Date of Award
Doctor of Philosophy in Computing Sciences - (Ph.D.)
Soon Ae Chun
Nowadays, patient-generated social health data are abundant and Healthcare is changing from the authoritative provider-centric model to collaborative and patient-oriented care. The aim of this dissertation is to provide a Social Health Analytics framework to utilize social data to solve the interdisciplinary research challenges of Big Data Science and Health Informatics. Specific research issues and objectives are described below.
The first objective is semantic integration of heterogeneous health data sources, which can vary from structured to unstructured and include patient-generated social data as well as authoritative data. An information seeker has to spend time selecting information from many websites and integrating it into a coherent mental model. An integrated health data model is designed to allow accommodating data features from different sources. The model utilizes semantic linked data for lightweight integration and allows a set of analytics and inferences over data sources. A prototype analytical and reasoning tool called “Social InfoButtons” that can be linked from existing EHR systems is developed to allow doctors to understand and take into consideration the behaviors, patterns or trends of patients’ healthcare practices during a patient’s care. The tool can also shed insights for public health officials to make better-informed policy decisions.
The second objective is near-real time monitoring of disease outbreaks using social media. The research for epidemics detection based on search query terms entered by millions of users is limited by the fact that query terms are not easily accessible by non-affiliated researchers. Publically available Twitter data is exploited to develop the Epidemics Outbreak and Spread Detection System (EOSDS). EOSDS provides four visual analytics tools for monitoring epidemics, i.e., Instance Map, Distribution Map, Filter Map, and Sentiment Trend to investigate public health threats in space and time.
The third objective is to capture, analyze and quantify public health concerns through sentiment classifications on Twitter data. For traditional public health surveillance systems, it is hard to detect and monitor health related concerns and changes in public attitudes to health-related issues, due to their expenses and significant time delays. A two-step sentiment classification model is built to measure the concern. In the first step, Personal tweets are distinguished from Non-Personal tweets. In the second step, Personal Negative tweets are further separated from Personal Non-Negative tweets. In the proposed classification, training data is labeled by an emotion-oriented, clue-based method, and three Machine Learning models are trained and tested. Measure of Concern (MOC) is computed based on the number of Personal Negative sentiment tweets. A timeline trend of the MOC is also generated to monitor public concern levels, which is important for health emergency resource allocations and policy making.
The fourth objective is predicting medical condition incidence and progression trajectories by using patients’ self-reported data on PatientsLikeMe. Some medical conditions are correlated with each other to a measureable degree (“comorbidities”). A prediction model is provided to predict the comorbidities and rank future conditions by their likelihood and to predict the possible progression trajectories given an observed medical condition. The novel models for trajectory prediction of medical conditions are validated to cover the comorbidities reported in the medical literature.
Ji, Xiang, "Social analytics for health integration, intelligence, and monitoring" (2015). Dissertations. 130.