Date of Award

2023

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Complex Systems and Data Science

First Advisor

Donna M. Rizzo

Second Advisor

Kristen L. Underwood

Abstract

Studying and evaluating time series signals that emerge when monitoring complex phenomena requires fusing, visualizing, and often reducing the dimensionality of large amounts of data to reveal the patterns and relationships that appear at different scales. In this work, we develop methods for monitoring, visualizing, and identifying the complex relationships that appear in time series data collected from two very different domains – health care conversations and river networks – to facilitate large-scale understanding of these systems. Fostering connection between clinicians and patients and their families in the context of serious illness is a fundamental component of good clinical communication that unfortunately is rare in modern healthcare. This is especially true for serious and life-threating illness, where a sense of trust is necessary for patients to fully express their treatment preferences. In river basins, frequent exchange of water and sediment between channel and floodplain leads to societal benefits including sediment and nutrient storage and attenuation of flood waves. However, human activities can disconnect floodplains, leading to adverse impacts on floodplain function and downstream water quality.My research combines existing and newly developed machine learning (ML) tools to investigate three very different sets of times series information. First, I constructed a ML pipeline to identify and classify moments of human connection in conversations between palliative care physicians and patients. Next, I developed novel terrain-based measures of channel/floodplain connectivity and erosion hazard in river systems. Finally, we applied unsupervised time series clustering methods to lexical and auditory features extracted from patient/therapist conversations recorded as part of psilocybin-assisted therapy study. The ML pipeline runs efficiently and the channel/floodplain connectivity and erosion hazard measures, developed as part of this research, help fuse disparate and complex data sets into measures that are easier to understand and that can be calculated rapidly at the catchment scale. Timeseries clustering of conversational features in the psilocybin-assisted therapy study showed differences in lexicon between the pre-dosing therapy sessions where the study team worked to develop the “set and setting” for the dosing day, the dosing day, and two post-dosing integration therapy sessions. I first show that fully automated detection and classification of connectional moments in serious illness conversations is feasible. This is a necessary step for conducting larger-scale multi-site studies in natural hospital environments. Next, I show proof-of-concept of my connectivity and erosion hazard measures and discuss how these can be used by water resource planners at the watershed scale. Lastly, I discuss the patterns revealed by timeseries analysis of conversational data associated with the psilocybin-assisted therapy study.

Language

en

Number of Pages

232 p.

Share

COinS