Acquiring Plausible Predications from MEDLINE by Clustering MeSH Annotations

Area of Science:

• Biomedical Informatics
• Medical Subject Headings (MeSH)
• Unified Medical Language System (UMLS)

Background:

• The biomedical literature, exemplified by MEDLINE, is vast, containing over 23 million records.
• MeSH descriptors, often refined by subheadings, are used for manual indexing of these records.

Purpose of the Study:

• To leverage MeSH subheading information for clustering MeSH descriptor co-occurrences.
• To infer plausible predicates for these clusters, aiming to extract meaningful biomedical relationships.

Main Methods:

• Utilized UMLS to process MeSH descriptor co-occurrence data.
• Grouped disease-pharmacologic substance co-occurrences into six clusters in an initial experiment.
• A domain expert manually assigned predicates to the identified clusters.

Main Results:

• Achieved a mean accuracy of 85% for the top ten generated biomedical facts per cluster.
• Demonstrated the effectiveness of MeSH subheadings in identifying relationships.

Conclusions:

• MeSH subheadings hold significant potential for extracting plausible medical predications from the MEDLINE database.
• This approach can aid in the discovery and organization of biomedical knowledge.