Feature engineering for MEDLINE citation categorization with MeSH.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Machine Learning

Background:

• Biomedical text categorization traditionally relies on bag-of-words models.
• Sophisticated text representations (syntactic, semantic, argumentative) are underexplored.
• Current research evaluates advanced features for reproducing Medical Subject Headings (MeSH) annotations.

Purpose of the Study:

• To assess the impact of diverse biomedical text representations on MeSH annotation accuracy.
• To compare traditional features (unigrams, bigrams) with advanced features (noun phrases, metadata, citation structure, semantic annotations).
• To evaluate the effectiveness of combining various feature sets and machine learning algorithms.

Main Methods:

• Utilized unigrams, bigrams, noun phrases, citation metadata, citation structure, and semantic annotations (MetaMap, UMLS concepts) as text features.
• Evaluated feature performance in reproducing MeSH annotations for frequent MeSH headings.
• Tested machine learning algorithms, focusing on those robust to class imbalance.

Main Results:

• Unigrams and bigrams demonstrated strong performance, outperforming noun phrases (too sparse) and citation metadata/structure.
• MetaMap conceptual annotation showed performance comparable to unigrams; UMLS concept integration did not yield further improvements.
• Combining all feature sets significantly outperformed individual sets and enhanced a state-of-the-art MeSH indexer. Algorithms resilient to class imbalance performed better.

Conclusions:

• Combining traditional text features with advanced representations effectively boosts bag-of-words performance for MeSH annotation.
• The synergy between feature sets and machine learning algorithms is critical for system performance.
• Algorithms robust to class imbalance improve results, but careful selection is needed to prevent overfitting with large feature sets.