Unsupervised Medical Subject Heading Assignment Using Output Label Co-occurrence Statistics and Semantic Predications

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Medical Subject Headings (MeSH)

Background:

• The National Library of Medicine (NLM) uses Medical Subject Headings (MeSH) for indexing biomedical literature.
• Automated methods for MeSH term assignment often rely on supervised machine learning, requiring extensive labeled training data.
• Biomedical data privacy concerns limit the availability of such training datasets.

Purpose of the Study:

• To develop and evaluate a novel unsupervised approach for assigning MeSH terms to biomedical abstracts.
• To explore the utility of named entity recognition, relationship extraction, and MeSH term co-occurrence frequencies.
• To assess the potential of unsupervised methods in overcoming data limitations in biomedical indexing.

Main Methods:

• Utilized named entity recognition and relationship extraction on a large corpus of NLM-indexed articles.
• Calculated co-occurrence frequencies for MeSH term pairs from existing indexed articles.
• Developed an unsupervised model based on extracted relationships and co-occurrence statistics.

Main Results:

• The unsupervised approach achieved a micro F-score comparable to supervised methods.
• Demonstrated the effectiveness of using output label co-occurrence statistics for indexing.
• Showcased the feasibility of unsupervised term recommendation without sensitive training data.

Conclusions:

• Unsupervised methods utilizing text-derived relationships and label co-occurrences show significant promise for biomedical article indexing.
• This approach offers a viable alternative to supervised methods, especially when labeled data is scarce or sensitive.
• Further research into exploiting free-text relationships and co-occurrences can enhance automated indexing systems.