Year of Publication

2020

Degree Name

Doctor of Philosophy (PhD)

Document Type

Doctoral Dissertation

College

Engineering

Department

Computer Science

First Advisor

Dr. Ramakanth Kavuluru

Abstract

The Internet provides an alternative way to share health information. Specifically, social network systems such as Twitter, Facebook, Reddit, and disease specific online support forums are increasingly being used to share information on health related topics. This could be in the form of personal health information disclosure to seek suggestions or answering other patients' questions based on their history. This social media uptake gives a new angle to improve the current health communication landscape with consumer generated content from social platforms. With these online modes of communication, health providers can offer more immediate support to the people seeking advice. Non-profit organizations and federal agencies can also diffuse preventative information in such networks for better outcomes. Researchers in health communication can mine user generated content on social networks to understand themes and derive insights into patient experiences that may be impractical to glean through traditional surveys. The main difficulty in mining social health data is in separating the signal from the noise. Social data is characterized by informal nature of content, typos, emoticons, tonal variations (e.g. sarcasm), and ambiguities arising from polysemous words, all of which make it difficult in building automated systems for deriving insights from such sources.

In this dissertation, we present four efforts to mine health related insights from user generated social data. In the first effort, we build a model to identify marketing tweets on electronic cigarettes (e-cigs) and assess different topics in marketing and non-marketing messages on e-cigs on Twitter. In our next effort, we build ensemble models to classify messages on a mental health forum for triaging posts whose authors need immediate attention from trained moderators to prevent self-harm. The third effort deals with models from our participation in a shared task on identifying tweets that discuss adverse drug reactions and those that mention medication intake. In the final task, we build a classifier that identifies whether a particular tweet about the popular Juul e-cig indicates the tweeter actually using the product. Our methods range from linear classifiers (e.g., logistic regression), classical nonlinear models (e.g., nearest neighbors), recent deep neural networks (e.g., convolutional neural networks), and ensembles of all these models in using different supervised training regimens (e.g., co-training). The focus is more on task specific system building than on building specific individual models. Overall, we demonstrate that it is possible to glean insights from social data on health related topics through natural language processing and machine learning with use-cases from substance use and mental health.

Digital Object Identifier (DOI)

https://doi.org/10.13023/etd.2020.057

Funding Information

This research was supported by the National Cancer Institute grant R21CA218231 (2017-2019)

Available for download on Thursday, March 04, 2021

Share

COinS