Automatic Assignment of Non-Leaf MeSH Terms to Biomedical Articles.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Information Retrieval

Background:

• Hierarchical classification is a specialized multi-label classification task, often optimized for micro F1-score.
• Accurate binary classifiers for challenging non-leaf labels in hierarchies can enhance overall classification performance.
• Medical Subject Headings (MeSH) is a crucial biomedical ontology used for indexing articles in PubMed, with non-leaf nodes presenting significant classification challenges.

Purpose of the Study:

• To propose and evaluate strategies for building improved binary classifiers for non-leaf MeSH terms.
• To reduce false positives in non-leaf MeSH term prediction by utilizing hierarchical information.
• To demonstrate the effectiveness of incorporating descendant node instances in classifier training.

Main Methods:

• Developed classification strategies that exploit training examples of descendant terms for non-leaf nodes.
• Focused on improving binary classifiers for non-leaf MeSH terms within the MeSH hierarchy.
• Evaluated methods on a set of 500 non-leaf MeSH terms, specifically targeting false positive reduction.

Main Results:

• Achieved a 7.5% relative improvement in macro-F1 score for non-leaf MeSH terms.
• Increased the micro-F1 score by 1.6% for the evaluated non-leaf terms.
• Demonstrated significant performance gains by incorporating hierarchical information.

Conclusions:

• Exploiting hierarchical information, specifically descendant node instances, is critical for improving binary classifiers of non-leaf MeSH terms.
• The proposed methods effectively reduce false positives and enhance both micro and macro F1-scores.
• This study is the first to highlight the role of hierarchical information in boosting binary classifiers for non-leaf MeSH terms.