TL;DR
This paper introduces a novel method combining Formal Concept Analysis (FCA) and Semantic Query Expansion (SQE) to help researchers and clinicians define complex health outcomes of interest (HOIs) from Electronic Health Records (EHRs). By leveraging ontologies and user feedback, it aims to improve the precision and recall of HOI searches, making it easier to identify patient cohorts for studies.
What Problem Does This Paper Address?
Identifying specific groups of patients based on health outcomes (like "myocardial infarction" or "diabetes mellitus") is crucial for research using Electronic Health Records (EHRs). However, defining these "Health Outcomes of Interest" (HOIs) is a complex and time-consuming task. Clinicians and researchers often struggle with the sheer volume and intricate structure of medical terminologies and ontologies (like UMLS, which contains millions of concepts). Simply searching for terms can lead to too many irrelevant results or miss important variations. Existing methods, like the "2-hop" approach, require significant manual effort from medical experts to identify concept unique identifiers (CUIs). This paper addresses the need for a more automated and user-friendly way to define HOIs, reducing the burden on experts and improving search accuracy.
Methods
The core of this paper is the integration of two techniques:
- Semantic Query Expansion (SQE): This process takes a user's initial free-text query (e.g., "pituitary cancer") and expands it by traversing the hierarchical relationships within ontologies. It identifies all "descendant" concepts (more specific terms) and "ancestor" concepts (broader terms) to broaden the search space and capture related terms that might not be explicitly in the initial query.
- Formal Concept Analysis (FCA): FCA is used to analyze the relationships between concepts and attributes within the ontologies. It builds a "concept lattice," which is a structured representation of these relationships. This lattice helps to:
- Identify improbable relations: By analyzing the structure, the system can flag relationships that might not be intuitively obvious.
- Prune irrelevant concepts: FCA helps to identify concepts that are too broad or too specific and don't align well with the user's intent. A "pruning threshold" is used to automatically remove less relevant branches of the lattice.
The system works by first performing a lexical match of the user's query against ontologies. Then, SQE expands this set of initial matches by navigating up the ontology hierarchy. FCA is then applied to this expanded set to structure the concepts and identify the most relevant ones. The user provides feedback on the proposed concepts, which helps to refine the search space further by pruning less relevant parts of the FCA lattice. This iterative process aims to balance broad coverage with high precision.
Datasets
The primary dataset used for evaluation is the 2008 i2b2 Obesity Challenge corpus. This corpus is a gold-standard benchmark that includes manual annotations for 16 conditions and their associated events in clinical notes. The authors used this dataset to evaluate their method's sensitivity and specificity, comparing it against the existing "2-hop" method.
Key Insights and Ideas
- Bridging the gap between plain language and complex ontologies: The system allows users to start with simple, natural language queries, abstracting away the need for deep knowledge of ontology structures and CUIs.
- FCA as a powerful pruning mechanism: FCA's ability to create structured concept lattices is key to efficiently identifying and removing irrelevant concepts that SQE might introduce, thus improving precision.
- Iterative refinement with user feedback: The system is designed to be interactive. User feedback on proposed concepts is crucial for guiding the FCA pruning process and ensuring the final HOI definition accurately reflects the user's intent.
- Discovery of novel positive results: The method demonstrated the ability to find relevant concepts that were not present in the original i2b2 Obesity NLP reference set, suggesting it can uncover valuable connections.
Key Results
The paper reports positive results for their FCA and SQE cooperation approach:
- Improved Sensitivity and Specificity: The method achieved an average sensitivity of 77.3% and specificity of 99.1% on the i2b2 Obesity dataset, which is an improvement over the previously published "2-hop" method (74% sensitivity, 96% specificity).
- Better Recall with High Precision: The approach provides better recall (finding more relevant items) while maintaining high precision (minimizing irrelevant items).
- Efficient Processing: While processing times varied (0.6 to 67.3 seconds), the authors note that these are for complex HOIs and that there's potential for performance gains.
Limitations and Future Work
- Limitations:
- The current evaluation is based on a specific dataset (i2b2 Obesity).
- The full evaluation of the Web interface's efficiency and interactivity with physicians in real-world scenarios is still pending.
- The processing duration can be significant for very complex HOIs due to the size of the ontologies and lattices.
- Future Work:
- Conduct user-driven evaluations of the Web interface with physicians.
- Analyze physician acceptance/rejection patterns to further refine the method.
- Explore active learning mechanisms to improve future queries based on past refinements.
- Apply the method to mining EHR clinical notes and cohort building tools.
- Extend the approach to support other domains beyond health outcomes.
Summary updated