Comparative analysis of classification techniques for topic-based biomedical literature categorisation.

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Pharmacovigilance

Background:

• Scientific literature volume growth challenges information extraction.
• Accurate identification of drug-induced liver injury (DILI) is critical for patient safety and drug development.
• Automated classification methods are needed to efficiently process biomedical texts for DILI information.

Purpose of the Study:

• To compare various text classification methods for identifying DILI information in scientific articles.
• To develop a fast, interpretable text classification approach for domain-specific categorization.
• To address data imbalance challenges in biomedical text classification.

Main Methods:

• Comparison of transformers, LSTMs, information theory, and statistics-based text classification methods.
• Development of a novel, interpretable text classification model.
• Implementation of techniques to handle imbalanced datasets.

Main Results:

• Transformers perform best when training and test data distributions match.
• Simple statistical and information theory models surpass complex transformers on imbalanced data, offering superior interpretability.
• Domain-specific pre-training and tailored loss functions improve neural network performance for imbalanced biomedical data.

Conclusions:

• While transformers are powerful, simpler statistical methods can be more efficient and explainable for topic classification in biomedicine.
• Combining strengths of complex and simple models offers promising avenues for future research.
• Further development of neural network architectures and training strategies is needed for robust classification of imbalanced biomedical data.