Semantic Similarity

Calculate similarity by analyzing the shortest path or distance between two concept nodes in the ontological graph. The fundamental assumption is that concepts closer in the hierarchy are more semantically related.

• Shortest Path Length: A simple edge count between two nodes.
• Wu & Palmer: Scales the depth of the Least Common Subsumer (LCS) by the sum of depths of the two concepts.
• Leacock & Chodorow: Normalizes the shortest path length by the maximum depth of the taxonomy.

Example: In SNOMED CT, the path between Pneumonia and Lung Cancer is shorter (both are lung disorders) than the path between Pneumonia and Femur Fracture.

Weight concepts based on their specificity or rarity within a corpus. A rare, deeply nested concept carries more information than a common, shallow one. Similarity is a function of the shared information between two concepts.

• Resnik: Similarity equals the IC of the Least Common Subsumer (LCS).
• Lin: Normalizes the shared IC by the sum of the individual ICs of the two concepts.
• Jiang & Conrath: Uses the difference between the sum of individual ICs and twice the shared IC to compute a distance.

Corpus Dependency: IC values can be derived intrinsically from the ontology structure or extrinsically from term frequency in a clinical text corpus.

Represent concepts as sets of defining characteristics or logical properties rather than just nodes in a graph. Similarity is computed using set theory operations on these feature sets.

• Tversky Index: An asymmetric model where the weight of common and distinctive features can be tuned, reflecting human similarity judgments.
• Common Features: The intersection of properties shared by two concepts.
• Distinctive Features: Properties unique to one concept but not the other.

Application: Highly effective in Description Logic-based ontologies like OWL, where concepts have rich axiomatic definitions beyond simple hierarchical placement.

Combine structural ontology knowledge with distributional semantics from neural language models to overcome the limitations of purely graph-based or purely text-based approaches.

• Ontology-Aware Embeddings: Graph neural networks encode the hierarchical structure and relational edges of an ontology into dense vector representations.
• BERT-based Alignment: Contextual embeddings from transformer models capture nuanced synonymy and paraphrasing that rigid ontology structures might miss.
• Cross-Lingual Mapping: Embedding spaces allow similarity computation between concepts from ontologies in different natural languages.

Advantage: Mitigates the lexical gap where semantically identical concepts have completely different surface forms or labels.

Leverages formal Description Logic reasoning to determine semantic relationships beyond simple distance. A reasoner infers implicit subsumption relationships based on the logical axioms of the ontology.

• Equivalence: The reasoner determines if two concepts are logically identical based on their necessary and sufficient conditions.
• Subsumption Testing: Checks if one concept is a strict subclass of another, implying a directed similarity relationship.
• Satisfiability: Ensures a concept definition is logically consistent before computing similarity.

Clinical Relevance: Critical for ensuring that a mapped concept in ICD-10-CM is a valid, more granular subtype of a broader SNOMED CT concept, preventing semantic drift in data aggregation.

Adjusts similarity scores based on the specific task, user intent, or clinical context rather than relying solely on universal ontological distance. A measure of fitness-for-purpose.

• Task-Specific Weighting: A path-based measure might be weighted differently for billing code selection vs. clinical research cohort identification.
• Temporal Context: The similarity of two diagnoses might increase if they occur within the same patient encounter.
• Value Set Binding: Similarity is constrained by whether a candidate concept is a member of a specific, authorized value set for a given data element.

Example: Diabetes Mellitus and Hyperglycemia are highly similar in a lab result context but less similar in a primary diagnosis context where specificity is required.