MeSH2Matrix: combining MeSH keywords and machine learning for biomedical relation classification based on PubMed.

Area of Science:

• Bioinformatics
• Computational Biology
• Medical Informatics

Background:

• Advanced machine learning improves biomedical relation classification from scholarly texts.
• Reliance on raw text limits generalization, precision, and reliability of current methods.
• Bibliographic metadata offers potential for enhanced performance in this task.

Purpose of the Study:

• To introduce an approach for biomedical relation classification using qualifiers of co-occurring Medical Subject Headings (MeSH).
• To present MeSH2Matrix, a novel dataset of 46,469 biomedical relations from PubMed.
• To evaluate the effectiveness of MeSH qualifiers in improving classification models.

Main Methods:

• Developed MeSH2Matrix dataset mapping MeSH qualifier associations and Wikidata relation types.
• Curated 46,469 biomedical relations from PubMed publications.
• Trained and evaluated three machine learning models (SVM, D-Model, C-Net) on the MeSH2Matrix dataset.

Main Results:

• The best model achieved 70.78% accuracy for 195 relation classes and 83.09% for five superclasses.
• Feature analyses examined the relationship between MeSH qualifiers and classified biomedical relations.
• Confusion matrices provided detailed insights into model performance.

Conclusions:

• Utilizing MeSH qualifiers offers a promising avenue for enhancing biomedical relation classification.
• The MeSH2Matrix dataset and approach can lead to more accurate biomedical ontology classification.
• Publicly accessible dataset and code facilitate reproducibility and further research.