Date of Award

Fall 2007

Document Type

Dissertation

Degree Name

Doctor of Philosophy in Computing Sciences - (Ph.D.)

Department

Computer Science

First Advisor

Yehoshua Perl

Second Advisor

James Geller

Third Advisor

James J. Cimino

Fourth Advisor

Barry Cohen

Fifth Advisor

Huanying Gu

Sixth Advisor

Michael Halper

Abstract

The Unified Medical Language System (UMLS) is a two-level biomedical terminological knowledge base, consisting of the Metathesaurus (META) and the Semantic Network (SN), which is an upper-level ontology of broad categories called semantic types (STs). The two levels are related via assignments of one or more STs to each concept of the META.

Although the SN provides a high-level abstraction for the META, it is not compact enough. Various metaschemas, which are compact higher-level abstraction networks of the SN, have been derived. A methodology is presented to evaluate and compare two given metaschemas, based on their structural properties. A consolidation algorithm is designed to yield a consolidated metaschema maintaining the best and avoiding the worst of the two given metaschemas. The methodology and consolidation algorithm were applied to the pair of heuristic metaschemas, the top-down metaschema and the bottom-up metaschema, which have been derived from two studies involving two groups of UMLS experts. The results show that the consolidated metaschema has better structural properties than either of the two input metaschemas. Better structural properties are expected to lead to better utilization of a metaschema in orientation and visualization of the SN. Repetitive consolidation, which leads to further structural improvements, is also shown.

The META and SN were created in the absence of a comprehensive curated genomics terminology. The internal consistency of the SN's categories which are relevant to genomics is evaluated and changes to improve its ability to express genomic knowledge are proposed. The completeness of the SN with respect to genomic concepts is evaluated and conesponding extensions to the SN to fill identified gaps are proposed.

Due to the size and complexity of the UMLS, errors are inevitable. A group auditing methodolgy is presented, where the ST assignments for groups of similar concepts are audited. The extent of an ST, which is the group of all concepts assigned this ST, is divided into groups of concepts that have been assigned exactly the same set of STs. An algorithm finds subgroups of suspicious concepts. The auditor is presented with these subgroups, which purportedly exhibit the same semantics, and thus he will notice different concepts with wrong or missing ST assignments. Another methodology partitions these groups into smaller, singly rooted, hierarchically organized sets used to audit the hierarchical relationships. The algorithmic methodologies are compared with a comprehensive manual audit and show a very high error recall with a much higher precision than the manual exhaustive review.

Share

COinS