Automatic MeSH Indexing: Revisiting the Subheading Attachment Problem

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Medical Information Retrieval

Background:

• The National Library of Medicine (NLM) relies on indexers to process a large volume of articles annually.
• The Medical Text Indexer (MTI) system assists indexers by recommending MeSH (Medical Subject Headings) main heading/subheading pairs.
• Previous research has primarily focused on improving main heading prediction, with less attention given to automated fine-grained subheading attachment.

Purpose of the Study:

• To address the under-researched problem of automated subheading attachment in medical text indexing.
• To significantly improve the performance of automated fine-grained indexing using advanced machine learning techniques.
• To evaluate the effectiveness of modern Convolutional Neural Network (CNN) classifiers for subheading prediction.

Main Methods:

• Utilized modern Convolutional Neural Network (CNN) classifiers to tackle the subheading attachment problem.
• Developed and trained CNN models on a dataset of medical articles for predicting MeSH main heading/subheading pairs.
• Compared the performance of the developed CNN models against the existing Medical Text Indexer (MTI) system.

Main Results:

• Achieved significant performance improvements in automated fine-grained indexing.
• The best performing CNN model demonstrated a 3.7% absolute improvement in precision compared to the current MTI.
• The best performing CNN model showed a 27.6% absolute improvement in recall compared to the current MTI.
• A manual review indicated that 70% of false positive predictions from the best model were acceptable as indexing.

Conclusions:

• Modern Convolutional Neural Networks offer a substantial advancement for automated subheading attachment in medical text indexing.
• The proposed CNN-based approach significantly outperforms the current Medical Text Indexer (MTI) in terms of precision and recall.
• The findings suggest a promising direction for enhancing the efficiency and accuracy of the NLM's indexing process.