A lexical-based approach for exhaustive detection of missing hierarchical IS-A relations in SNOMED CT.

Fengbo Zheng, Jay Shi, Licong CuiPubMed 2020

TL;DR

This paper introduces a novel lexical-based method to automatically discover missing hierarchical relationships (IS-A relations) within SNOMED CT, a massive medical ontology. By analyzing concept names and their ancestors, the approach identifies potentially overlooked connections, significantly improving the completeness of this critical healthcare knowledge base.

What Problem Does This Paper Address?

Ontologies, like SNOMED CT, are foundational for many biomedical applications, including data annotation, integration, and clinical decision support. However, these ontologies are rarely perfect and can suffer from incompleteness, meaning crucial relationships between concepts might be missing. This incompleteness can lead to flawed analyses, missed diagnoses, and suboptimal healthcare delivery. For example, if a specific type of cancer isn't correctly classified as a subtype of a broader cancer category, patients with that specific cancer might be excluded from relevant clinical trials or treatment protocols. Existing methods for finding these missing links often rely on complex structural analysis or are limited in scope. This paper aims to address this by providing a more exhaustive and efficient way to uncover these missing IS-A relations using only the textual information within the ontology.

Methods

The core of this approach is a sophisticated lexical analysis of SNOMED CT concepts. Instead of just looking at individual words, the method considers both words and noun phrases within a concept's "Fully Specified Name" (FSN) and, crucially, inherits these features from its ancestors.

Here's a breakdown of the key steps:

  1. Stop Word and Antonym Identification: To prevent spurious suggestions, the researchers first identify common "stop words" (e.g., "due to," "and") and antonym pairs (e.g., "right" vs. "left") that could lead to incorrect inferences.
  2. Lexical Feature Construction:
    • Preprocessing FSNs: Concept names are cleaned, removing semantic tags and special characters.
  3. Initialization with Noun Phrases: The Stanford CoreNLP parser is used to identify noun phrases within the FSN. This is a key innovation, as it captures more nuanced meaning than a simple bag-of-words model. For instance, "acute sensitivity" is treated as a distinct feature from just "acute" and "sensitivity."
    • Enrichment with Ancestor Features: The method iteratively enriches a concept's feature set by incorporating features from its ancestors. This is done in two steps: first, by adding noun phrases from non-hierarchically linked concepts, and second, by inheriting features from all ancestors (excluding those with stop words). This creates a rich, context-aware representation for each concept.
  4. Subset Inclusion Checking: Candidate concept pairs (those within the same sub-hierarchy, not already linked, sharing a semantic tag, and free of stop words/antonyms) are compared. If the enriched feature set of one concept is a superset of another, an IS-A relation is suggested.
  5. Redundancy Removal: A final step filters out redundant suggestions, ensuring that only direct or implied relationships are proposed.

Datasets

The primary dataset used is the September 2017 release of SNOMED CT (US edition). This is a large-scale, widely used clinical healthcare terminology. The paper does not introduce a new benchmark dataset but rather applies its method to an existing, well-established ontology.

Key Insights and Ideas

  • The Power of Noun Phrases: Moving beyond simple word counts to incorporate noun phrases significantly improves the accuracy of identifying semantic relationships. This captures how words modify each other to create specific meanings.
  • Inheritance is Key: Lexical features inherited from ancestors are crucial for understanding a concept's place in the hierarchy. This approach systematically leverages this inheritance, creating a more comprehensive feature set.
  • Lexical Clues to Ontology Quality: The method not only finds missing links but also implicitly highlights existing problematic IS-A relations. When the algorithm produces false positives, it often points to an incorrect existing relationship that needs review.
  • Exhaustive Search is Feasible: The paper demonstrates that a comprehensive, ontology-wide search for missing IS-A relations is achievable using lexical analysis, rather than being limited to specific problematic subgraphs.

Key Results

The approach was applied to SNOMED CT, suggesting a total of 38,615 potentially missing IS-A relations. For evaluation, a domain expert reviewed a random sample of 100 suggestions from the "Clinical finding" sub-hierarchy, confirming 90 as valid, yielding a precision of 90%. This high precision indicates the effectiveness of the lexical feature enrichment and comparison strategy. The analysis also revealed that a significant portion (94.6%) of the detected missing relations were novel compared to previous work by the same research group.

Limitations and Future Work

The authors acknowledge several limitations:

  • Evaluation Scope: The manual evaluation was limited to a single sub-hierarchy ("Clinical Finding"). A broader evaluation across all sub-hierarchies would be beneficial.
  • Manual Review Burden: Manual review by domain experts is time-consuming. Future work could involve leveraging external knowledge sources (like biomedical literature or other ontologies) to automatically provide supporting evidence for suggested relations, reducing the manual effort.
  • Noun Phrase Recognition: The reliance on the Stanford CoreNLP Parser means that inaccuracies in noun phrase identification can affect results.
  • Remediation Complexity: Simply adding suggested IS-A relations might not be sufficient, as SNOMED CT's hierarchy is also based on logical definitions. Future work should explore how to infer missing relations from logical definitions themselves.
  • Recall Measurement: Measuring recall is challenging due to the lack of a ground truth for "false negatives." The authors suggest using changes in SNOMED CT versions over time as a potential surrogate.

Future work also includes applying a recent approach for suggesting missing IS-A relations in Gene Ontology to SNOMED CT and comparing the results.

Summary updated

Abstract

Incompleteness of ontologies affects the quality of downstream ontology-based applications. In this paper, we introduce a novel lexical-based approach to automatically detect potentially missing hierarchical IS-A relations in SNOMED CT. We model each concept with an enriched set of lexical features, by leveraging words and noun phrases in the name of the concept itself and the concept's ancestors. Then we perform subset inclusion checking to suggest potentially missing IS-A relations between concepts. We applied our approach to the September 2017 release of SNOMED CT (US edition) which suggested a total of 38,615 potentially missing IS-A relations. For evaluation, a domain expert reviewed a random sample of 100 missing IS-A relations selected from the "Clinical finding" sub-hierarchy, and confirmed 90 are valid (a precision of 90%). Additional review of invalid suggestions further revealed incorrect existing IS-A relations. Our results demonstrate that systematic analysis of the enriched lexical features of concepts is an effective approach to identify potentially missing hierarchical IS-A relations in SNOMED CT.
BibTeX
@inproceedings{zheng2020lexical,
  author = {Fengbo Zheng and Jay Shi and Licong Cui},
  title = {{A lexical-based approach for exhaustive detection of missing hierarchical IS-A relations in SNOMED CT.}},
  booktitle = {PubMed},
  year = {2020},
}
Figure