Thesaurus-based word embeddings for automated biomedical literature classification.

Area of Science:

• Biomedical informatics
• Natural Language Processing
• Machine Learning

Background:

• Biomedical literature presents unique challenges for automated classification due to its volume and complexity.
• Manual indexing is labor-intensive, expensive, and prone to errors.
• Existing automated methods struggle with the broadness and multi-label nature of the domain.

Purpose of the Study:

• To investigate the efficacy of current word embedding algorithms for biomedical text classification.
• To leverage the Medical Subject Headings (MeSH) ontology for machine-readable labels and problem dimensionality.
• To explore deep and shallow network architectures, including transfer learning, for improved classification.

Main Methods:

• Utilized word embedding algorithms to extract features from contextualized representations of abstracts.
• Employed deep and shallow neural network approaches for classification.
• Integrated the Medical Subject Headings ontology for label representation.
• Evaluated a transfer learning approach with a separate classifier.
• Trained and tested models on large datasets of biomedical citations.

Main Results:

• Word embedding algorithms show promise in supporting multi-label biomedical text classification.
• Deep and shallow network approaches, combined with MeSH ontology, provide effective feature extraction.
• Transfer learning can further enhance classification performance.
• Automated methods, while not replacing experts, offer valuable validation and recommendation capabilities.

Conclusions:

• Word embedding techniques offer an efficient way to enhance automated biomedical text classification, particularly in multi-label scenarios.
• The integration of MeSH ontology and advanced network architectures improves classification accuracy and utility.
• These automated systems serve as valuable tools for assisting, not replacing, human experts in biomedical literature analysis.