=Paper=
{{Paper
|id=Vol-2137/paper_34.pdf
|storemode=property
|title=Scrutinizing the Relationships Between SNOMED CT Concepts and Semantic Tags
|pdfUrl=https://ceur-ws.org/Vol-2137/paper_34.pdf
|volume=Vol-2137
|authors=Jonathan Bona,Werner Ceusters
|dblpUrl=https://dblp.org/rec/conf/icbo/BonaC17
}}
==Scrutinizing the Relationships Between SNOMED CT Concepts and Semantic Tags==
https://ceur-ws.org/Vol-2137/paper_34.pdf
Bona and Ceusters
Scrutinizing the relationships between SNOMED CT concepts and
semantic tags
Jonathan Bona1,* and Werner Ceusters2
1
Department of Biomedical Informatics, University of Arkansas for Medical Sciences, Little Rock, USA
2
Department of Biomedical Informatics, University at Buffalo, Buffalo, NY
ABSTRACT mately subsumed by the highest-level concepts for morpho-
The fully specified name of a concept in SNOMED CT is formed by a term logic abnormalities and disorders respectively. Because STs
to which in the typical case is added a semantic tag (ST). An ST is meant to are substrings added to names inside FSNs and are not repre-
disambiguate homonymous terms and indicate where that concept fits into sented separately as part of SNOMED CT’s formal model, it
SNOMED’s massive concept hierarchy. We have developed a method to de- is not easy to determine whether a tag on a concept should be
termine whether or not a concept’s tag correctly identifies its place in the taken to mean that the concept is necessarily part of the same
hierarchy, and applied this method to an analysis of all active concepts in sub-hierarchy as others with that tag. A concept’s ST would
every SNOMED CT release from January 2003 to January 2017. Our results strictly identify its place within the hierarchy if each tag had
show that there are concepts in every release whose tags do not match their a single, high-level corresponding concept that used it, and
placement in the hierarchy. These tag/hierarchy mismatches appear to be er- every concept using the tag was below that high-level concept
rors. The number of such errors is increasing in recent versions. in the hierarchy. For instance: in the clinical finding hierar-
chy the highest finding concept, [404684003 | Clinical find-
1 INTRODUCTION ing (finding)], subsumes all other findings.
SNOMED CT is a large reference terminology for the clinical The exact relationship between SNOMED CT’s STs and
domain made up of 300,000+ active concepts with machine- concepts has thus far not been widely researched. In
readable logical definitions that can be used for logical infer- (Ceusters & Bona, 2016) we explored how the STs of con-
ence (IHTSDO, 2015). SNOMED concepts are organized cepts changed over time. We found in total 285 patterns ac-
into a hierarchy of ‘Is-a’ relations. The top concept, cording to which SNOMED CT concepts underwent changes
138875005 | SNOMED CT Concept (SNOMED in the STs assigned to them -- a change from no ST at all to a
RT+CTV3) directly subsumes 19 high level concepts. This ST (43 patterns) counted also as a change. There were no pat-
includes first order concepts such as 404684003 | Clinical terns with more than 3 changes over time. Changes in STs
finding (finding), and 123037004 | Body structure (body were found to happen for a number of reasons. One is a
structure), which serve as the root of subhierarchies of con- change in SNOMED CT’s concept model, for instance when
cepts about entities directly relevant to and within the domain distinctions are made that didn’t exist in earlier versions, or
of healthcare. It includes also relations used amongst con- different interpretations were introduced (e.g. the product /
cepts in SNOMED CT as well as second order concepts that substance distinction). Such changes have a global impact on
describe the structure of SNOMED CT rather than the struc- large parts of the ontology. Another reason is that concepts
ture of what the first-order concepts of SNOMED CT are were in one or other way erroneous and had to be corrected.
about. Every SNOMED CT concept comes with descriptions While doing these analyses, we were nevertheless hampered
one of which is selected as the Fully Specified Name (FSN) by the fact that the SNOMED CT documentation available
and which typically ends in a semantic tag (ST) that disam- from the IHTSDO webserver provides insufficient infor-
biguates it from other concepts that may have similar names mation on what the precise set of STs the SNOMED CT edi-
(IHTSDO, 2015, p41). The ST also serves to indicate where tors are working with might be. The information that a ST is
the concept fits into the SNOMED CT concept hierarchy that what appears at the end of a FSN between brackets
(IHTSDO, 2017). For example, the concepts [35566002 | He- (IHTSDO, 2015, p41) turned out not to be reliable. Histori-
matoma (morphologic abnormality)] and [385494008 | cally, FSNs didn’t have a ST at all as this was apparently in-
Hematoma (disorder)] have the STs ‘morphologic abnor- troduced later as witnessed by the many changes in descrip-
mality’ and ‘disorder’ attached to the name they have in tions to that end. It was found that parsing anything that ter-
comon: hematoma. In the hierarchy, these concepts are ulti- minates a FSN between brackets leads to many false positives
in older concepts, thus requiring manual inspection for dis-
ambiguation. The work presented here examines the January
* To whom correspondence should be addressed: jonathanbona@gmail.com
31, 2017 International Release of SNOMED CT to investi-
1
K.Takahashi et al.
gate the extent to which SNOMED CT’s use of STs is sys- We define therefore the corresponding concept for any ST
tematic and consistent with its placement of concepts that use t as: the highest concept in the hierarchy that is tagged with
those STs within the concept hierarchy. Research hypotheses t. Note that this definition does not require tags to keep the
driving this work are: same corresponding concept across releases.
Based on this we determine the corresponding concept Ct
(1) All STs are related to the concept system through a
for each ST in a SNOMED release by:
one-to-one correspondence between the ST and
some high-level concept. Every concept that uses a (1) Calculating the whole number depth for each concept
particular ST t should be subsumed by that ST’s C as the length of the shortest Is-a path from the top
‘corresponding concept’ Ct, where Ct is the highest concept to C.
level concept that uses t. This hypothesis is moti- (2) For each ST t, select from the set of concepts tagged
vated by the apparent change in terminology from with t the concept with the least depth, Xt.
‘semantic tag’ in (IHTSDO, 2015) to ‘hierarchy (3) Let Ct = Xt if none of Xt’s ancestors is tagged with t.
tag’ in (IHTSDO, 2017). Otherwise let Ct be the ancestor of Xt that has the
(2) We consider a concept to be ‘mismatched’ if it has the least depth.
ST t but is not subsumed by the corresponding con-
Step 3 is necessary to handle special cases (Fig. 1) that arise
cept Ct.
from SNOMED CT’s use of multiple inheritance caused by
(3) Where such mismatches exist, they are due to errors its Is-a hierarchy forming a directed acyclic graph with a sin-
in the concept’s placement in the hierarchy or in its gle root node (SNOMED CT Concept) that has no edges com-
ST, and should be corrected in future releases. ing into it (i.e. is not subsumed by any other concept). Such
This paper reports on techniques we have developed to de- special cases occur whenever there is a concept with some ST
tect mismatched concepts, categorize them, and extract pat- t that is closest to the top as compared to all other concepts
terns to understand how they change over time as new ver- with ST t, and at the same time is also subsumed by another
sions of SNOMED CT are released. concept with ST t that has a longer shortest path to the top
concept. Such patterns were found in some releases making
2 METHODS it thus possible for a more general concept – i.e. higher up in
the hierarchy – for a ST to be subsumed by less general con-
We have developed computational procedures (1) to identify
cepts that use the same ST.
the concept that corresponds to an ST and (2) to facilitate an-
swering questions about subsumption that involve consider- 123037004
Body structure
ing all SNOMED CT concepts in each release. These are de- (body structure)
scribed in detail below.
442083009
2.1 Identifying tag corresponding concepts Anatomical or acquired
body structure
In order to determine whether a concept C is mismatched it (body structure)
is necessary to know which concept is the corresponding con-
91723000
cept for C’s ST. There does still not appear to be an official Anatomical structure
91832008
published mapping that lists the ST / concept correspond- Anatomical (body structure)
ences for SNOMED CT. In many cases this correspondence organizational pattern
may seem obvious to a human observer since for some tags (body structure) 4421005
Cell structure
there is a single high-level concept that uses the tag and (cell structure)
whose name is the same as the tag. For example, one direct
sub-concept of the top SNOMED CT Concept is 71388002 | 67185001
Subcellular structure
Procedure (procedure). This concept is has the ST ‘proce- (cell structure)
dure’ and its name in the FSN is the word ‘Procedure’.
11874005
In other cases, the correspondence is less obvious. For in- Distinctive arrangement
stance, no direct sub-concept of the top concept is tagged of cytoplasmic filaments
‘morphologic abnormality’, nor is there any concept whose (cell structure)
name is exactly ‘Morphologic abnormality’. The concept
Fig. 1. Effect of concept multiple inheritance on ST hierarchy.
118956008 | Body structure, altered from its original an-
atomical structure (morphologic abnormality) is a child of The output of this process is a mapping of STs to corre-
123037004 | Body structure (body structure) and appears sponding concepts for each release. This mapping is fairly
to be the highest concept (i.e. closest to the top) tagged with stable across releases, though there are some changes, which
‘morphologic abnormality’. we discuss more in the results section below.
2
Scrutinizing the relationships between SNOMED CT concepts and semantic tags
2.2 Identifying mismatched concepts cepts among their subsumers These were organized into a ta-
Once a corresponding concept has been identified for each ble with concepts as rows and SNOMED CT release dates as
tag in each release, it is possible to find mismatched concepts columns, with each cell indicating the concept’s category for
by looking at each concept in turn to see whether it is sub- that release. We took of course into account that SNOMED
sumed by the corresponding concept for its ST. We devel- CT concepts can be either active or inactive in a release and
oped computational procedures to do this. that a concept that is active in one release may be deactivated
In order to make use of the built-in subsumption reasoning in the next one, for instance if the concept was deemed by
provided by standard semantic web tools, we constructed an SNOMED CT’s editors to be no longer accurate or useful.
RDF/OWL model that represents SNOMED CT’s concept Less commonly, a concept that is inactive at one release may
hierarchy (300,000+ concepts connected by the is-a relation) be (re)activated at the next. We consider concepts to be not
and STs. Each concept is represented as an OWL class with active in releases that precede their addition to SNOMED CT.
separate annotations for its FSN and ST. Each of SNOMED’s The categories into which concepts were classified were
Is-a relations between concepts has a corresponding rdf:sub- constructed by building up a three-character code ‘_ _ _’
ClassOf assertion this representation. We built one such where each character is a flag indicating whether a certain
OWL file for each SNOMED CT release from January 2003 condition holds of the concept in that release. If a concept is
to January 2017. The identifiers (URIs) for each concept use inactive or did not yet exist at a release, then that concept was
a namespace that indicates the release version, e.g. marked with the three-character empty code ‘ ‘ for that re-
is an identifier for the lease. The following construction principles were used:
concept with concept id 64572001 in the January 31 2017 re- The first character is ‘Y’ if the concept is subsumed by
lease. These files were loaded into a single repository in a its ST’s corresponding concept in this release (i.e. if it is
triple store database (Bishop et al., 2011) configured for NOT mismatched in the release), and ‘N’ otherwise.
RDFS+ inference that, upon loading, pre-computed sub-
The second character is ‘Y’ if the concept has any ances-
sumption for each hierarchy, resulting in a total of 185 mil-
tor concept that is NOT mismatched. It is ‘N’ if every
lion triples. This facilitates very fast retrieval of subsumption
ancestor of this concept is mismatched.
information using simple SPARQL queries, and allows us to
instantly answer questions such as: given a release R, a tag t, The third character is ‘Y’ if the concept has any ancestor
and a concept C, which concept - if any - are tagged with t in concept that IS mismatched. It is ‘N’ if no ancestor of
R, but not subsumed by C in R? As an example, the following this concept is mismatched.
query retrieves the concept URI, label, and ST for every con- Combinatorically, this would allow us to code for nine dif-
cept that is not subsumed by 64572001 | Disease (disorder) ferent situations including the inactive concepts. However,
even though it uses the corresponding tag. given the meanings assigned to these codes, some combina-
tions are impossible. Ideally, every active concept in
PREFIX corr:
PREFIX tagged: < http://ex.com/r20170131#tagged>
SNOMED would be in the ‘YYN’ category, indicating that
PREFIX : the concept is properly matched to its ST’s corresponding
SELECT ?conc ?l ?tag concept, as are all of the concepts above it. Possible codes for
WHERE { mismatched concepts are 'NYY' and 'NYN' while for non-
?conc rdfs:label ?l .
?conc tagged: ?target_tag . mismatched concepts 'YYN' and 'YYY'. The latter indicates
corr: tagged: ?target_tag . a concept that itself is not mismatched, but it is subsumed by
?conc tagged: ?tag . at least one mismatched concept.
FILTER NOT EXISTS {?conc rdfs:subClassOf corr: }
}
Two flavors of mismatches were then looked for: (a) ‘local’ 3 RESULTS
mismatching as defined in assumption (2) above which oc- 3.1 Corresponding concept mappings
curs within the scope of a specific release, and (b) ‘global’
We used the corresponding concept discovery procedure pre-
mismatching in which the reference is the most recent release
sented above to construct a table with [STconcept] pairs
investigated. The group of globally mismatched concepts in-
and inspected this table manually to assess whether the map-
cludes thus those concepts which have in at least one version
pings made sense. In the majority of cases, the ST turned out
V a semantic tag which is different from the one it has in the
to be identical to the name of the corresponding concept mod-
last version, whether or not it is locally mismatched in V.
ulo capitalization and spacing. Exceptions were: [SNOMED
2.3 Characterizing mismatched concepts RT+CTV3SNOMED CT Concept], [metadata
We then group locally mismatched concepts into categories SNOMED CT Model Component], [Environment / lo-
based on the presence or absence of other mismatched con- cationEnvironment or geographical location], [Staging
3
Bona and Ceusters
Table 1: Per-tag counts of globally mismatched concepts for the January release of each year
0301 0307 0401 0407 0501 0507 0601 0607 0701 0707 0801 0807 0901 0907 1001
disorder 0 0 0 0 0 0 0 0 0 44 8 19 0 0 0
finding 4 1 0 5 0 0 0 0 0 0 0 0 0 0 0
observable entity 0 0 0 2 0 0 0 0 0 0 0 0 0 0 0
product 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1
regime/therapy 0 0 0 0 0 0 0 0 0 37 259 123 0 0 0
substance 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0
1007 1101 1107 1201 1207 1301 1307 1401 1407 1501 1507 1601 1607 1701 Total
disorder 0 0 0 0 2 4 10 26 32 44 24 74 78 83 188
finding 0 0 0 0 0 0 0 0 0 0 0 0 0 0 10
observable entity 0 0 0 0 0 0 0 0 0 0 0 0 0 0 2
product 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
regime/therapy 0 0 0 0 0 0 0 0 0 0 3 3 3 4 263
substance 0 0 0 0 0 0 0 0 0 0 0 0 0 1 2
Table 2: Per-tag counts of locally mismatched concepts per release.
scaleStaging and scales], [situationSituation with ex- release and examined this table for changes. We found that
plicit context], [assessment scaleAssessment scales], [re- the majority of tag to corresponding concept pairings are sta-
gime/therapyRegimes and therapies], [cellEntire cell], ble over all releases. A few are absent initially but appeared
[morphologic abnormalityBody structure, altered from its when their tag was added to SNOMED, e.g.415229000 | Ra-
original anatomical structure], [geographic locationGeo- cial group (racial group) appears for the first time in Janu-
graphical and/or political region of the world], [prod- ary 2005.A few are present initially but disappeared when
uctPharmaceutical / biologic product], and [disorderDis- they were removed from SNOMED, e.g. 304813002 | Ad-
ease]. These seem to be plausible mappings. We then con- ministrative values (administrative concept), which was
structed a table with all tag corresponding concepts in every removed as of the July 2010 release.
4
Scrutinizing the relationships between SNOMED CT concepts and semantic tags
Some corresponding concepts had minor edits made to Release NE Inact. N|I NYN NYY YYN YYY #Err. %Err. Total
their FSNs, but we do not count this as a change in corre- 20030131 78 0 78 0 0 110 0 0 0 110
sponding concept. Finally, one tag switched its correspond- 20030731 78 0 78 0 0 110 0 0 0 110
ing concepts from one release to the next: the tag ‘finding’ 20040131 75 0 75 0 0 113 0 0 0 113
initially had as its corresponding concept 246188002 | Find- 20040731 75 0 75 0 0 113 0 0 0 113
20050131 75 0 75 0 0 113 0 0 0 113
ing (finding) but this concept was deactivated in the January
20050731 75 0 75 0 0 113 0 0 0 113
2004 release and the ‘finding’ corresponding concept
20060131 75 0 75 0 0 113 0 0 0 113
changed to 404684003 | Clinical finding (finding). 20060731 75 0 75 0 0 113 0 0 0 113
3.2 Mismatched concepts 20070131 73 0 73 0 0 115 0 0 0 115
20070731 66 0 66 18 26 78 0 44 36.07 122
After identifying all mismatched concepts for every ST in 20080131 61 0 61 6 2 119 0 8 6.3 127
every release, we organized counts of mismatched concepts 20080731 60 0 60 1 18 109 0 19 14.84 128
into two tables, one for global mismatching (Table 1) and 20090131 60 0 60 0 0 128 0 0 0 128
one for local mismatching (Table 2), with one row per ST 20090731 60 0 60 0 0 128 0 0 0 128
and one column per release (Table 1, for readability, contains 20100131 60 15 75 0 0 113 0 0 0 113
only the counts for the January versions of each release). The 20100731 59 15 74 0 0 114 0 0 0 114
number of globally mismatched concepts per release is gen- 20110131 59 15 74 0 0 114 0 0 0 114
erally decreasing over time, and has gone from 14,814 (5% 20110731 59 15 74 0 0 114 0 0 0 114
20120131 59 15 74 0 0 114 0 0 0 114
of active concepts) in the January 2003 release to 89
20120731 58 15 73 2 0 113 0 2 1.74 115
(0.027%) in the January 2017 release. The number of global
20130131 56 15 71 4 0 113 0 4 3.42 117
mismatches dropped dramatically from the July 2005 release 20130731 50 15 65 10 0 113 0 10 8.13 123
(14,715 mismatches – 4.83% of active concepts) to the next 20140131 40 15 55 17 9 107 0 26 19.55 133
release in January 2006 (1522 – 0.5%). This improvement is 20140731 35 15 50 22 10 106 0 32 23.19 138
likely attributable to large changes in the hierarchy that in- 20150131 21 15 36 34 10 108 0 44 28.95 152
volved changes in semantic tags for three hierarchies: ‘disor- 20150731 12 16 28 24 0 136 0 24 15 160
der’, ‘event’, and ‘finding’. This reorganization is docu- 20160131 6 16 22 65 9 92 0 74 44.58 166
mented in the SNOMED CT Editorial Guide’s section on 20160731 5 16 21 67 11 89 0 78 46.71 167
Changes and historical notes: ‘In January 2006, a number of 20170131 0 16 16 69 14 89 0 83 48.26 172
concepts from the | Clinical finding | hierarchy were moved Table 3. Locally mismatched ‘disorder’ concepts
to the Event hierarchy’ (IHTSDO, 2015, p294).
In the January 2017 release there are only four tags with mis-
matched concepts for a total of 89 mismatches: 83 are tagged 4 DISCUSSION
‘disorder’, four are tagged ‘regime/therapy’, one is tagged Our hypothesis that SNOMED CT intends its STs to have a
‘product’, and one is tagged ‘substance’. one-to-one correspondence between tags and certain high-
level concepts is supported by:
3.3 Mismatched disorders
(1) the very existence of identifiable tag corresponding
Table 3 provides more detail about the categorization of the
concepts (a single ‘highest’ concept for each tag that
188 locally mismatched ‘disorder’ concepts by release. This
is close to the top concept and that in each case sub-
table was constructed by collecting all concepts that at least
sumes the vast majority of concepts that use the tag);
once in their lifetime were locally mismatched while having
the semantic tag ‘disorder’. (2) the generally very low occurrence of mismatched con-
The colum marked ‘NE’ has for each release counts of the cepts.
number of ‘disorder’ concepts that appear and are mis- Errors, however, remain present and sometimes are even
matched in some later SNOMED release, but that did not yet introduced, interestingly as witnessed by our analysis of the
exist at the release for that row. locally mismatched disorder concepts (Table 2) increasingly
The ‘Inact.’ column counts how many were active concepts more in more recent versions. A particularly illustrative ex-
in an earlier release but were inactive at the row release. ample of this in the January 2017 release is the concept
The column ‘N | I’ is a sum of the previous two columns. 109186003 | Sickle cell test kit (substance) which turned out
The ‘NYN’, ‘NYY’, ‘YYN’, and ‘YYY’ columns count the to be newly mismatched as there were no ‘substance’ con-
number of concepts that fall into each of those categories. In cepts mismatched globally from 2009 until this release, and
the January 2017 release, the 83 mismatched concepts fall not even a locally mismatched ‘substance’ concept ever be-
into two of our categories: ‘NYN’ (69) and ‘NYY’ (14). fore. It is mismatched because it is not subsumed by the ‘sub-
stance’ tag’s corresponding concept 105590001 | Substance
(substance). Indeed, the sickle cell test kit concept is directly
5
K.Takahashi et al.
subsumed by 385387009 | Test kit (physical object), which level concepts. The occurrence of mismatches between the
has 29 other children that all have the words ‘test kit’ in their semantic tags of lower-level concepts and their placement in
FSN and are correctly tagged with ‘physical object‘ (e.g. another hierarchy than where expected according to the se-
1109190001 | Virus test kit (physical object)). mantic tag is a sign that the SNOMED CT authoring tool is
There are a number of ways for a mismatched concept to not equipped with a formal mechanism to keep the hierarchy
appear in a release. These include the addition of a new con- consistent with the semantic tags. It is our recommendation
cept, re-activation of an old concept, and changes in the con- that such mechanism would be implemented and the method
cept’s subsumption hierarchy. In the sickle cell test kit case developed here might be a good starting point in addition to
changes in the hierarchy are responsible: in 2016 and earlier, other mechanisms for quality control that have been devel-
the concept 385387009 | Test kit (physical object) was itself oped by third parties (Geller, Ochs, Perl, & Xu, 2012; Ochs
mismatched, being subsumed by 105590001 | Substance et al., 2015).
(substance) and not by the ‘physical object’ tag’s corre-
sponding concept, 260787004 | Physical object (physical ACKNOWLEDGEMENTS
object). The test kit concept’s children were all tagged ‘sub- This work was supported in part by Clinical and Translational
stance’. Science Award NIH 1 UL1 TR001412-01 from the National
In 2017 the test kit concept was (correctly) moved to the Institutes of Health. The content of this paper is solely the
physical object hierarchy, and it went from being mismatched responsibility of the authors and does not necessarily repre-
to being not mismatched. 29 of its children had their FSNs sent the official views of the NIDCR, the NLM or the Na-
changed to use the tag ‘physical object’. Though this move tional Institutes of Health.
resolved one mismatch, another appeared: the sickle cell test
kit concept became mismatched as a result, as shown in the REFERENCES
two hierarchy excerpts in Fig.4. Most of the child concepts
Bishop, B., Kiryakov, A., Ognyanoff, D., Peikov, I., Tashev, Z., &
of 385387009 | Test kit (physical object) are omitted here in Velkov, R. (2011). OWLIM: A family of scalable semantic
the interest of space, as are child concepts of all the other con- repositories. Semant. web, 2(1), 33-42.
cepts that appear here. Ceusters W. (2009) Applying Evolutionary Terminology Auditing
to the Gene Ontology. Journal of Biomedical Informatics
2016 2017 42:518–529.
138875005 138875005 Ceusters, W., & Bona, J. P. (2016). Analyzing SNOMED CT's
SNOMED CT SNOMED CT Historical Data: Pitfalls and Possibilities. AMIA Annu Symp
Concept Concept
Proc, 2016, 361-370.
105590001 260787004 105590001 260787004 Geller, J., Ochs, C., Perl, Y., & Xu, J. (2012). New abstraction
Substance Physical object Substance Physical object networks and a new visualization tool in support of auditing the
(substance) (physical object) (substance) (physical object)
SNOMED CT content. AMIA Annu Symp Proc, 2012, 237-246.
385387009 385387009 IHTSDO. (2015). International Health Terminology Standards
Test kit Test kit
Development Organization - SNOMED CT® Technical
(physical object) (physical object)
+28 +28
Implementation Guide - January 2015 International Release (US
109186003 109190001 109186003 109190001 English).
Sickle cell test kit Virus test kit Sickle cell test kit Virus test kit
(substance) (substance) (substance) (physical object)
IHTSDO. (2017). SNOMED CT Editorial Guide: 7.3.4. Hierarchy
tag. Retrieved from https://confluence.ihtsdotools.org/display/
DOCEG/7.3.4+Hierarchy+tag
Fig.4. Test kit concept changes 2016 - 2017
Ochs, C., Geller, J., Perl, Y., Chen, Y., Xu, J., Min, H., . . . Wei, Z.
(2015). Scalable quality assurance for large SNOMED CT
5 CONCLUSION hierarchies using subject-based subtaxonomies. J Am Med
Inform Assoc, 22(3), 507-518. doi:10.1136/amiajnl-2014-
We have successfully demonstrated that it is possible to im- 003151
plement an algorithm that maps semantic tags to correspond-
ing SNOMED CT concepts. We applied this mapping in an
analysis of all active concepts across SNOMED CT releases,
assessing the extent to which the tags as used reflect the
placement within the hierarchy of the concepts that use them
both locally, and, in the spirit of Evolutionary Terminology
Auditing (Ceusters 2009) with respect to the last version
which functions as a gold standard. The results support our
hypothesis that SNOMED CT indeed intends its semantic
tags to have a one-to-one correspondence with certain high-
6