Please use this identifier to cite or link to this item:
|Title:||Text Mining in Biomedical and Healthcare Domains|
|Keywords:||Computer Science & Engineering|
|Abstract:||A significant amount of knowledge from biomedical research, clinical practice and crowdintelligence stay concealed inside the unstructured contents of scientific articles, electronic medical records, and immensely popular social media platform. Extracting relevant information from the biomedical/medical contents can significantly expedite the growth of research and development in the related disciplines. However, because of the high variability and creativity in natural language, it limits our ability to perceive when diverse looking proclamations are stating the same thing. Advancement in Natural Language Processing (NLP) technology is one of the most promising avenues for discovering vital biomedical/clinical information from such data. Substantial progress has been made in providing technologies in biomedical natural language processing such as information extraction, relationship discovery, and question answering. However, the solutions are often based on the various machine learning models involving structured lexicons and ontologies which are not only expensive and time-consuming to generate but are often restricted to a specific domain. Taking into consideration, the language of biomedical text that do not conform to the uniform annotation guidelines, the techniques, developed by targeting a specific domain, often fails to show reasonable performance when it is evaluated for some other domains. Furthermore, most standard approaches make strong simplifying language assumptions require welldesigned feature representations which again is manual and domain-dependent. This dissertation focuses on developing a domain adaptable framework based on the concept of evolutionary algorithm and neural network technique to solve various biomedical/clinical NLP tasks with minimal or no human effort. Firstly, we propose a new evolutionary feature selection method to automatically select the relevant features for biomedical entity extraction task. As a use-case, we used the biomedical scientific literature derived from PubMed and MEDLINE to extract biomedical entities. Secondly, we analyze the limitations of the current techniques for entity extraction and propose three new methods based on deep learning, that outperforms the state-of-the-art technique. As a use-case, we conduct our study on electronic medical records to extract patient health information. We further applied our method on biomedical literature to curate the protein interacted scientific articles and identify the interacting relationship between the protein mentions. Finally, we develop a novel multi-task learning framework that improves sharing across various domains/tasks and outperforms single-task learning as well as state-of-the-art. We validated the proposed multi-task learning method on pharmacovigilance mining task using social media text such as Twitter and medical blog-posts as well as on the biomedical literature. As our second use-case, we explore the proposed method for classifying the medical blog-post based on sentiment it possesses.|
|Appears in Collections:||01. CSE|
Files in This Item:
|Text Mining in Biomedical and Healthcare Domains.pdf||2.51 MB||Adobe PDF||View/Open Request a copy|
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.