Effective biomedical document classification for identifying publications relevant to the mouse Gene Expression Database (GXD)
Year: 2017
Type: other
Abstract: The Gene Expression Database (GXD) is a comprehensive online database within the Mouse Genome Informatics (MGI) resource, aiming to provide available information about endogenous gene expression during mouse development. The information stems primarily from many thousands of biomedical publications that database curators must go through and read. Given the very large number of biomedical papers published each year, automatic document classification plays an important role in biomedical research. Specifically, an effective and efficient document classifier is needed for supporting the GXD annotation workflow. We present an effective – yet relatively simple – biomedical document classification scheme to assist curators in identifying publications that are relevant to GXD. We use a large manually curated dataset, consisting of more than 25,000 PubMed abstracts, of which about half are curated as relevant to GXD while the other half as irrelevant to GXD. In addition to text from title-and-abstract, we also consider image captions, an important information source that we integrate into our method. The classification scheme we propose uses readily-available tools while employing several of our strategies of statistical feature selection to reduce the document representation size, and to focus on terms that support the GXD classification task. As image captions in biomedical documents typically carry significant and useful information for determining the relevance of a publication to a topic, we select text-features obtained from image captions as well as from the title and the abstract. We employ and compare well-used classifiers such as Random Forest and Naïve Bayes, and assess them via widely accepted measures for document classification evaluation including precision and recall, as well as the utility measure, which favors high recall. Our experiments over the large GXD dataset show that our classifiers retain a high level of performance across a variety of cross-validation settings and performance measures. This indicates that our classifier is effective and can indeed be useful in practice. We also ran multiple experiments using different sets of features, showing that classifiers relying on features obtained from title, abstract and image captions, perform better according to every measure, which demonstrates that image captions indeed provide valuable information supporting the GXD document classification task. Our classifier retains similar level of performance on documents outside the training/test set, demonstrating that our classifier is stable and practically applicable. The biomedical document classification method we present is effective and robust over a realistically large biomedical document dataset derived from the GXD. Moreover, the performance of our classifier affirms the utility of image captions as a substantial evidence source for automatically determining the relevance of biomedical publications to a specific subject area.
Institution: University of Delaware
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W2734569213