CAREER: Scalable and Adaptable Cross-Domain Autonomous Health Assessment

点击翻译按钮获取中文摘要

The wide availability of commodity smart home sensor systems (Google Home, Amazon Echo, etc.) and internet-of-things (IoT) devices (Fitbit, Actigraph, etc.) is making it easier to continuously monitor individuals' health-related vital signals, activities, and behaviors to provide just-in-time health intervention to the aging population. This CAREER project seeks to design, implement, and evaluate heterogeneous sensor systems in smart homes that help ameliorate the progressive functional and behavioral health decline of older adults. The work specifically looks at cross-domain approaches that can accommodate variability in behavior, activity, and physiological health conditions across a large population and diverse set of smart home sensor systems. The inability to build scalable and adaptable activity and behavior monitoring models across domains such as multi-occupant homes with heterogeneous internet-of-things devices is a major impediment to adoption of smart home technologies for healthcare applications. The project develops novel deep transfer learning techniques, optimization-based heuristics, opportunistic sensing architecture, and spatiotemporal dynamical systems-based approaches to address the diversity, adaptability, and reliability of activity and behavior recognition models across different users and technologies, while leveraging a human-in-the-loop control for improving the performance of the sensor systems. These techniques will help automate activity and physiological health monitoring at scale, and thereby improve the design and study of adaptive interventions for elderly people, their families, and professional caregivers. In order to realize autonomous health assessment methodologies in practice, it is necessary to build an activity and behavior recognition system across multiple inhabitants and various connected consumer devices that can select, adapt, and cope with device and user heterogeneities, privacy characteristics, resource constraints and scarcity of labeled data. To address the above-mentioned problems, this research project contributes to new methodology in four ways. First, it is introducing deep transfer learning activity recognition model and multi-user multi-device optimization-based heuristics that automatically help adapt the inherent variations across different domains, including user/device-type/device-instance. Second, it is designing a spatio-temporal dynamical system approach based on fractal dynamics to mitigate the variability in various sensor signals, and capture the self-similarity of human physiological health markers and establish the parametric task performance dependency between functional and behavioral health measurements. Third, it posits an opportunistic sensing architecture and human-in-the loop activity model for real-time data sharing and annotation that help optimize the user interruption and system performance. Fourth, it is designing a distributed implementation of tailored-computational techniques in actual smart home deployments, and evaluating the effectiveness of sensor-based functional and behavioral models and algorithms for just-in-time health assessment in actual living environments. In addition to the targeted focus on education, an ongoing collaboration with the University of Maryland, School of Nursing is being leveraged for real deployment of smart home sensor systems and technologies at three retirement community centers and senior homes in the greater Baltimore area to compound the impact of proposed evidence-based research efforts.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.