Automated Mapping of Observation Archetypes to SNOMED CT Concepts - Foundations on Natural and Artificial Computation - page 557

and recall. For the other nodes, when the algorithm uses the SNOMED CT
relationships (Algorithm 2 for element nodes and Algorithm 3 for value nodes),
we obtain an important precision and low recall which is improved with lexical
mapping techniques to SNOMED CT.
5 Discussion and Future Work
After creating the gold standard manually, we detected that 27.7% of nodes of
the selected archetypes have no correspondence with SNOMED CT. There are
two reasons for it:
- A 19.7% are nodes whose text are tagged with phrases like: comment, a-
dditional information, any event, normal statement, row, new element which
are too general or their tag does not represent clinical information and,
consequently, it have not correspondence with SNOMED concepts. We could
skip mapping this kind of nodes because we extracted the list of avoidable
phrases.
- The other 80.3% are nodes uncovered by SNOMED. Our method is focussed
on mapping archetype nodes to pre-coordinated SNOMED-CT concepts.
However, in some cases more than one SNOMED CT concept is needed to
explain the archetype node. For example, the node waist and hip circum-
ference will map to the SNOMED concepts waist circumference (observable
entity) and hip circumference (observable entity) . So, post-coordination tech-
niques are needed for mapping composite expressions of archetype nodes.
Other approaches [20], [16] and [21] use lexical and context methods for the
mapping process. Our approach, in addition to the mentioned techniques, takes
advantage of the SNOMED CT relationships interprets and IS A .Ontheone
hand, our algorithm exploits the similarity between the relationships element-
values of the archetypes nodes and the SNOMED CT relationship interprets .
For example, in Fig. 3 the values shallow , normal and deep of the element node
depth would never lexically match to the corresponding SNOMED CT concepts
because of the word breathing . However, our algorithm can traverse the relation
interprets of the SNOMED CT concept depth of respiration , and it applies par-
tial lexical matching with the value nodes of element Depth in the archetype.
On the other hand, the algorithm also profits from the correspondence between
the relationships root-elements in archetype nodes and the SNOMED CT rela-
tionship IS A . For example (Fig. 3), after mapping the root node respiration to
the SNOMED CT concept respiration observable ,weusethe IS A relationship
for mapping the elements depth , rate and rhythm to the SNOMED CT concepts
depth of respiration , respiratory rate and rhythm of respiration . In spite of ob-
taining a low recall with the relationship traversing application, the concepts
retrieved offer a more appropriate semantics to the archetype. Therefore, this
technique provides us a method of validation and disambiguation.