Data Informed Health Simulation Modeling
Combining reliable data with dynamic models can enhance the understanding of health-related phenomena. Smartphone sensor data characterizing discrete states is often suitable for analysis with machine learning classifiers. For dynamic models with continuous states, high-velocity data also serves an important role in model parameterization and calibration. Particle filtering (PF), combined with dynamic models, can support accurate recurrent estimation of continuous system state. This thesis explored these and related ideas with several case studies. The first employed multivariate Hidden Markov models (HMMs) to identify smoking intervals, using time-series of smartphone-based sensor data. Findings demonstrated that multivariate HMMs can achieve notable accuracy in classifying smoking state, with performance being strongly elevated by appropriate data conditioning. Reflecting the advantages of dynamic simulation models, this thesis has contributed two applications of articulated dynamic models: An agent-based model (ABM) of smoking and E-Cigarette use and a hybrid multi-scale model of diabetes in pregnancy (DIP). The ABM of smoking and E-Cigarette use, informed by cross-sectional data, supports investigations of smoking behavior change in light of the influence of social networks and E-Cigarette use. The DIP model was evidenced by both longitudinal and cross-sectional data, and is notable for its use of interwoven ABM, system dynamics (SD), and discrete event simulation elements to explore the interaction of risk factors, coupled dynamics of glycemia regulation, and intervention tradeoffs to address the growing incidence of DIP in the Australia Capital Territory. The final study applied PF with an SD model of mosquito development to estimate the underlying Culex mosquito population using various direct observations, including time series of weather-related factors and mosquito trap counts. The results demonstrate the effectiveness of PF in regrounding the states and evolving model parameters based on incoming observations. Using PF in the context of automated model calibration allows optimization of the values of parameters to markedly reduce model discrepancy. Collectively, the thesis demonstrates how characteristics and availability of data can influence model structure and scope, how dynamic model structure directly affects the ways that data can be used, and how advanced analysis methods for calibration and filtering can enhance model accuracy and versatility.
Machine Learning, Agent-based Modeling, Social Network, Multi-Scale Hybrid Simulation Modeling, Particle Filtering, System Dynamics
Master of Science (M.Sc.)