Assistant Professor Sevgi Gurbuz received an NSF award for the project titled “Collaborative Research: ECCS: Small: Personalized RF Sensing: Learning Optimal Representations of Human Activities and Ethogram on the Fly”.
Radio Frequency (RF) sensing can be a game changer to reduce healthcare costs and disparities, improve quality of care, and facilitate aging-in-place because they are non-contact, low-power devices that are effective in the dark, do not limit or alter freedom of movement, and do not acquire private visual or audio recordings. However, a significant impediment to the advancement of RF technologies for recognition and health assessment of human gait is the continuous and sequential nature of human movement, which can be characterized by periods of activity and transitions between activities that depend upon a person’s mobility. Gait is a person-specific trait that embodies important health information for many disorders as well as aging related impacts. Thus, a second important challenge is the development of personalized machine learning (ML) models that continually learn from person-specific RF data to improve health-driven gait analysis.
The goal of this project is the design of a personalized RF-sensing framework for the monitoring of activities of daily living (ADL), detection and characterization of pathological, gaits round-the-clock 24/7 in a natural setting. The proposal?s two main objectives are: 1) Create a human ethogram via the formulation of a new and general framework for interpretation, segmentation, and categorization of a broad swath of human activities based on modeling the structure and dynamics of individual mobility; 2) Design a new approach for personalization of deep neural networks, where consecutive and contiguous observations are used to increase the classification accuracy of ADL for the monitored person. The ethogram is a quantitative, structured approach to describing daily human behavior in terms of ?body states? while activities are then modeled as transitions between states. By introducing the concept of personalization to RF sensing, this project will broaden the realm of personalized devices beyond its current scope of wearable and implantable devices to now include RF sensors. The proposal integrates sensor and kinematic knowledge into DNN design, resulting in novel architectures with greater accuracy that will advance the state-of-the-art in RF signal classification more generally. Moreover, the central ideas in this proposal are independent of device specifications and can be generalized to other sensing modalities, paving the way for non-contact, ubiquitous, fine-grained personalized gait classification and analysis. The outcomes of this project will pave the way for timely interventions and more effective treatments in both home and clinical settings, reducing the costs and improving the accessibility of health care.