Data analysis and system identification algorithms for behavioral dynamical modeling
Open Access
- Author:
- Hojjatinia, Sahar
- Graduate Program:
- Electrical Engineering
- Degree:
- Doctor of Philosophy
- Document Type:
- Dissertation
- Date of Defense:
- October 06, 2022
- Committee Members:
- Minghui Zhu, Major Field Member
Constantino Lagoa, Chair & Dissertation Advisor
Vishal Monga, Major Field Member
David Conroy, Outside Unit & Field Member
Thomas La Porta, Program Head/Chair - Keywords:
- System Identification
Sparsity
Behavioral Modeling - Abstract:
- In the past few years, the development and advancements in new sensing technologies and wearable activity monitors have enabled the collection of high frequency individual data. This opens the exciting opportunity of identifying the relation between different variables and modeling the dynamic responses of the output to input variables. Such models then can be used in designing and developing targeted adaptive micro-treatments that use collected data to determine which is the best treatment option and when it should be delivered. However, there are significant challenges in the analysis of this type of data. First, due to the high rate of data collection, it can no longer be assumed that the input or the micro-intervention only has an instantaneous effect on the output; there are also some delayed effects. Moreover, such data frequently suffers from data fragmentation, i.e., poor placement of sensors, non-wear of the data collecting device and/or external disturbances that can lead to intervals of time where the data collected is not reliable, that is missing or corrupted. Literature in behavioral medicine, a multidisciplinary field focusing on health aspects of biological, behavioral, psychological, and social sciences have shown that a variety of medical and mental health conditions can be prevented or treated if intervened before or in early stages of the condition. For example, studies have shown a relationship between physical inactivity and a large range of diseases such as cardiovascular and metabolic diseases, etc. As another example, cigarette smoking is the leading preventable cause of death in the United States, responsible for about one in five deaths annually, with the yearly medical and economic burden of more than $\$300$ billion. The availability of large amounts of intensive longitudinal data and the limitation of currently used methods for analyzing such data has been the motivation behind this work. The main objective of this dissertation has been to deal with such challenges by leveraging concepts from the areas of control systems engineering, dynamical models, machine learning, and signal processing, and the application of these tools and techniques in modeling the human behavior. To achieve this goal, the main focus was to develop different algorithms and tools to identify models that consider the delayed effects of inputs and output and are capable of handling the fragmented data. Finally, the standard algorithms used and the developed ones were implemented and applied to identify the models describing the relation between different variables and analyze health problems in behavioral medicine. The drawbacks of system identification methods mostly used in literature and the efficiency of the atomic norm minimization technique in identifying parsimonious models from experimental data have motivated us to develop novel identification algorithms that consider sparsity using the atomic norm concept, in addition to using and implementing some of the standard algorithms. That is, by developing parsimonious model identification algorithms, we seek to identify models that provide a desired explanation of the data with as few parameters as possible. As an application of the algorithms used or developed in this dissertation, we identified personalized dynamic models of physical activity in response to digital messaging interventions to promote physical activity and reduce sedentary behavior. Identified switched system parameters provide the basis to tailor decision rules that can be used for future just-in-time adaptive interventions. Additionally, we proposed a new methodology to identify relation between stress and smoking considering sparsity. This is most important at the moments preceding and following the smoking episodes since it provides the basis for designing person-specific, tailored smoking cessation interventions from the parameters linking vulnerable moments to intervention decisions in the future works. As another project, a time-varying version of logistic regression model in combination with regularized l1-norm, to induce sparsity, is developed to identify the models that best describe the dynamics of the data and can predict the future outcomes.