dynamic Time Warping (DTW) shows that our algorithms outperform other classifiers from 5% to 26% in terms of classification recall and 51 to 739 times faster in terms of classification processing time.