Integrating image caption information into biomedical document classification in support of biocuration

Area of Science:

• Biomedical Informatics
• Computational Biology
• Scientific Literature Analysis

Background:

• Scientific literature growth necessitates automated methods for information retrieval in biomedical research.
• Biocuration tasks, like those in the Gene Expression Database (GXD), are labor-intensive due to the vast number of publications.
• Effective automation of document classification is crucial for efficient curation of biological databases.

Purpose of the Study:

• To develop and evaluate a document classification scheme for identifying topic-relevant papers within large biomedical literature collections.
• To enhance existing meta-classification frameworks by incorporating features from figure captions alongside titles and abstracts.
• To support the biocuration classification task by improving the efficiency and accuracy of identifying relevant scientific articles.

Main Methods:

• A meta-classification framework was adapted to include textual features from titles, abstracts, and figure captions.
• The classifier was trained and tested on a large, imbalanced dataset of approximately 60,000 documents curated by the Gene Expression Database (GXD).
• The dataset spanned documents from 2012-2016, with each document represented by its title, abstract, and figure caption text.

Main Results:

• The developed classifier achieved a precision of 0.698, recall of 0.784, f-measure of 0.738, and Matthews correlation coefficient of 0.711.
• The results demonstrate the framework's effectiveness in handling imbalanced datasets common in biocuration.
• Performance significantly improved when incorporating figure caption information compared to using only titles and abstracts.

Conclusions:

• Figure captions contain substantial information that significantly enhances biomedical document classification and curation.
• The proposed framework offers an effective solution for automated identification of relevant scientific literature, aiding biocurators.
• This approach addresses the challenges posed by the increasing volume of biomedical publications and supports efficient knowledge discovery.