- SNOMED CT is a large and complex medical terminology. Thousands of editing operations are applied to its content for each new release. Understanding what changed in a release is important for the end user and SNOMED CT editors. Each SNOMED CT release comes with release notes that provide a brief description of the changes that occurred and a set of “delta” files that identify individual changes in the content. The release notes are brief and changes to thousands of concepts may be described in a few sentences, whereas the delta files contain tens of thousands of individual changes. To better identify how SNOMED CT content changes between releases we introduce a methodology of creating a descriptive delta that captures the editing operations that were applied to SNOMED CT content in a given release in a more comprehensible form. We use this methodology to analyze editing operations that were part of a recent remodeling effort of the Congenital disease and Infectious disease subhierarchies in the large Clinical finding hierarchy.
SNOMED CT [
From the perspective of an end user, it is important to know how (and where) SNOMED CT content is changing. Applications that utilize SNOMED CT will be affected by the changes made between each release. The IHTSDO provides release notes with each SNOMED CT release that briefly describe the major changes that occurred. For example, the Jan 2016 release notes mention that “A project to improve the consistency and quality of the Infectious disease hierarchy has begun with the remodeling of over 2,000 of the estimated 6,400 concepts that are subtypes of Infectious disease.” However, typically, no details are provided about which concepts were modified and how they were modified.
In this paper we describe the creation of a human readable descriptive delta, based on SNOMED CT delta files, to provide a list of editing operations that were applied to individual concepts. A descriptive delta provides the semantics for the changes listed in the delta files, better fitting the needs of the initiators of the change (i.e., SNOMED CT editors) and the observers of the change (i.e., the end users of SNOMED CT). The descriptive delta methodology can be applied to the entirety of SNOMED CT, a specific subhierarchy of interest, or an individual concept. We use this methodology to investigate the editing operations that were applied to the Clinical finding hierarchy throughout 2015 (i.e., in the July 2015 and Jan 2016 versions of the hierarchy). We identify and characterize significant changes in the Infectious disease and Congenital disease subhierarchies that resulted from a concentrated effort to remodel these hierarchies due to errors and inconsistencies in their content.
SNOMED CT [
SNOMED CT is released in a format called RF2 (for
details and a review of this format see [
SNOMED CT is based on description logics (specifically, the EL subset of DL which allows only existential qualifiers). The stated version of SNOMED CT includes all of the relationships that were defined by SNOMED CT editors. The inferred version of the terminology is obtained by applying a classifier on the stated relationships. Each RF2 release includes the stated relationships and the relationships inferred by the classifier.
A concept may be primitive or fully defined. A concept is primitive if its defining characteristics are not sufficient to uniquely distinguish its meaning from other similar concepts. A concept is fully defined if its defining characteristics are sufficient to distinguish its meaning from other similar concepts. If a concept is fully defined, the SNOMED CT classifier can add inferred hierarchical relationships according to the logical definitions of the concepts.
A concept may be modeled using several attribute
relationships of the same type. For example, in Fig 2a
Congenital cyst of iris has three Associated morphology and
Finding site attribute relationships. To clarify the meaning of a
group of attribute relationships, SNOMED CT provides a
mechanism of relationship groups [
Ceusters et al. [
A. Congenital Disease and Infectious Disease Remodeling
As noted in SNOMED CT July 2015 and Jan 2016 release notes, the Congenital disease and Infectious disease subhierarchies are being remodeled to address inconsistencies and errors in their content. Both subhierarchies each contain over 6,000 concepts and are in the large Clinical finding hierarchy, which has 103,912 concepts in the Jan 2016 release. While this editing effort was initiated by the IHTSDO, change requests may be submitted by individual users or SNOMED CT national release centers.
Many SNOMED CT concepts follow the closest proximal
primitive parent stated concept model [
For example, in Fig 1a, Congenital cyst of iris has two stated parents, Congenital anomaly of iris and Cyst of iris, both fully defined concepts. Thus, this concept does not follow the proximal primitive parent. In Fig 1c, which shows the stated modeling of Congenital cyst of iris in the July 2015 release, the concept has just one stated parent, Disease, the proximal primitive parent as of the Jan 2016 release.
A second issue is inconsistent grouping of attribute relationships. In the proximal primitive parent model, attribute relationships must be properly assigned to relationship groups. For example, in Fig 1a Congenital cyst of iris’s Occurrence attribute relationship was not assigned to a relationship group. In the new stated modeling of the attribute relationships are all assigned to the same relationship group.
As of the Jan 2015 release, the Congenital disease hierarchy had 4,566 concepts with a stated fully defined parent and 2,466 concepts with a stated attribute relationship that wasn’t in a relationship group. Similarly, in the Infectious disease hierarchy 1,065 concepts had a stated fully defined parent and 4,198 had a stated attribute relationship in no relationship group.
By modeling concepts using the closest proximal primitive parent design pattern, the SNOMED CT classifier can more accurately infer parents and descendants and remove the incorrect inferences. Furthermore, SNOMED CT editors will only need to assign the proper proximal primitive parent. There will be no need to try and determine where the concept should be in the inferred hierarchy. The other advantage of this editing approach is that it has allowed a substantial number of primitive concepts to become fully defined. B. Descriptive Delta
We will now describe a methodology to extract a human readable descriptive list of editing operations that were applied between two SNOMED CT releases. We call this list a “descriptive delta.” This process utilizes the delta files provided with each SNOMED CT release. Our focus is on capturing the editing operations that were applied to the relationships used in concept definitions.
A SNOMED CT relationship can be represented as a (source concept, relationship type, destination concept, relationship group) tuple. For a relationship this can be abbreviated as ( , , , ). The relationship delta files (for stated relationships and inferred relationships) also indicate if the given relationship is active (i.e., added) or inactive (i.e., removed) in the current release (see Fig 1b, Fig 2b).
Given the set of active and inactive relationships, as expressed in a relationship delta file, and the inferred concept hierarchy for the current release, the editing operations that were applied to the active concepts can be derived. Below we identify five important examples of editing operations and how they are identified using the delta file.
Added relationship: Given an activated relationship 1 there exists no retired relationship 2 where s1 = s2, t1 = t2, and d1 = d2. d1 is also not an ancestor or descendant of d2 in the current release’s inferred concept hierarchy.
Removed relationship: Given a retired relationship 1 there exists no active relationship 2 where s1 = s2, t1 = t2, and d1 = d2. d1 is also not an ancestor or descendant of d2 in the current release inferred concept hierarchy.
Relationship Target More Refined: Given an inactive relationship 1 there exists an active relationship 2 where s1 = s2, t1 = t2, and d2 is a descendant of d1 in the current release’s inferred concept hierarchy.
Relationship Target Less Refined: Given an inactive relationship 1 there exists an activated relationship 2 where s1 = s2, t1 = t2, and d2 is an ancestor of d1 in the current release’s inferred concept hierarchy.
Relationship Group Changed: Given an active relationship 1 there exists an inactive relationship 2 where s1 = s2, t1 = t2, and d1 = d2 but g1 does not equal to g2.
The process of converting a delta file into a descriptive delta may be a one-to-one process, where one delta entry is captured as one editing operation (e.g., for an added relationships), or a many-to-one process (e.g., for relationship target more refined).
The first four editing operations can be further refined by distinguishing between changes to Is a relationships and changes to attribute relationships (e.g., Added parent versus Added attribute relationship and More refined parent versus Attribute relationship target more refined). This distinction provides a separation between changes in the hierarchy and changes in the relationships used to define each concept.
We note that it is possible to capture additional kinds of editing operations, and further refinements of the above editing operations, when necessary. If a concept has only one parent, and there exists both an inactive Is a relationship for that concept and an active Is a relationship for that concept, and the new parent is not an ancestor or descendant of the old parent in the inferred hierarchy, this could be identified as a changed parent editing operation. For example, the concept Midline sinus of the upper lip had one parent, Lip pits, in the Jan 2015 release. This concept was deactivated in July 2015 and Midline sinus of the upper lip parent was changed to Congenital lip pits. However, in this study we focus on the core editing operations listed above.
Two or more editing operations may relate to the same modeling change. For example, in Fig 1, the two stated parents of Congenital cyst of iris, Congenital anomaly of iris and Cyst of iris, were replaced by a single stated parent, Disease (the proximal primitive parent), an ancestor of both of the inactivated parents (only the Is a relationships were inactivated in this example, not the parent concepts themselves). Furthermore, a given concept may have multiple instances of a given kind of editing operation. For example, in its stated modeling, Congenital cyst of iris has two instances of stated attribute relationships being added.
Applying this process on a stated relationships delta file identifies the changes that were intentionally applied by a SNOMED CT editor. Consider the changes listed in the stated relationships delta file for Congenital cyst of iris, as illustrated in Fig 1c. In the descriptive delta, these are identified as: Less refined parent (2): Congenital anomaly of iris and Cyst of iris were replaced by Disease.
Added attribute relationships (2): Associated morphology with a target of Embryonic cyst and Occurrence with a target of Congenital.
Removed attribute relationship (1): Associated morphology with a target of Congenital anomaly
Similarly, the descriptive delta process can be applied on the inferred relationships delta file. This will highlight the implicit changes that occurred between the inferred versions of the concepts. Consider the changes listed in the July 2015 inferred relationships delta file for Congenital cyst of iris, as shown in Fig 1d. These implicit changes are captured as: More refined attribute relationship target (2): Associated morphology with a target of Developmental anomaly and Associated morphology with a target of Cyst replaced by Associated morphology with a target of Embryonic cyst. Removed attribute relationship (3): Associated morphology with target Congenital anomaly and two Finding site relationships with target Iris structure.
Relationship group changed: Finding site with target Iris structure moved to relationship group 4.
The descriptive delta for a stated relationships file and the descriptive delta for an inferred relationships file will often be very different, as the classifier will make significant changes based on the stated relationships. A concept that had no stated editing operations applied may have several implicit changes due to editing operations at its ancestor(s). Given a concept, one can use the inferred concept hierarchy to identify which ancestors were modified and which editing operations were applied at those ancestors. Alternatively, a concept may have stated editing operations applied but this does not affect the concepts in the inferred version of SNOMED CT, and thus, it would have no implicit changes in the descriptive delta created from the inferred delta file.
There may be tens of thousands of editing operations identified. To reduce the amount of information that has to be reviewed, a user can focus on only a subset of the content that was edited. For example, a user can choose a concept and look at only the editing operations that affected the subhierarchy rooted at that chosen concept. Alternatively, they can obtain a summary of the descriptive delta by only looking at the number of editing operations applied. For example, the stated descriptive delta for Congenital cyst of iris could be summarized as “Less refined parent (2), Added attribute relationship (2), Removed attribute relationship (1).”
If a user is interested in identifying which portions of a SNOMED CT hierarchy underwent significant change they could review the number of concepts that were affected by stated editing operations in certain subhierarchies (e.g., a user can see that relatively many editing operations occurred in the Congenital disease and Infectious disease subhierarchies in July 2015, see Results).
To create descriptive deltas we have developed a software tool that takes as input a relationship delta file and a concept of interest c. The output identifies all of the editing operations that affected the subhierarchy rooted at c.
In the initial version of this tool the output is presented tab-delimited text, where one line consists of the concept identifier, the name of the concept, the kind of editing operation applied, and a textual description of the change (e.g., “Parent Cyst of iris was replaced by Disease”). This output can be inserted into an Excel spreadsheet to obtain aggregated information. In Discussion we mention a future version of this tool which will provide a visual change browser.
We obtained the July 2015 and Jan 2016 International SNOMED CT releases in RF2 format and created descriptive deltas using the stated relationships delta file and inferred relationships delta file provided in each release. We focused our analysis on concepts that changed, as opposed to concepts that were added and removed, since we concentrated our attention to the remodeling efforts.
First, we investigated the editing operations that were applied to the concepts in the complete Clinical finding hierarchy. The July 2015 and Jan 2016 stated relationships delta files contained 7,737 and 10,190 entries for the Clinical finding hierarchy, respectively. The inferred relationships delta files contained 33,146 and 39,380 entries, respectively. We identified 5,728 and 7,524 descriptive delta editing operations in the July 2015 and Jan 2016 release, respectively. We identify which subhierarchies (rooted at the children of Clinical finding and Disease) had the most edited concepts.
Disease Finding by site Clinical history and observation findings
In both releases the majority of edited concepts were in the Disease and Finding by site subhierarchies (Table 1). This is not surprising, as these subhierarchies both contain over 60,000 concepts and both contain many of the same concepts (i.e., there are many diseases that can be found in many body sites). The remaining subhierarchies had few changes. The effect of the remodeling of Congenital disease and Infectious disease is reflected in Table 1.
About 1,500 of the 2,586 edited Disease concepts (over 58%) were in the Congenital disease and Infectious disease subhierarchies (these are mostly disjoint; they share only 75 concepts). On the other hand, the Disorder by body site and Inflammatory disease subhierarchies contain relatively many concepts that are also in either the Infectious disease or Congenital disease subhierarchies.
Table 2 provides metrics from the descriptive deltas for July 2015 and Jan 2016. It lists the number of concepts that had a particular type of editing operation applied. We found that an edited concept typically undergoes multiple editing operations. Each edited concept had, on average, 3-4 editing operations applied. The most common editing operations applied in these hierarchies was making a parent more general (i.e., assigning the proximal primitive parent), adding attribute relationships, and assigning relationship groups to existing stated attribute relationships. These editing operations are inline with the goals of the remodeling effort being undertaken in these hierarchies. Table 3 provides three examples of the stated modeling of concept changing.
Identifying changes between two versions of SNOMED CT is important for end users and for SNOMED CT editors. The descriptive delta described in this paper provides a human readable, actionable format that is not currently available. As analogy, the delta files provided in SNOMED CT release are like machine code and the descriptive delta is like a higher level programming language.
Sometimes a user realizes certain changes occurred in a SNOMED CT subhierarchy that they are interested in but they do not know the nature and motivation for the changes. The descriptive delta, with its semantic view of change, can better support such a user in finding out the nature of the changes.
The descriptive delta analysis reflects the extensive work that went into remodeling portions of the Clinical finding hierarchy in July 2015 and Jan 2016. Namely, thousands of editing operations were applied manually to the concepts in the Congenital disease and Infectious disease subhierarchies. These changes affected thousands of concepts directly and thousands more implicitly (Table 2). Whenever this much change occurs one needs to assess the overall impact (and correctness) of the changes. During this study we identify concepts that underwent significant changes (e.g., Congenital heart block) but still were not yet completely remodeled.
From Tables 1 and 2 we notice that concerted editing efforts have a detectable impact on SNOMED CT’s content. For example, the Congenital disease subhierarchy underwent extensive remodeling in July 2015 but almost no remodeling in Jan 2016. In comparison, the Infectious disease subhierarchy saw significant remodeling in both releases. The Congenital disease remodeling project is still in progress, but the focus of the Jan 2016 editing cycle was on Infectious diseases. One limitation is the amount of resources (in terms of editor’s time) that can be dedicated to these large remodeling projects.
From Table 2 two we observe that the number of primitive concepts in Infectious disease has decreased significantly. This is the desired result of the remodeling effort. In the Congenital disease subhierarchy, however, there was no significant reduction. Many concepts that were remodeled could not be fully defined due to limitations in the concept model (e.g., lack of finding site and associated morphology target values of sufficient specificity, SNOMED CT does not currently contain concepts specific enough to be proper targets). From Table 2 we also see that thousands of concepts were affected by inferred changes. By removing the stated fully defined parents from concepts in these subhierarchies, and remodeling the concepts using the proximal primitive parent, the classifier significantly affected the contents.
In future work we will continue to observe the effects of the remodel effort being applied to the Congenital disease and Infectious disease subhierarchies. We will perform a review of sets of concepts that underwent remodeling to evaluate the correctness of the changes. We will compare the results of this review with the output of the descriptive delta to determine if concepts that undergo certain editing operations are more prone to needing additional remodeling.
While our current descriptive deltas are output in a
tabdelimited file format, we plan to provide them in an
interactive application-based display that will enable end users
to easily investigate and search the changes based on the type
of change, areas of interest, specific concepts, and changes in
the inferred version resulting from stated changes. We will
also investigate combining the descriptive delta technique
with diff partial-area taxonomies [
In this paper we introduced a method of creating descriptive deltas for SNOMED CT to identify the editing operations that were applied to its concepts. The descriptive delta helps to identify the nature of the remodeling the large amount of change in the Clinical finding hierarchy that occurred due to the still-in-progress remodeling of the Congenital disease and Infectious disease subhierarchies.
Research reported in this publication was supported by the National Cancer Institute of the National Institutes of Health under Award Number R01CA190779. The content is solely the responsibility of the authors and does not necessarily represent the views of the National Institutes of Health.