Identifying Missing IS-A Relations in SNOMED CT with Fine-Tuned Pre-trained Language Models and Non-lattice...

Area of Science:

• Medical Informatics
• Computational Linguistics
• Knowledge Representation

Background:

• Completeness of IS-A relations in SNOMED CT is vital for clinical accuracy.
• Existing methods may not fully capture all hierarchical relationships.

Purpose of the Study:

• To develop and evaluate a hybrid approach using pre-trained language models (PLMs) to identify missing IS-A relations in SNOMED CT.
• To assess the performance of various fine-tuned PLMs in this task.

Main Methods:

• A hybrid approach combining non-lattice subgraphs and PLMs was employed.
• Eight PLMs (BERT, DistillBERT, DeBERTa, BioClinicalBERT, BioMistral, Llama3, Gemma2, Phi-4) were fine-tuned.
• Consensus predictions from all models were used to identify missing IS-A relations.

Main Results:

• DeBERTa demonstrated superior performance with precision of 0.96, recall of 0.97, and an F1-score of 0.965.
• The approach identified 678 potential missing IS-A relations in SNOMED CT (March 2023 US Edition).
• Manual review of 100 selected cases confirmed 93 as valid (93% precision).

Conclusions:

• Fine-tuned PLMs are effective in detecting missing IS-A relations within non-lattice subgraphs.
• This method offers a promising strategy for enhancing the quality and completeness of SNOMED CT.
• The findings support the use of AI-driven approaches for clinical terminology maintenance.