Wearable computers are gaining significant attention due to their capability to enable a wide variety of new applications in domains such as wellness and health care. Despite their tremendous potential to impact our lives, wearable health monitoring systems face a number of hurdles before becoming a reality. The enabling processors and architectures demand a large amount of energy, requiring sizable batteries. This creates challenges for further miniaturization of the wearable units. This EAGER award is pursuing preliminary research in tiered, model based signal processing that can exploit pre-determined signal templates to enable extreme power optimization. In this approach, signal processing can be performed at several levels, where in each level, only the hardware for a specific template is active. If the template of interest is present, the next level of signal processing will be activated, otherwise hardware components corresponding to the next and the remaining levels will remain inactive. This approach, however, highly depends on the effectiveness of templates. In monitoring human movements, if templates do not accurately represent the physical activity of interest, the system will not exhibit acceptable accuracy. The goal is to develop effective techniques and methodologies to ensure templates adapt to remain valid throughout the operation of the system, accurately representing the corresponding physical movements.
The research focuses on speed-insensitive template matching architectures that can reduce the effects of movement variations on signal processing. Timing models for movements and user activity profiles are exploited to monitor the correctness of the signal processing, and tunable parameters decrease or increase the sensitivity of the signal processing. For example, if the user is expected to perform sit to stand at least once every two hours in the day time, and the tiered signal processing has not detected the movement in the past few hours, the sensitivity will be increased, or user interaction and template retraining can be initiated. When performing a movement that has been determined to be of interest, the user can initiate (re)training if the system does not recognize the movement. Effective template generation and on-line retraining are expected to open opportunities to individualize systems and signal processing and to reduce the complexity of storage and processing architectures. This research is expected to provide the groundwork for ongoing design and development of practical ultra low power signal processing architectures, reduce costs of computing platforms for medical sensing, and to enable future progress in areas such as gait and balance monitoring for fall prevention, and in-home movement monitoring for Parkinson?s disease.
Wearable computers are gaining significant attention due to their capability to enable a large variety of new applications in wellness and health care domains. Despite their tremendous potential to impact our lives, wearable health monitoring systems face a number of hurdles before becoming a reality. The enabling processors and architectures demand a large amount of energy, requiring sizable batteries. This creates challenges for further miniaturization of the wearable units. Model based signal processing techniques adapting to the physics and kinematics of human body will create attractive solutions for power optimization. This approach, however, highly depends on the effectiveness of models. If the models do not accurately represent the physical activities of interest, the system will not exhibit acceptable accuracy and the power optimization techniques will not operate to the expectations. The goal of this research is to investigate effective techniques and methodologies to ensure models remain valid throughout the operation of the system, and accurately represent their corresponding physical movements. The intellectual merit of this project includes preparing the ground work for a new set of novel algorithms that perform the signal processing in a tiered fashion eventually enabling ultra-low power processing units. The research created several models for periodic (e.g., walking, running), and transitional (e.g., sit to stand, lying) movements. Variations in models of activities (e.g., speed of movements, effects of environment) were investigated and techniques to overcome the limitations were developed. Finally methods that enhance the accuracy of signal processing algorithms in the presence of variations in the models of the activities were proposed. The proposed methodologies were validated in the context of two health monitoring and wellness related case studies: 1) gait and balance monitoring for fall prevention, 2) monitoring activities of daily living (ADL). The broader impact of this research includes creation of techniques for effective and low-power tracking and monitoring of human movements that will be very effective for kinesiologists who are interested in longevity studies on human movements as well as neurologists who are interested to observe the trends of changes in gait parameters. This project targeted a very important health care application with several hundreds of billion dollars annual financial toll. The concepts introduced here will create opportunities to individualize systems and signal processing and reduce the complexity of the processing architectures. The cost reduction associated with this will enable semiconductor companies to produce billions of chips for wearable computers. Growing demand for emerging monitoring applications of health care requires students, engineers and health care professionals to design, develop, deploy and operate wearable systems. This project mentored two undergraduate and five graduate as well as 2 high school students.
