=Paper=
{{Paper
|id=Vol-2518/paper-ODLS8
|storemode=property
|title=Aligning an Administrative Procedure
Coding System with SNOMED CT
|pdfUrl=https://ceur-ws.org/Vol-2518/paper-ODLS8.pdf
|volume=Vol-2518
|authors=Stefan Schulz,Johannes Steffel,Peter Polster,Matvey Palchuk,Philipp Daumke
|dblpUrl=https://dblp.org/rec/conf/jowo/0001SPPD19
}}
==Aligning an Administrative Procedure
Coding System with SNOMED CT==
https://ceur-ws.org/Vol-2518/paper-ODLS8.pdf
Aligning an Administrative Procedure
Coding System with SNOMED CT
Stefan SCHULZ a,c,1, Johannes STEFFEL a, Peter POLSTER a,
Matvey PALCHUK b and Philipp DAUMKE a
a
Averbis GmbH, Freiburg, Germany
b
TriNetX, Cambridge, MA, USA
c
Institute for Medical Informatics, Statistics and Documentation,
Medical University of Graz, Austria
Abstract. OPS, the German coding system for therapeutic and diagnostic
procedures, is a large and complex classification system. Its main purpose is to
provide codes for billing. Like other systems of this type (e.g. ICD-10) it follows
the principle of class disjointness and exhaustiveness. SNOMED CT, on the other
hand, aims at providing standardised terms, together with logic-based descriptions,
and pursues the goal to make the electronic health record (EHR) computable and
interoperable across languages and jurisdictions. We investigated the feasibility of
aligning OPS with SNOMED CT, based on the 1000 most frequently used OPS
codes. A team of three terminologists performed the mapping (partially overlapping),
using the first hundred codes to determine guidelines. From the work, which is
currently being extended, we can draw the following conclusions: (i) for less than
half of the OPS codes, a semantically equivalent SNOMED CT code can be found;
(ii) many maps require SNOMED CT post-coordination but remain approximate;
(iii) the mapping work is impaired by imprecise descriptions in either terminology
system.
Keywords. Medical Procedures, Medical Classifications, SNOMED CT
1. Introduction
Most artefacts that provide a semantic reference – generically referred to as terminology
systems [1, 2] – for the organisation of biomedical data are restricted to a well-defined
scope regarding the types of referents these data denote. Examples are drug
terminologies, which offer codes and definitions for pharmaceutical products and
chemicals (e.g., ATC, RxNorm, ChEBI), terminologies for everything that can be
observed and measured (LOINC), bodily conditions like disorders and injuries (ICD-10),
cell components, molecular functions and biological processes (Gene Ontology), just to
name some of the most important ones. There is an increasing momentum towards
international standardisation and cross-border use. It has a long tradition in the case of
ICD for health statistics, promoted by the WHO in 42 language versions, and is more
1
Corresponding Author: Stefan Schulz, Averbis GmbH, Salzstraße 15, 79098 Freiburg im Breisgau;
E-mail: stefan.schulz@averbis.com. Copyright © 2019 for this paper by its authors. Use permitted under
Creative Commons License Attribution 4.0 International (CC BY 4.0).
recent in the case of ongoing standardisation efforts for drug products (IDMP),
responding to a worldwide demand for internationally harmonized medicinal product
specifications. LOINC is being translated into more and more languages. Regarding
terminology support for biomedical research, resources like the Gene Ontology, ChEBI
and other bio-ontologies have been created from the very beginning for semantic
annotations of bio-molecular data across international research communities.
There is one remarkable exception to this trend, viz. terminologies for medical
procedures. Procedure terminologies encompass operations, drug and other therapies, as
well as imaging and other diagnostic procedures. Here, a multitude of national
terminologies coexists. Most of them are better described as catalogues than as
terminologies or ontologies, i.e., flat lists of medical interventions, often completed by
weight values from which the price of each procedure can be computed. Rather than
terminologies or ontologies (which systematise terms and/or their referents), they
resemble classification systems (like ICD) with mutually disjoint classes, supported by
rules that clarify under which conditions a given procedure belongs to a given class.
Billing is therefore a major driver for procedure coding systems, and the
heterogeneity of national health systems (enhanced by different systems co-existing
within jurisdictions) explains the special status of these artefacts. While there are good
examples that semantic harmonisation of data across national borders is possible when
the data is captured by standards that enjoy global adoption, semantic harmonization of
medical procedures remains a problem [3], as they stand out as a large data domain that
does not have a globally adopted standard.
Medical procedures also constitute an important portion of SNOMED CT (58,213
concepts) [4], the “common global language for health terms”, an international standard
that enjoys increasing acceptance around the globe. Best described as an ontology-based
terminology, SNOMED CT aims at providing standardised terms, together with logic-
based descriptions, and pursues the goal to make electronic health records (EHR)
computable and interoperable across languages and jurisdictions. The focus of the
SNOMED procedure hierarchy (about one sixth of all SNOMED CT concepts) is to
provide fine-grained, standardised descriptions of clinical procedures.
Therefore, SNOMED CT is often seen as a strong candidate to represent medical
procedures, but capturing procedures with SNOMED CT has never been current practice.
An important step towards having procedures coded in SNOMED CT, could consist in
the construction of mappings to SNOMED CT from the local procedure coding systems,
as there does not seem to be any other suitable standard. ICD-10-PCS, developed for use
in the U.S. and adopted in few additional countries, is typically considered to be very
difficult to map, ICHI (International Classification of Health Interventions [5]), currently
under development by the WHO is far too coarse-grained, which is also true for the
Procedures section of the Medical Subject Headings (MeSH) primarily targeting
biomedical literature indexing.
2. Background
There are ongoing efforts in some countries to harmonize procedure standards. In the
US, the National Library of Medicine is working on a way to map ICD-10-PCS to
SNOMED. Instead of a map [6] they chose to focus on a tool that requires human
intervention to generate an equivalence mapping from an ICD-10-PCS procedure to
SNOMED. In the UK, the OPCS standard has been mapped to SNOMED CT by the
NHS. However, efforts requiring the mapping of very large (tens of thousands of codes)
and very complex terminologies such as the ones used for procedures (as well as
SNOMED itself) are being undertaken by governmental agencies more so than private
companies or collaborations. Governments tackle these projects because the effort
required is high and because supporting and promoting interoperability is being
understood as a global need.
Compared to other biomedical terminology systems, SNOMED CT is unique not
only due to its scope and size, but also regarding its ontological foundation, based on the
description logics [7] profile OWL-EL [8]. This allows for logically defining its
representational units (SNOMED CT concepts), e.g. in
Tonsillectomy equivalent to Procedure and
has-part some ((method some Excision - action) and
(‘procedure site direct’ some ‘Tonsillar structure (palatine)’))
This is the case of a so-called pre-coordinated concept (identified by the code
173422009 and the label ‘Tonsillectomy (procedure)’). The syntax also allows for
constructing more detailed expressions for which no code is available, e.g. tonsillectomy
with an ultrasonic scalpel.
Procedure and has-part some
((method some Excision - action) and
(‘procedure site direct’ some ‘Tonsillar structure (palatine)’) and
(‘using device’ some ‘Ultrasonic scalpel’))
This mechanism, called post-coordination, allows for an increased level of coverage
and detail, however, at the price of increased complexity, which means that
documentation systems have to deal with logical expressions instead of just codes.
A typical classification system also provides a code for tonsillectomy, but often
comes with additional instructions, such as using different codes for tonsillectomy with
or without adenoidectomy, or demanding additional codes for, e.g. haemostasis after
tonsillectomy, or requiring different codes for different age groups. Table 1 provides a
juxtaposition of a procedure ontology with a procedure classification (see also [9]).
Table 1. Procedure ontologies compared to procedure classifications
Procedure ontology Procedure classification
Semantics Open world, classes (extension of Closed world (disjoint classes)
concepts) often overlap
Structure Multiple hierarchy, intensional Single hierarchy, extensional
Constructors Subclass, Equivalence, conjunction Subclass, excludes
(“and”), existential quantification
(“some”)
Ontological Classes of medical procedures, i.e. Purpose-oriented standardised information
commitment (parts of) actions performed on a patient objects related to medical procedures (see
by a health professional left)
Purpose Provision of representational units Provision of correlates to medical procedures
within a formal account of the electronic and their parts, from which the monetary
health record value of a procedure can be derived
Several scenarios of use justify an alignment between classification-type coding
systems and SNOMED CT:
1. Primary effort put into manual administrative encoding using a procedure
classification or catalogue. Then, the alignment resource can be used to infer
SNOMED codes for adding semantic annotations to the EHR, supporting a
broad range of primary or secondary use scenarios (decision support, prediction,
cohort building, health statistics), capitalizing on the ontological structure of
SNOMED CT.
2. Primary effort put into manually annotating EHR content with SNOMED CT
codes. However, inferring the full meaning of a procedure classification code
would then require representing also the disjointness conditions and exclusions.
For several reasons this exceeds the power of SNOMED CT post-coordination.
As an alternative, each procedure code could be expressed as a query on a
SNOMED-CT annotated record.
3. The same as 2, but using natural language processing (NLP) for annotating
clinical narratives with SNOMED CT codes. This scenario, as well as the
previous one, would realistically require additional human encoding efforts,
given the high quality required for codes that are used for billing.
Our work described in this paper pursues the first of these three goals, i.e. the
direction from the classification to the ontology. Out of a ranked list of approximately
24,000 codes from the German procedure classification system OPS [10] (Versions 2004
- 2019) it takes the most frequent 1000 and attempts to map them to SNOMED codes or
post-coordinated expressions. The main criterion underlying this effort is the following:
given an OPS procedure code pi attached to an EHR: for which SNOMED CT concepts
ci1… cin (or OWL class-like post-coordinated expressions), instance(s) can be assumed to
exist in the health care episode described by that EHR?
3. Materials and Methods
3.1. OPS and SNOMED CT
OPS, used in Germany for encoding therapeutic and diagnostic procedures, is a fine-
grained classification system with 35,641 codes distributed across seven hierarchical
levels. Its purpose is to provide codes for billing. Like other systems of this type (e.g.
ICD-10) it follows the principle of class disjointness and exhaustiveness. The
“Systematic Version” PDF file (2019 release) was used as a reference. In this version,
formatted like a book, certain naming principles had to be considered. E.g., the code 5-
790.26: “Geschlossene Reposition einer Fraktur oder Epiphysenlösung des distalen
Radius mit Osteosynthese unter Verwendung eines intramedullären Drahts” [Closed
reduction of a fracture or slipped epiphysis of the distal radius by internal fixation using
an intramedullary wire]. This label is not pre-synthesised; it has to be constructed by
5-790 Closed reduction of a fracture or slipped epiphysis by internal fixation
**5-790.2 By intramedullary wire
6 ↔ Distal radius
The sixth digit is taken from a list with anatomical sites that can be combined with
several 5-character codes. However, completely pre-synthesised texts are also available.
Nevertheless is the inspection of the Systematic Version indispensable in order to get
access to exclusions, inclusions, and scope notes at all hierarchical levels. E.g., the
subchapter 5-79 Reduction of Fracture and Dislocation is preceded by nearly one page
of such additional information (e.g. excludes therapy of pseudarthrosis, requires separate
encoding of nerve sutures). In addition, under the heading the subsubchapter 5-790, more
additional information is given, e.g. that child fractures are included, closed reductions
dislocations of joints are excluded, or that arthroscopic assistance requires an additional
code. The level of detail is often only fully understandable by specialist surgeons.
Expertise and a certain degree of subjective interpretation are also needed in order to
understand the meaning of certain terms, which lack precise definition, such as
“Epiphyseolyse”.
The SNOMED CT Procedures hierarchy provides formal concept definitions, which
can be interpreted as description logics axioms (cf. Background). In contrast, scope notes
or text definitions are completely missing, which is particularly challenging where
formal definitions refer to undefined primitives, e.g. from the SNOMED CT Qualifier
Value hierarchy, e.g. Preperitoneal approach without connection to any anatomy
reference. Another peculiarity is what SNOMED CT calls role grouping. This can most
straightforwardly be interpreted as asserting a mereological order between procedures
and their processual parts. However, these role groups also occur solely in, e.g. in the
concept Sigmoidoscopy, which is therefore classified as a taxonomic parent of, e.g.
Sigmoidoscopy with biopsy. Therefore, the precise meaning of Sigmoidoscopy would
then be “Procedure with sigmoidoscopy”. The background of this is to optimise term
retrieval: searching for “sigmoidoscopy” would then retrieve also data annotated with
the concept Sigmoidoscopy with biopsy. Negation cannot be expressed by the SNOMED
CT syntax. However, we find “reified” negations in several concepts, such as Computed
tomography of head without contrast.
These examples demonstrate, in addition to the basic distinctions exposed in Tab. 1,
the wide discrepancy of SNOMED CT and OPS in particular (which can also be extended
to classification-like coding systems in general). Whatsoever alignment between OPS
and SNOMED CT has to be aware of this. Simple lexical mapping is not sufficient. It
would lead to numerous wrong equivalence statements: an OPS code, with its meaning
restricted by numerous exclusion rules and with the underlying close-value assumption
is rarely fully semantically equivalent to any SNOMED CT concept.
3.2. Dataset for OPS code ranking
The dataset was provided by TriNetX. They harvested it from two German hospitals,
where OPS is established as the official coding system (together with ICD-10) for the
German DRG (Diagnosis-related groups) payment system [12]. The datasets consists of
6,892,330 single codes out of 23,985 different OPS code types. This corresponds to a
coverage of 67.3%, i.e. about one third of OPS codes were never used. For our mapping
project we selected the 1,000 most frequent codes, which still correspond to 5,580,702
code assignments, i.e. 80.9%.
3.3. Coders, pilot SNOMED mappings and mapping schema
Two coders were recruited (2nd and 3rd authors), both of them final year medical students,
one of which already held a degree in nursing. They were trained supervised by the first
author, MD and experienced terminologist / ontologist. Each coder was hired for 26
hours a month over three months. The supervisor had the same time budget. This period
covered the whole cycle from guideline creation, training, mapping, validation to the
delivery of the map and the final report (this paper).
The 1000-code set OPS1000 was ordered by random. The first author, together with
the coders analysed the first 100 codes regarding their alignment with SNOMED CT, i.e.
the mapping of one OPS code to one or more SNOMED CT codes. The main purpose of
this initial step was to reach a consensus regarding a meaningful, simple and reusable
mapping scheme. First, the underlying assumption of the mapping process was
formulated: According to assumption 3 in the Background chapter, we defined the
mapping task as follows: Given a patient record annotated with the code OPSi , which
SNOMED CT expression(s) SCT1...j can be reliably assumed to be instantiated. The range
of the mapping should contain one or more SNOMED CT codes or SNOMED CT post-
coordinated expressions belonging to one of the semantic types "Procedure",
"Regime/Therapy" or "Situation". In addition, a scoring system is used to distinguish
either the quality of the mapping or “no mapping”.
For each map, the OPS code is analysed in its hierarchical context, taken into
consideration scope notes, inclusion and exclusion statements. Elements of the OPS label
are translated into English, eliminating doubts regarding the appropriate translation in
online sources whenever necessary. In case of doubt, a search with more general terms
is done. Words or word stems are entered into the SNOMED browser [13]. In order to
find the best matching term, also sibling, super and subconcepts are inspected.
As a rule of thumb, maps are preferred that were as close as possible to the original
wording. In case no map is achieved, a compositional approach is pursued to approximate
the meaning. Full post-coordination using the SNOMED CT compositional grammar is
not aimed at, due to its complexity, its experimental status (especially regarding its
closeness to description logics) and its irrelevance for current implementations. Post-
coordination is therefore restricted to logical conjunction (AND), disjunction (OR) and
addition (ADD). The latter is preferred in case the OPS code stands for clearly distinct
entities, of which a conjunction (even given the large tolerance how SNOMED CT
handles logical conjunctions) is considered inappropriate. E.g., if there is no SNOMED
CT concept for an OPS code ‘Procedure P on body site B’, then the post-coordination P
AND ‘Procedure on B’ would be an appropriate representation, because it is still one
procedure. In contrast, if an OPS code stands for ‘Procedure P followed by Procedure
Q’, the preferred SNOMED map would then be P ADD Q, which means that it is
represented by actually two separate procedures.
OPS codes are often defined by numerical values or value ranges such as number of
therapy units, dosage, frequency, or implicitly by age groups (e.g. adults, children).
SNOMED CT procedure concepts never include such criteria, i.e. an exact mapping
cannot be expected in these cases. Table 2 gives an overview of the mapping scores we
elaborated.
Table 2. Scoring of OPS – SNOMED CT mappings
Score Meaning regarding source OPS code (S) and target code or expression (T)
Exact T holds for the same (individual) procedures as S
Exact-Q T holds for the same (individual) procedures as S, when quantitative restrictions on
S are neglected
Broader The individual procedures denoted by S are a (still significant) subset of those
denoted by T
No mapping There is no code or expression T that allows any of the above judgement.
For practical purposes, there is still the preliminary category “revisit”, which is set
in case a coder is not sure about the decision and wants to mark the code for a group
discussion. Each OPS code is seen three times, with the following roles: C: coder - the
person who does the OPS-SNOMED mapping; R1: first reviser - the person who checks
the decision taken by C, R2: second (senior) reviser: the person (mostly the supervisor)
who takes the final mapping decision, sometimes as a result of a group discussion.
Comment fields are available for each of the three experts; whenever a map is changed,
this is documented by an entry. Each expert uses a different colour for his comments.
The pilot phase also yielded the following exclusion recommendations:
1. OPS Codes containing the administration of medicines for which no SNOMED
CT procedure codes exists
2. Codes that contain extremely detailed descriptions of a "complex therapy"
3. Planning phase A of a procedure B if only SNOMED codes for B are present
4. Procedures in cases of doubt
5. Supplementary OPS codes (“Zusatzcodes"), unless containing significant
information of the type procedure
6. Retired OPS codes
The exclusion criteria were reassessed at the end of the mapping phase.
3.4. Mapping process
As a collaborative environment, a Google spreadsheet was created and filled by the
coders with OPS codes, texts, logical operators, and comments. Table 3 provides the
stepwise approach on a randomised list of OPS codes, identified as OPS1 - OPS1000.
Table 3. Steps for mapping the most frequent OPS codes to SNOMED CT
OPS codes Step
OPS1 - OPS100 Collaborative, explorative (C, R1, R2) . Consolidation of the mapping scheme and mapping
guidelines.
OPS101 - OPS300 Students play the role of C and R1 for half of the codes
Thereafter group discussion including R2, adjudication of controversial decisions
OPS301 - OPS400 Performed in separate spreadsheet (and without communication) for first reliability testing. Both
C and R1 play the C role. Thereafter, calculation of inter-coder agreement, then adjudication
between students and with R2 for controversial cases. Adding dataset to main table
OPS401 - OPS800 Like in step 2, students play the role of C and R1 for half of the codes
Thereafter group discussion including R2, adjudication of controversial decisions
OPS1 - OPS800 Reordering of list by order of codes. Comparison of similar codes and related mapping decisions
by all C, R1, R2. Chat and phone discussions in case of inconsistent mappings of similar codes.
Revising and completing R2 decisions
OPS801 - OPS1000 Second reliability testing. Adjudication between students and with R2 for controversial cases.
Reassessment of the exclusions. Mapping of the re-included codes. Decision for all codes
marked as “revisit”
Coders were also asked to skip the mapping of a code whenever this takes more than
ten minutes. These codes were tagged as “revisit”. The revisiting of these codes is
scheduled to take place once all other codes are consolidated.
3.5. Quality assessment of mappings
Inter-coder agreement was measured at two points (Tab. 3): for the codes OPS301 - OPS400,
in order to achieve a preliminary estimation and at the end, for the codes OPS801 - OPS1000.
3.6. Prototypical cases of mapping issues
During the whole process, cases of difficult or controversial mappings were picked out
and discussed. Priority was given to those mapping problems that can be seen as
prototypical issues not only with regard to OPS, but also to classification-like coding
systems in general.
3.7. Final workup of top 1000 map
For a final quality check, the OPS codes were re-arranged from a random order to the
numeric order of OPS. This revealed many inconsistencies regarding the mapping of
similar codes. The mapping guidelines were adjusted in the sense that also
supplementary codes were mapped (as long as this yielded significant clinical meaning).
In addition, retired codes were mapped. As a matter of principle, codes that only
consisted of administration of drugs were not mapped, assuming that in EHRs there are
other, more complete and reliable sources of medication information.
4. Results
4.1. Metrics
The complete time spent amounted to 3 months x 3 experts x 26 hours per expert and
month, totalling 234 hours. This corresponds to an average effort of approx. 4.3 OPS
codes per hour (14 minutes per code). The mapping of the first 100 codes, including
guideline development and documentation required approx. one fourth of the total
time. A descriptive analysis of the mappings of the 1000 OPS codes is provided by
Table 4. Each code was seen which seen by the three experts and revisited in their
original order by at least one expert.
Table 4. Descriptive analysis of mappings
Cardinality of map (SNOMED CT codes per OPS codes)
0 1 2 3 4
48 617 282 42 11
Quality of mapping
Broader Exact Exact-Q No mapping Revisit
610 310 32 48 0
Type of logical combination
None AND ADD OR Complex
665 178 79 56 22
The results of inter-coder agreement are provided in Table 5. We compute simple
percentage agreements, because agreement by chance is negligible.
Table 5. Inter-coder agreement in percent (100 mappings evaluated: 301-400; 200 mappings: 801-1000)
Agreement [95% CI]
Type of agreement
OPS301 - OPS400 OPS801 - OPS1000
Coders agree on at least one core SNOMED CT concept per OPS 68% [58%; 76%] 65% [58%; 71% ]
code
Coders agree on the same set of SNOMED CT concepts per OPS 54% [44%; 63%] 46% [38%; 52% ]
code
Coders agree on the same set of SNOMED CT concepts per OPS 41% [31%; 50%] 36% [30%; 43%]
code and agree regarding the mapping quality
4.2. Typical cases
In Table 6, typical mapping phenomena are presented. All examples are instances of
frequently recurring phenomena. The OPS labels, which are only available in German,
were translated to English for better understanding.
Table 6. Instances of recurring mapping phenomena. Left column: OPS codes, central columns: mapped
SNOMED CT codes, right column: logical connection between SNOMED CT codes
1 Procedure with finding: finding is not represented in the map (would require complex post-
coordination)
1-265.4 Electrophysiological 175131000 Percutaneous transluminal
examination of the heart, electrophysiological studies on
catheter-assisted: In conducting system of heart
tachycardia with narrow QRS
complex or atrial tachycardia
2 Procedure with device: device is not represented in the map (would require complex post-coordination)
1-266.1 Electrophysiological 252425004 Cardiac electrophysiology
examination of the heart, not
catheter-assisted: implanted
cardioverter defibrillator
(ICD)
3 Procedure with body part: body part is not represented in the map (would require complex post-
coordination)
1-268.3 Cardiac Mapping: 21032000 Cardiac mapping
Right Ventricle
4 Part of the procedure requires separate coding
1-430.1 Endoscopic biopsy of 312849006 Biopsy of 10847001 ADD
respiratory organs: bronchus bronchus Bronchoscopy
312849006
5 Logical conjunction of specific procedure and anatomy-related procedure
1-490.6 Biopsy without 287538006 Non- 118714000 AND
incision on skin and surgical skin biopsy Procedure on
subcutaneous tissue: lower leg lower leg
6 Different granularity in SNOMED requires post-coordination in one case but not in another
3-825 Magnetic resonance 432369004 Magnetic resonance
imaging of the abdomen with imaging of abdomen with contrast
contrast
3-826 Magnetic resonance 58713006 Magnetic 51619007 Magnetic resonance AND
imaging of the musculoskeletal resonance imaging of imaging with contrast
system with contrast agent musculoskeletal
structures
7 Coordination needed to add the feature that a procedure is a diagnostic one:
1-631 Diagnostic 392153002 103693007 AND
Esophagogastroscopy Esophagogastroscopy Diagnostic
procedure
8 Missing of aggregations at the level “vessel” (regardless of whether artery or vein)
3-611.x Phlebography of 4008007 60006002 OR
cervical and thoracic vessels: Phlebography of neck Intrathoracic
Other phlebography
9 Exclusion statements for OPS codes (cannot be expressed by SNOMED CT semantics)
1-207.2 Video-EEG (10/20 252738008 Video
Electrodes). Excl.: Video-EEG electroencephalogram
during pre and intraoperative
epilepsy assessment
10 Residual class “other”, i.e. logical complement (cannot be expressed by SNOMED CT semantics)
1-273.x Right heart 40403005 Catheterization of right heart
catheterization: Other
11 Explicit definition “without” (cannot be expressed by SNOMED CT semantics)
1-275.0 Transarterial Left 33367005 Coronary 67629009 AND
Heart Catheter Examination: angiography Catheterization
Coronary angiography without of left heart
further action
12 Distinction between logical conjunction “AND” and addition (more than one instance in the target
representation (“ADD”))
1-650.2 Diagnostic 174184006 235150006 265387003 (X AND
Colonoscopy: Total, with Diagnostic endo- Total Diagnostic endo- Y) ADD Z
Ileoscopy scopic examination colonoscopy scopic examination of
on colon ileum
5. Discussion
Given the size of the two terminologies and the fact that the mapping was done only with
the most frequent 1000 codes (i.e. the most frequent medical procedures, covering 80.9%
of the procedure coding results used in German university hospitals), the relatively low
amount of exact mappings and the frequent need of post-coordination of two or more
SNOMED CT concepts may be surprising. However, knowing the large structural
differences between these two coding systems and their distinct scenarios of use explains
the differences. For instance, many frequent OPS codes contain numeric criteria (number
of treatment sessions, duration of interventions, number of sites where a complex
intervention takes place, dosage of drugs). This is mainly because complexity and the
treatment costs grow with these numeric values. By including them in the definition of
codes, the use of OPS as a tool for billing becomes more convenient, since multiple
assignments of the same code, e.g. for each single application of a drug, are not
necessary. A similar case is the use of single codes for complex treatments, e.g. stroke.
Here the assignment of one single code depends on fine-grained rules (for stroke,
comprising 592 words), including the frequency of monitoring, the required diagnostic
measures and the specialty of the clinicians involved. In rheumatology, complex
therapies with integrated function-oriented and pain-therapeutic treatment sections, often
lasting a week or more, are prerequisite for an efficient acute care of chronically ill
patients. The combination of a multitude of “small” clinical procedures requires
appropriate codes that represent the overall effort without coding each single procedure
[14]. There are also cases, in which the exact OPS code depends on the computation of
a score that estimates the overall effort spent in complex treatments (32 among 1000 OPS
codes).
On the other hand, many of the simple procedures, which are mentioned as
constituents of complex therapies, like blood pressure measurement or blood sampling,
as well as most lab procedures are missing in OPS. The reason is simple: the effort
needed for these actions, in isolation, is just too insignificant. As much a single procedure
may be relevant for clinical documentation, if it is cheap, there is no OPS code.
For SNOMED CT, such an overloading of procedure concepts would contradict its
main destination as a standard for fine-grained clinical documentation, where
reimbursement is not the focus. Although concepts for combined procedures exist, the
focus is on encoding every single procedure.
This explains why our map required so many SNOMED CT co-ordinations, even
for apparently simple concepts, and why the mapping could often not be considered
exact, given the exclusion rules that assure the non-overlapping of OPS classes.
The ontological structure of SNOMED CT allows for complex logical expressions
for concept refinement by its so-called post-coordination mechanism. Because such
expression cannot be processed by any routine implementation, we decided to restrict
ourselves to simple post-coordination patterns. This explains inexact mappings, e.g.
when an indication, a body part of a device was missing for a perfect semantic match
between an OPS code and a SNOMED CT expression (examples 1 – 3 in Table 6). In
other cases, the meaning of an OPS code could be represented by the logical combination
of two or three SNOMED CT classes (example 4 - 7). The same meaning could also be
achieved by using the SNOMED compositional syntax. The semantic equivalence
between such different syntactic forms could be ascertained by a description logics
reasoning engine like SNOROCKET [15]. Example 6 shows variation in the degrees of
pre-coordination: “Magnetic resonance imaging of the abdomen with contrast” maps to
one SNOMED CT concept, whereas “Magnetic resonance imaging of the
musculoskeletal system with contrast” requires post-coordination.
Pre-coordinated SNOMED CT concepts often skip certain anatomical hierarchy
levels, e.g. they require a distinction between arteries and veins, whether OPS often just
refers to “vessels” (Example 8). In other cases, the anatomic delineation of an OPS code
(e.g. knee + thigh for certain procedures on skin) has no correlate in SNOMED CT. Post-
coordination by simple conjunction is here not possible; the way out would be a
disjunctive expression (‘procedure X on skin of knee’ or ‘procedure X on skin of thigh’),
which is, however, not supported by SNOMED CT logics.
A tricky issue is the distinction between one single procedure and a set of
procedures. This was the reason for the “ADD” operation, which in contrast to “AND”
just lumps codes together. Given the special semantics of SNOMED CT procedures (see
comment to “role groups”), one could argue that even complex procedures could be
expressed as logical conjunctions. This might be an argument in favour of substituting
all “ADD” statements by “AND” statements in order to simplify implementations.
The limits of SNOMED CT’s post-coordination power are also reached when it
comes to negation, such as in explicit exclusion rules, residual classes (“others”) and
implicit negations in labels including “without” (examples 9 - 11). This is explained by
the OWL EL profile used for SNOMED CT, which lacks negation.
Finally, a complicating factor was the language gap (German vs. English) and the
lack of clear term definitions. The coders, who were not specialists in any surgical or
diagnostic discipline, depended on medical textbooks and online references, in order to
make clear whether a German term meant the same as its supposed English translation.
In many cases, this was difficult.
In the light of all these factors, the relatively low inter-coder agreement rates were
not surprising. Especially the distinction between “exact” mapping and “broader” was
not easy. That the inter-coder agreement in the last phase was even lower than in the
early phase (see Tables 3, 5) is explainable by the fact that in the early phase many
complex mappings had been left out and encoded as “revisit”.
Our results reinforce the large difference between medical coding systems even
between those that cover the same domain. The dependency of coded information on the
specifics of the vocabulary used and the purpose of the codes cannot be emphasized
enough. The problem of re-use of administrative codes for other scenarios has been
repeatedly addressed [16]. So far, idiosyncratic procedure codes are normally the only
source of procedure information, beyond EHR narratives. Mappings to an international
standard as SNOMED CT is, in theory, a partial solution to this problem; but to be of
high quality it will require leveraging its complete post-coordination mechanism,
supported by description logics reasoning, in order to obtain a high coverage and
convincing retrieval results. Another way to achieve interoperable EHR information
would be the application of natural language processing technology to clinical narratives
in which medical procedures are referenced. The quality and comprehensiveness of such
data depend, however, on natural language resources like lexicons and annotated corpora
for training. Both, however, still constitute a major bottleneck, given the dynamics of
medical language on the hand and the difficulties to share clinical real-world data due to
privacy issues, on the other hand.
Once the content of EHRs is comprehensively represented by standardised
information models and terminologies, a new mapping challenge will arise, viz. inferring
administrative codes like OPS from SNOMED CT codes in context. This could put a
new task on the agenda, viz. the construction of mappings in the inverse direction. This
has already been discussed in the context of ICD-11; a suggested formalism was to
express classification codes as queries on SNOMED CT coded EHR data [17]. Thus, a
new generation of medical classifications could be rooted in and maintained by using
international EHR standards like SNOMED CT.
6. Conclusions and Outlook
The exchange of clinical real-world data is a major desideratum; however, no universal
coding standards are currently used for medical procedures. We report on mapping codes
from OPS, the German coding system for therapeutic and diagnostic procedures, to
SNOMED CT concepts, under the hypothesis that this might be a route towards
worldwide interoperability of clinical data. A team of three terminologists has mapped
the 1000 most frequently used OPS codes to SNOMED CT. After analysing a pilot set
of 100 codes, mapping guidelines were derived. An intermediate analysis of the
mappings showed that about one third of OPS codes could precisely mapped to one
SNOMED CT code or a conjunction of up to three codes. A higher degree of precise
maps would require sophisticated post-coordination but even in these cases, maps are
still approximate. About 5% of the codes could not be mapped mostly due to complex
rules for codes that aggregate many elements of complex therapies or codes for
medication administration. The mapping result is affected by several factors such as lack
of precise definitions in either terminology system, translation problems, or the need of
fine-grained specialist knowledge in some areas. This is also one reason behind the rather
low inter-coder agreement.
The team is currently extending the mapping, covering the most frequent 2,125 OPS
codes, which correspond to 90% of the encodings in the underlying clinical dataset. This
map will be published in early 2020 by TriNetX.
References
[1] S. Schulz, P. Daumke, M. Romacker, P. López-García. Representing oncology in datasets: Standard or
custom biomedical terminology? Informatics in Medicine Unlocked 15 (2019), 100186.
[2] S. Schulz, J. Ingenerf, S. Thun, P. Daumke. German-Language Content in Biomedical Vocabularies. In
CLEF (Working Notes), Valencia, Spain, 2013.
[3] C. Maier et al. Towards Implementation of OMOP in a German University Hospital Consortium. Appl Clin
Inform. 9 (2018), 54–61.
[4] SNOMED International. SNOMED CT - The global language of Health, 2019, www.snomed.org .
[5] A. Zaiss, H.P. Dauben. Prozedurenklassifikation im Spagat zwischen Statistik und Abrechnung. [ICHI-
International Classification of Health Interventions : A balancing act between the demands of statistics
and reimbursement]. Bundesgesundheitsblatt 61 (2018), 778–786.
[6] MAGPIE – Map Assisted Generation of Procedure and Intervention Encoding. National Library of
Medicine, 2019, https://magpie.nlm.nih.gov .
[7] F. Baader, D. Calvanese, D.L. McGuinness, D. Nardi, P.F. Patel-Schneider. The Description Logic
Handbook. Theory, Implementation and Applications. 2nd edition. Cambridge University Press. Online
publication July 2010.
[8] B. Motik, B. Cuenca Grau, I. Horrocks, Z. Wu, C. Lutz. OWL 2 Web Ontology Language Profiles (2nd
Edition) W3C Recommendation 11 December 2012. https://www.w3.org/TR/owl2-profiles .
[9] J. Ingenerf, W. Giere. Concept-oriented standardization and statistics-oriented classification: continuing
the classification versus nomenclature controversy. Methods Inf Med 37 (1998), 527–539.
[10] Deutsches Institut für Medizinische Dokumentation und Information (DIMDI). Operationen- und
Prozedurenschlüssel. Version 2019. https://www.dimdi.de/static/de/klassifikationen/ops/kode-
suche/opshtml2019 .
[11] S. Schulz, S. Hanser, U. Hahn, J. Rogers. The semantics of procedures and diseases in SNOMED CT.
Methods Inf Med. 45 (2006), 354–358.
[12] S. Müller-Bergfort, J. Fritze. Diagnose- und Prozedurendaten im deutschen DRG-System [Extent and
use of administrative hospital data in the German DRG system]. Bundesgesundheitsblatt 50 (2007),
1047–1054.
[13] SNOMED CT Browser. https://browser.ihtsdotools.org/.
[14] H.J. Lakomek et al. Die multimodale rheumatologische Komplexbehandlung (OPS 8-983). [The
multimodal rheumatologic complex treatment (OPS 8-983)--challenges, solutions and perspectives]. Z
Rheumatol. 64 (2005), 557–563.
[15] M.J. Lawley, C. Bousquet. Fast classification in Protégé: Snorocket as an OWL 2 EL reasoner. Proc. 6th
Australasian Ontology Workshop (IAOA’10). Conferences in Research and Practice in Information
Technology 122 (2010).
[16] W.R. Hersh et al. Caveats for the use of operational electronic health record data in comparative
effectiveness research. Medical care 51(8 Suppl 3) (2013), S30-7.
[17] S. Schulz. J.M. Rodrigues, A. Rector, C.G. Chute. Interface Terminologies, Reference Terminologies
and Aggregation Terminologies: A Strategy for Better Integration. Stud Health Technol Inform. 245
(2017), 940–944.