Computer-aided approaches to enhance systematic review development

Medical knowledge learned in medical school can become quickly outdated given the tremendous growth of the biomedical literature. It is the responsibility of medical practitioners to continuously update their knowledge with recent, best available clinical evidence to make informed decisions about patient care. However, clinicians often have little time to spend on reading the primary literature even within their narrow specialty. As a result, they often rely on systematic evidence reviews developed by medical experts to fulfill their information needs. At the present, systematic reviews of clinical research are manually created and updated, which is expensive, slow, and unable to keep up with the rapidly growing pace of medical literature. This dissertation research aims to enhance the traditional systematic review development process using computer-aided solutions. The first study investigates query expansion and scientific quality ranking approaches to enhance literature search on clinical guideline topics. The study showed that unsupervised methods can improve retrieval performance of a popular biomedical search engine (PubMed). The proposed methods improve the comprehensiveness of literature search and increase the ratio of finding relevant studies with reduced screening effort. The second and third studies aim to enhance the traditional manual data extraction process. The second study developed a framework to extract and classify texts from PDF reports. This study demonstrated that a rule-based multipass sieve approach is more effective than a machine-learning approach in categorizing document-level structures and iv that classifying and filtering publication metadata and semistructured texts enhances the performance of an information extraction system. The proposed method could serve as a document processing step in any text mining research on PDF documents. The third study proposed a solution for the computer-aided data extraction by recommending relevant sentences and key phrases extracted from publication reports. This study demonstrated that using a machine-learning classifier to prioritize sentences for specific data elements performs equally or better than an abstract screening approach, and might save time and reduce errors in the full-text screening process. In summary, this dissertation showed that there are promising opportunities for technology enhancement to assist in the development of systematic reviews. In this modern age when computing resources are getting cheaper and more powerful, the failure to apply computer technologies to assist and optimize the manual processes is a lost opportunity to improve the timeliness of systematic reviews. This research provides methodologies and tests hypotheses, which can serve as the basis for further large-scale software engineering projects aimed at fully realizing the prospect of computer-aided systematic reviews.