MEDRank: using graph-based concept ranking to index biomedical texts.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Information Retrieval

Background:

• The exponential growth of biomedical text necessitates efficient automatic indexing.
• Current methods struggle to differentiate central concepts from incidental mentions.
• The National Library of Medicine (NLM) uses the Medical Text Indexer (MTI) to assist human indexers in labeling MEDLINE records.

Purpose of the Study:

• To enhance the Medical Text Indexer's (MTI) capability in identifying core concepts, termed "major headings," within MEDLINE abstracts.
• To introduce and evaluate MEDRank, a novel graph-based methodology for ranking biomedical concepts.
• To improve the precision and recall of automatic indexing for biomedical literature.

Main Methods:

• Integrating the MEDRank algorithm into the existing MTI system.
• Comparing the indexing output of MTI with and without MEDRank against human-selected terms for 11,803 PubMed Central articles.
• Conducting a human-rated evaluation of indexing terms generated by MEDLINE indexers, MTI without MEDRank, and MTI with MEDRank on 36 articles.

Main Results:

• MEDRank significantly boosted the recall of "major headings" by 30% compared to MTI alone (0.489 vs. 0.376).
• Overall recall saw a modest increase of 6.5% (0.490 vs. 0.460), with F(2) improving by 3% (0.408 vs. 0.396).
• Human raters showed a preference for terms generated by MTI with MEDRank, indicating improved relevance and alignment with expert indexing.

Conclusions:

• MEDRank substantially enhances the retrieval of core concepts in MEDLINE abstracts, outperforming MTI without the enhancement.
• The MEDRank-enhanced MTI system demonstrates improved recall and better alignment with human indexing expectations.
• This graph-based approach offers a promising advancement for automatic indexing in the biomedical domain.