Multilabel associative classification categorization of MEDLINE articles into MeSH keywords

Area of Science:

• Medical Informatics
• Data Mining
• Natural Language Processing

Background:

• MEDLINE database documents require multi-label classification due to multiple category assignments.
• Scalable methods are essential for handling hundreds of thousands of medical documents.

Purpose of the Study:

• To develop a novel, scalable system for automated multi-label classification of MEDLINE documents to MeSH keywords.
• To adapt the ACRI data mining algorithm for effective multi-label classification tasks.

Main Methods:

• Modification of the ACRI algorithm to support multi-label classification.
• Testing five distinct classification configurations and various quality measurement methods.
• Experimental comparison of word reoccurrence-based methods against non-recurrent associative classification.

Main Results:

• The proposed system achieved a macro F1 score of 46%, demonstrating high quality on the challenging MEDLINE dataset.
• The classifier's accuracy reached 90%, calculated as the ratio of true positives and true negatives to total examples.
• Word reoccurrence-based methods showed superiority over non-recurrent associative classification.

Conclusions:

• The developed system offers a high-quality solution for the multi-label classification of MEDLINE documents.
• Different configurations optimize for classifying numerous documents (micro F1) or categories with fewer documents (macro F1).
• A balanced approach optimizing the average of macro and micro F1 provides a trade-off between classification objectives.