A deep learning approach to identify missing is-a relations in SNOMED CT.

Area of Science:

• Medical Informatics
• Computational Linguistics
• Artificial Intelligence

Background:

• SNOMED CT is the world's largest clinical terminology, essential for accurate healthcare applications.
• Ensuring the quality and completeness of SNOMED CT is vital for reliable clinical decision support and data interoperability.
• Identifying missing hierarchical relationships (is-a relations) is a key aspect of SNOMED CT quality assurance.

Purpose of the Study:

• To develop and evaluate a deep learning-based method for uncovering missing is-a relations within SNOMED CT.
• To specifically target missing relations between concept pairs exhibiting a containment pattern.

Main Methods:

• A binary classification model was trained using concept name, hierarchical, lexical, and logical definition features.
• Hierarchically related containment concept pairs were used as positive training instances, and unrelated pairs as negative instances.
• A cross-validation inspired approach was employed to identify potential missing relations.

Main Results:

• The model achieved high performance on the Clinical Finding subhierarchy, with a precision of 0.8164, recall of 0.8397, and F1 score of 0.8279.
• A total of 1661 potential missing is-a relations were identified.
• Expert evaluation confirmed the validity of 83.48% of a randomly selected subset of these potential relations.

Conclusions:

• The deep learning approach effectively identifies missing is-a relations in SNOMED CT, particularly within containment patterns.
• This method contributes to enhancing the accuracy and completeness of the SNOMED CT knowledge base.
• The findings support the use of AI for automated quality assurance in large-scale clinical terminologies.