Automatic Annotation of PubMed Articles with MeSH Qualifiers.

Area of Science:

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

Background:

• PubMed utilizes Medical Subject Headings (MeSH) for indexing millions of biomedical documents.
• Manual MeSH annotation is time-consuming and can be challenging for complex subspecialties.
• Accurate indexing is crucial for effective clinical and research article retrieval.

Purpose of the Study:

• To develop and evaluate a machine learning procedure for predicting MeSH qualifiers.
• To compare the performance of machine learning models against manual PubMed annotation for heart transplantation articles.

Main Methods:

• Trained binary classifiers (logistic regression with tfidf, DistilBERT) to predict MeSH qualifiers for Heart Transplantation descriptor.
• Utilized PubMed abstracts for model training and evaluation.
• Performed a small-scale evaluation using a test set manually re-annotated by a cardiac surgeon.

Main Results:

• The DistilBERT model achieved a macroaveraged F1 score of 0.85, outperforming manual PubMed annotation (0.76).
• The logistic regression model achieved a macroaveraged F1 score of 0.81.
• Both models demonstrated superior performance compared to the existing manual annotation process.

Conclusions:

• Machine learning models can accurately predict MeSH qualifiers, potentially assisting human annotators.
• This approach promises to enhance the efficiency and quality of biomedical literature indexing.
• The method is extensible to other MeSH descriptors with sufficient training data, improving information retrieval for clinicians and researchers.