Semi-Parametric Continuous-Time Models for Adverse Health Behaviors
Moreno, Alexander F.
MetadataShow full item record
Two important problems in behavioral medicine are modeling the risk of adverse health behaviors such as smoking, and classifying longitudinal health data to determine health conditions. Models for each can be used for prediction in order to inform just in time adaptive interventions (JITAI) and inference to help behavioral scientists understand processes driving these behaviors. In this thesis we propose several models for both event risk and factors that contribute to risk for behaviors such as smoking, alcohol and drug use, and one model that can be applied to sensor time series classification. We use ecological momentary assessments (EMAs), which are surveys delivered repeatedly throughout the day, and wearable sensors to model risk factors. This is challenging as in order to inform intervention these models need good predictive accuracy, while in order to inform science they need to be interpretable from a behavioral standpoint and ideally have theoretical guarantees. Further, because human behaviors are often very complex, the flexibility of semi and non-parametric models is needed to adequately capture their dynamics. This thesis makes the following contributions: 1. A method to model the latent dynamics of ordinal longitudinal data using a discrete state latent variable model. This can model the latent dynamics of EMA, allowing us to forecast risk factors and describe how they evolve. 2. A joint model for longitudinal data and time-to-event with a discrete state latent variable model. This can jointly model the dynamics of EMA and a first lapse process, allowing us to describe how latent risk relates to both. 3. A method to describe the mean function of a counting process from discretely observed event counts, some of which may be missing. This can be used to model the expected number of cigarettes over time from EMA data when some responses are missing, which can help behavioral scientists understand how smoking dynamics change over time as well as allow them to compare smoking dynamics across groups. 4. A method for multimodal continuous attention densities parametrized by functions in reproducing kernel Hilbert space. This can be used in wearable sensor time series classification to identify regions of a signal that contribute to a classifier's decision, which can help convince specialists that a classifier's decision aligns with their intuition.