Improving the utility of MeSH® terms using the TopicalMeSH representation

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Information Retrieval

Background:

• Biomedical document retrieval and classification are crucial for accessing medical literature.
• Current methods often rely on MeSH terms or text-based representations, with varying effectiveness.
• Latent topic models offer alternative ways to represent document content.

Purpose of the Study:

• To evaluate a novel vector representation, TopicalMeSH, which integrates MeSH terms with latent topic proportions.
• To compare TopicalMeSH performance against standard MeSH term-based and text-based representations in retrieval and classification tasks.
• To assess the utility of TopicalMeSH across different machine learning models and corpora.

Main Methods:

• Developed the TopicalMeSH representation by combining latent Dirichlet allocation (LDA) topics with MeSH terms.
• Evaluated performance on 15 systematic drug review corpora for document retrieval (precision, recall) and classification (F1 score).
• Compared TopicalMeSH against MeSH terms, text (bag-of-words), combinations thereof, and supervised LDA using SVM, logistic regression, and decision trees.

Main Results:

• TopicalMeSH consistently improved document retrieval precision across 11 of 15 corpora compared to MeSH alone.
• For classification, TopicalMeSH achieved higher F1 scores in 14/15 corpora with SVMs, 12/15 with logistic regression, and 12/15 with decision trees.
• TopicalMeSH outperformed LDA topics, tf-idf weighted MeSH terms, and their combinations in most corpora.

Conclusions:

• The TopicalMeSH representation offers a superior approach for biomedical document retrieval and classification.
• Integrating latent topics with MeSH terms enhances the effectiveness of standard information retrieval and machine learning models.
• This novel representation provides a more robust method for organizing and accessing biomedical information.