Transmission Modeling with Smartphone-based Sensing
Infectious disease spread is difficult to accurately measure and model. Even for well-studied pathogens, uncertainties remain regarding the dynamics of mixing behavior and how to balance simulation-generated estimates with empirical data. Smartphone-based sensing data promises the availability of inferred proximate contacts, with which we can improve transmission models. This dissertation addresses the problem of informing transmission models with proximity contact data by breaking it down into three sub-questions. Firstly, can proximity contact data inform transmission models? To this question, an extended-Kalman-filter enhanced System Dynamics Susceptible-Infectious-Removed (EKF-SD-SIR) model demonstrated the filtering approach, as a framework, for informing Systems Dynamics models with proximity contact data. This combination results in recurrently-regrounded system status as empirical data arrive throughout disease transmission simulations---simultaneously considering empirical data accuracy, growing simulation error between measurements, and supporting estimation of changing model parameters. However, as revealed by this investigation, this filtering approach is limited by the quality and reliability of sensing-informed proximate contacts, which leads to the dissertation's second and third questions---investigating the impact of temporal and spatial resolution on sensing inferred proximity contact data for transmission models. GPS co-location and Bluetooth beaconing are two of those common measurement modalities to sense proximity contacts with different underlying technologies and tradeoffs. However, both measurement modalities have shortcomings and are prone to false positives or negatives when used to detect proximate contacts because unmeasured environmental influences bias the data. Will differences in sensing modalities impact transmission models informed by proximity contact data? The second part of this dissertation compares GPS- and Bluetooth-inferred proximate contacts by accessing their impact on simulated attack rates in corresponding proximate-contact-informed agent-based Susceptible-Exposed-Infectious-Recovered (ABM-SEIR) models of four distinct contagious diseases. Results show that the inferred proximate contacts resulting from these two measurement modalities are different and give rise to significantly different attack rates across multiple data collections and pathogens. While the advent of commodity mobile devices has eased the collection of proximity contact data, battery capacity and associated costs impose tradeoffs between the frequency and scanning duration used for proximate-contact detection. The choice of a balanced sensing regime involves specifying temporal resolutions and interpreting sensing data---depending on circumstances such as the characteristics of a particular pathogen, accompanying disease, and underlying population. How will the temporal resolution of sensing impact transmission models informed by proximity contact data? Furthermore, how will circumstances alter the impact of temporal resolution? The third part of this dissertation investigates the impacts of sensing regimes on findings from two sampling methods of sensing at widely varying inter-observation intervals by synthetically downsampling proximity contact data from five contact network studies---with each of these five studies measuring participant-participant contact every 5 minutes for durations of four or more weeks. The impact of downsampling is evaluated through ABM-SEIR simulations from both population- and individual-level for 12 distinct contagious diseases and associated variants of concern. Studies in this part find that for epidemiological models employing proximity contact data, both the observation paradigms and the inter-observation interval configured to collect proximity contact data exert impacts on the simulation results. Moreover, the impact is subject to the population characteristics and pathogen infectiousness reflective (such as the basic reproduction number, $R_0$). By comparing the performance of two sampling methods of sensing, we found that in most cases, periodically observing for a certain duration can collect proximity contact data that allows agent-based models to produce a reasonable estimation of the attack rate. However, higher-resolution data are preferred for modeling individual infection risk. Findings from this part of the dissertation represent a step towards providing the empirical basis for guidelines to inform data collection that is at once efficient and effective. This dissertation addresses the problem of informing transmission models with proximity contact data in three steps. Firstly, the demonstration of an EKF-SD-SIR model suggests that the filtering approach could improve System Dynamics transmission models by leveraging proximity contact data. In addition, experiments with the EKF-SD-SIR model also revealed that the filtering approach is constrained by the limited quality and reliability of sensing-data-inferred proximate contacts. The following two parts of this dissertation investigate spatial-temporal factors that could impact the quality and reliability of sensor-collected proximity contact data. In the second step, the impact of spatial resolution is illustrated by differences between two typical sensing modalities---Bluetooth beaconing versus GPS co-location. Experiments show that, in general, proximity contact data collected with Bluetooth beaconing lead to transmission models with results different from those driven by proximity contact data collected with GPS co-location. Awareness of the differences between sensing modalities can aid researchers in incorporating proximity contact data into transmission models. Finally, in the third step, the impact of temporal resolution is elucidated by investigating the differences between results of transmission models led by proximity contact data collected with varying observation frequencies. These differences led by varying observation frequencies are evaluated under circumstances with alternative assumptions regarding sampling method, disease/pathogen type, and the underlying population. Experiments show that the impact of sensing regimes is influenced by the type of diseases/pathogens and underlying population, while sampling once in a while can be a decent choice across all situations. This dissertation demonstrated the value of a filtering approach to enhance transmission models with sensor-collected proximity contact data, as well as explored spatial-temporal factors that will impact the accuracy and reliability of sensor-collected proximity contact data. Furthermore, this dissertation suggested guidance for future sensor-based proximity contact data collection and highlighted needs and opportunities for further research on sensing-inferred proximity contact data for transmission models.
transmission modeling, proximity contact data, sensor data collection, Kalman filtering
Doctor of Philosophy (Ph.D.)