Convolutional Neural Networks for Biomedical Text Classification: Application in Indexing Biomedical Articles.

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Machine Learning

Background:

• Accurate text classification is crucial for extracting information from biomedical literature.
• Traditional methods like logistic regression and SVMs face challenges with large feature spaces.
• Automated Medical Subject Heading (MeSH) term assignment aids the National Library of Medicine's PubMed search system.

Purpose of the Study:

• To evaluate the effectiveness of Convolutional Neural Networks (CNNs) for biomedical text classification.
• To compare CNN performance against traditional discriminative approaches for MeSH term assignment.
• To demonstrate CNNs' potential in improving the accuracy of indexing biomedical articles.

Main Methods:

• Utilized Convolutional Neural Networks (CNNs) to build binary text classifiers.
• Applied CNNs to the task of assigning Medical Subject Headings (MeSH terms) to biomedical articles.
• Compared CNN performance against established methods on a public dataset, focusing on challenging MeSH terms.

Main Results:

• CNNs demonstrated superior performance compared to traditional methods in biomedical text classification.
• Achieved an absolute improvement of over 3% in macro F-score for classifying difficult MeSH terms.
• Showed performance gains with CNNs on experiments involving 50 high-frequency MeSH terms.

Conclusions:

• Convolutional Neural Networks (CNNs) offer a powerful and direct approach for biomedical text classification.
• CNNs significantly enhance the accuracy of automated Medical Subject Heading (MeSH) term assignment.
• The findings highlight the strong potential of CNNs for advancing biomedical literature indexing and information retrieval.