Training a Convolutional Neural Network with Terminology Summarization Data Improves SNOMED CT Enrichment

Area of Science:

• Biomedical Informatics
• Artificial Intelligence

Background:

• Automated concept insertion into biomedical ontologies is challenging.
• Verifying IS-A (is-a) links is a crucial sub-problem for ontology development.
• Machine learning, specifically Convolutional Neural Networks (CNNs), shows promise for this task.

Purpose of the Study:

• To investigate the automatic verification of IS-A links between new and existing concepts in a biomedical ontology.
• To evaluate the impact of different negative training data strategies on CNN performance.
• To compare the effectiveness of using the complete SNOMED CT hierarchy versus the Area Taxonomy for training.

Main Methods:

• Utilized Convolutional Neural Networks (CNNs) for IS-A link verification.
• Employed SNOMED CT (July 2017 release) for training and a subsequent release for testing.
• Experimented with two negative training data approaches using uncle-nephew concept pairs.
• Constrained training data using the Area Taxonomy ontology summarization mechanism.

Main Results:

• The primary challenge identified was the generation of high-quality negative training data.
• Using the Area Taxonomy to constrain training data significantly improved performance compared to using the complete SNOMED CT hierarchy.
• An improvement of 8.6% in IS-A link verification accuracy was achieved by employing the Area Taxonomy.

Conclusions:

• The Area Taxonomy summarization mechanism is an effective strategy for improving the quality of training data for CNN-based IS-A link verification.
• This approach enhances the accuracy of automated biomedical concept relationship identification.
• The findings contribute to the development of more robust automated ontology construction systems.