Events Calendar

Harshvardhan Vathsangam, USC (PhD Defense)

Speaker: Harshvardhan Vathsangam, USC

Talk Title: Sense and Sensibility: Statistical Techniques for Human Energy Expenditure Estimation Using Kinematic Sensors

Series: PhD Defense Announcements

Abstract: Sense and Sensibility: Statistical Techniques for Human Energy Expenditure Estimation Using Kinematic Sensors ==== Healthcare is undergoing a shift from the episodic, expert-driven, curative approaches of the past towards a self-empowered, preventative model for the future. Central to this is the treatment of chronic illnesses. This treatment will require the adoption of behavioral changes on one's lifestyle. In this thesis, we focus on the negative effects of one such chronic illness: physically inactivity.
Regular physical activity is associated with decreased mortality, lower risk of cardiovascular disease, diabetes mellitus, colon and breast cancer.
Despite this knowledge, physical activity levels are not adequate.
Central to the need to get people to be more active is the ability to accurately measure and characterize physical activity in a cost-effective yet ubiquitous manner. One dimension of characterization of physical activity is the energy expended as a result of that activity. In this dissertation, we aim to demonstrate how kinematic sensors in combination with statistical techniques can accurately predict energy expenditure due to physical activity.

We cast the problem of determining energy expenditure in a mathematical framework and discuss various functional maps. We derive a set of frequency-based features that are robust to location on the human body and orientation. We use these features to determine the most accurate 'per-person' technique to map movement to energy expenditure. We compare prediction accuracies using different sensor streams and algorithms. A comparative study of accuracy versus inference time is also performed. We extend this work to be able to generate maps given a minimal set of morphological descriptors such as height, weight, age etc. We present and compare a set of models including nearest neighbor models, weight-scaled models, a set of hierarchical linear models and speed-based approaches. We show how these approaches can be used to evaluate the best subset of morphological descriptors and the best individual descriptor to generate personalized maps across people. These contributions are a step towards designing cost-effective, accurate and ubiquitous solutions to estimate physical activity levels and designing interventions based on accurately measured data.

Host: Lizsl DeLeon

Location: Ronald Tutor Hall of Engineering (RTH) - 406

Audiences: Everyone Is Invited

Contact: Assistant to CS chair