2005 Scientific Session Abstracts
Alignment of Custom Semantic Spaces to Standardized Knowledge Sources
Bonnie L. Zeigler, PhD (bzeigler@ardais.com); David Aronow, MD, MPH; Kristel Hackett, BS Clinical Informatics Group, Ardais Corporation, Lexington, MA
Context: Ardais has developed custom terminologies (Ardais Terminology Service, or ARTS) supporting the phenotypic annotation of human tissue. We have applied ARTS in our commercial software environment for three years and have recently extended the terminologies in a research mode to capture clinical data pertinent to the understanding of the genetic basis of disease.
The original ARTS core, required for the attributes Diagnosis, Tissue, and Surgical Procedure, was SNOMED-based. As our clinical data expanded to new semantic domains (e.g., diagnostics), we added customized local concepts to ARTS. With the advent of SNOMED-CT, UMLS, and federally-driven informatics efforts (EVS, Ca-BIG), we deemed it necessary to examine the potential for alignment of ARTS with current standardized knowledge sources.
Technical Design: We applied a narrow-beam alignment approach to semantic domains that were well-standardized, especially those that were part of the ARTS core. The goal of narrow beam alignment was to achieve nearly 100% mapping of Ardais concepts to standard concepts, with minimal use of local customization. To identify candidates for alignment, we applied UMLS Norm to the ARTS concepts and compared our concepts with normalized concepts in UMLS.
We developed a broad-beam approach to less-well-standardized semantic domains. The goal of broad-beam alignment was to evaluate the potential for mapping Ardais concepts to standardized terminologies and to understand the pragmatics that would influence alignment in a particular context. Techniques for broad-band alignment included automated string comparisons between Ardais concepts and UMLS and LOINC knowledge sources, as well as use of exploration environments (Apelon Termworks) to identify potential matches.
Results: Narrow-beam analysis for Diagnosis, Procedure, and Tissue concepts revealed both semantic drift of existing SNOMED-based concepts and good potential for alignment of Ardais concepts to SNOMED-CT. Over ARTS as a whole, SNOMED alignment improved from 76% to 85%, with the most improvement in the area of Surgical Procedures (70% to 88%). Concepts originally in the SNOMED core were aligned from 93% to 100%.
Example results from our broad domain of Diagnostics suggest that Chemistry tests offer potential for about 70% alignment, IHC/Immunophenotype tests about 40%, and Genetic Studies about 20%. Factors impeding genetic-test mapping between local and standard knowledge sources included a rather low representation of genes in standard sources, as well as differences in granularity and modeling approach between custom and standard terminologies.
Conclusion: Vendors who develop customized terminology systems for targeted application areas must assess their ability to align with standard knowledge sources as such sources become more available and as their use is mandated by customers. Ardais has achieved close conformance with standard sources in traditional semantic domains. For newer domains, where conformance is more difficult, we have used our broad-beam strategy to target domains with the most potential for alignment and to understand how application context will influence our ability to exploit emerging terminology standards.