Integrating Deep Learning and Network Science to Support Healthcare Management

Abstract

The internet, online platforms, and open-source repositories provide an alternate way of sharing and spreading health information, knowledge, and topics. The usage of health-related documents, such as intellectual properties, biomedical research literature, and health social medial conversation, has grown rapidly in the last few years. The texts in the health domains are no longer being used only for storing data and conveying information for communication purposes but also used in healthcare research. The data recorded are often in an unstructured manner as in the free-text format. The unstructured textual data present various challenges to the researchers since the data are not primarily collected for research purposes. Text and data mining techniques, specifically in the fields of deep learning and network science, are increasingly being used to process large amounts of textual data for research purposes and uncover insights for healthcare management. This thesis concerns the use of deep learning and network science-based data mining and natural language processing techniques to process unstructured text data in the health domain, including USPTO patents issued within the category of health informatics, research abstracts in biomedical literature of MEDLINE, and question threads in online health communities (ASKDOCS, etc.). In this thesis, I present three efforts to mine health concerned insights from health-related text data. In the first effort, I propose a new framework that incorporates the patent heterogeneous network analysis and network community detection to track the technology evolution for the health informatics technology domain. In our next effort, I focus on the problem of ranking health responses/suggestions from an online health forum where users can ask health-related questions and responses are provided by qualified doctors or patients who have had similar conditions. I propose a novel Knowledge-Enhanced Response Ranking System based on knowledge components (based on user knowledge and external knowledge sources) and content features of each response. In the third effort, I introduce a novel word-level attention bi-directional LSTM (deep learning) based method to extract Drug-Drug Interactions (DDIs) from biomedical research publications and extract important interaction terms/words from sentences that indicate DDIs. Our methods range from heterogeneous network analytics, deep learning, and natural language processing, etc. Overall, I demonstrate that it is valuable to glean insights and knowledge from intellectual properties, scientific publications and online social media through machine learning and text mining methodologies. The automated knowledge uncovered can be used to facilitate the management and development within the health domains for healthcare professionals and researchers, companies, and even governments.