Automated Mapping of Observation Archetypes to SNOMED CT Concepts - Foundations on Natural and Artificial Computation - page 556

Table 2. Mapping results to SNOMED CT concepts

Retrieved Retrieved Relevant Precision Recall
Concepts # Concepts # %%

Alg. 1:

13 11 85 55

Exact Searches

Alg. 1: Exact and

55 17 31 85

Truncate Searches

Algorithm 1:

87 20 23 100

complete

Algorithm 2 11 11 100 55

Algorithm 2 10 9 90 9

Algorithm 3 55 50 91 51

Algorithm 3 6 6 100 4

Algorithm 3 +

127 121 95 83

mapping value nodes

project [3]. All of selected archetypes fulfill the following requirements: they are
observable archetypes and their root nodes map to a SNOMED CT concept
descendant of the top level concept observable entity . The reason of establishing
this requirement is that we have found a correspondence between data sections
nodes of this type of archetypes and the SNOMED CT concepts descendant of
the top level observable entity .
Thirteen root nodes were lexically mapped to some SNOMED CT concept
using the exact match Algorithm 1, although only eleven root nodes were correctly
match. This represents 85% precision and 55% recall of the Algorithm 1 (Table 2).
When approximate lexical techniques are applied (Algorithm 1: complete), 100%
recall is reached (against 23% precision).
Two hundred and fifty-one element and value nodes were structurally mapped
to some SNOMED CT concept. Nine element nodes were correctly mapped using
the algorithm 2 with 90% precision and 9% recall. Applying algorithm 3, 50 ele-
ment nodes were correctly mapped with 91% precision and 51% recall. These re-
sults confirm the need of applying a combination of named- and structure-based
method in order to increase the recall. On the other hand, only six value nodes were
correctly mapped using the algorithm 3 with 100% precision and 4% recall. When
value nodes were directly mapped to SNOMED CT by applying a named-based
method, 121 nodes were reached, with 95% precision and 83% recall.
For the root node mapping, we can see in table 2 that the Algorithm 1:
complete (using UMLS) obtains a very high recall and a low precision. Howe-
ver, Algorithm 2 (using lexical matching) procures opposite results for precision