This paper describes a methodology for encoding problem lists used in general practice with SNOMED CT. Our intent is to help general practitioners to incorporate SNOMED CT into their existing Electronic Medical Record (EMR) systems with minimal disruption as a first step, thus allowing them to assess its impact prior to full-scale conversion. We started with 1,713 original unique terms that made up the problem lists from the general practice EMR used in the study. We ended with 1,468 unique concepts after two cycles of matching and revisions that led to 1,347 or ~92% successful matches. The remaining terms were revised to tease out modifiers or secondary concepts that could be used to provide equivalency through post-coordination. While skeptics of reference terminology systems often balk at their unwieldy size and complexity for local adoption, this study has demonstrated that, using our methodology, it is possible to create a manageable subset of SNOMED concepts for problem lists used in general practice with immediate tangible value.
The problem list is the keystone of the medical record. In general practice settings, the type of problems presented by patients can be quite diverse. Examples range from non-specific symptoms such as headaches with unknown cause, to a diagnosis of coronary disease that can be expressed in different ways such as heart attack and myocardial infarction. The choice of terms used in problem lists becomes an important design issue for the electronic medical record (EMR), since the level of granularity selected for defining the problems and the actual terms entered into the system can affect one’s ability to retrieve the information afterwards, thus impacting the overall quality of the EMR system.
There have been many studies on the design and use
of controlled terminology to encode the problem lists
in EMR systems and their impact on practice [
For this study, we included all the problem list (PL) terms from the commercial EMR system used by a local general practice in British Columbia, Canada. This setting is typical of many general practices across the country, which are made up of small groups of general practitioners working in a private medical office, mostly on a fee-for-service basis. The medical office in this study has four general practitioners who have worked as a group for 30 years in a township with a population of 100,000 located east of Vancouver, British Columbia. The practice has had 8 years of experience using an EMR. At least two of the practitioners record all of the information on their patients on a daily basis at the time of encounter or shortly thereafter. Laboratory and imaging results and consult reports from external sources – both electronic and on paper – are entered into the EMR either by the practitioners themselves or the medical office assistant.
We applied four matching algorithms used in an
earlier SNOMED CT to ICD-10 mapping project to
find matching SNOMED concepts for each of the PL
terms [
Replaced-By to find current concepts 5. Unmatched Assigned when no match is found
Table 1. Matching algorithms used in this study
In addition to applying the matching algorithms to
the original PL terms, we reran the algorithms after
we normalized the PL and SNOMED terms to
remove “noise” using the Unified Medical Language
System (UMLS 2007 version) normalization steps,
shown in Table 2a [
No 1 2 3 4 5 6
Step Example Remove Hodgkin’s disease, NOS o Hodgkin genitive diseases, NOS Remove stop Hodgkin diseases, NOS o Hodgkin words diseases, Convert to Hodgkin diseases, o hodgkin diseases, lowercase Strip hodgkin diseases, o hodgkin diseases punctuation Uninflect hodgkin diseases o hodgkin disease phrase Sort hodgkin disease o disease hodgkin words
Table 2a. UMLS normalization steps [8, slide20]
The process of matching the PL terms involved cycling through the matching algorithms one at a time to find the best candidate SNOMED CT concepts. For each algorithm we always began with the original terms, then the UMLS normalized terms, followed by the stemmed terms. During each cycle, we would review the candidate concepts found to determine if it was a match, and if so, what type of match it was based on the algorithm applied. When no matching concepts were found, we would label the term as unmatched. Our experience with the matching algorithms had been that, the sooner we could find a match in the cycle, the greater confidence we would have that the candidate concept is appropriate. The preferred order of matching selected is always exact first, then all, followed by partial. For exact-match and match-all if only inactive concepts are found then a semantic-match is done to find their corresponding current concepts through the historical relationships.
SNOMED Prefixes Step-5 Stop words Exclude words
Explanation Frequent short words that do not affect the phrase: and, by, for, in, of, on, the, to, with, no, (nos) Words that may change meaning of the word but if ignored help to find a term otherwise missed: about, alongside, an, anything, around, as, at, because, before, being, both, cannot, chronically, consists, covered, does, during, every, find, from, instead, into, more, must, no, not, only, or, properly, side, sided, some, something, specific, than, that, things, this, throughout, up, using, usually, when, while [X] – concepts with ICD-10 codes not in ICD-9 [D] - concepts in ICD-9 XVI and ICD-10 SVII [M] – morphology of neoplasm concepts in ICD-O [SO] – concepts in OPCS-4 chapter Z in CTV3 [Q] – temporary qualifying terms from CTV3 [V] – concepts in ICD-9 and ICD-10 on factors influencing health status and contact with health services (V-codes and Z-codes)
Table 2b. Expanded UMLS normalization step-2
The process of encoding the problem lists extracted from the EMR followed these steps: (a) tabulating the frequency of occurrences for all of the original PL terms; (b) cataloguing all of the unique words across the PL terms present; (c) examining all unique words and PL terms to identify and revise for acronyms, abbreviations, spelling variants and errors; (d) matching the PL terms to SNOMED CT concepts using matching algorithms described earlier; (e) producing detailed and summary outputs to show the type of matches found; (f) reviewing/verifying the matched concepts one term at a time for accuracy; (g) repeating steps (c-f) until no further matches could be found; (h) examine remaining partial-matches for post-coordination; (i) create an index table of all PL and matched SNOMED terms. As part of this study, we also explored navigating within the SNOMED hierarchy to examine how the super-types and relations could be used to improve the quality of recall using the matched SNOMED concepts.
A total of 7,833 PL entries were extracted from the EMR for this study. The majority of these entries were recorded by one practitioner over a 7-year period. Of these entries, there were 1,713 unique PL terms present. Based on the frequency distribution of the entries, the top 10 PL terms were hypertension, hypercholesterolemia, diabetes mellitus, hypothyroid, asthma, atrial fibrillation, gastroesophageal reflux, depression, congestive heart failure and chronic kidney disease. After the second cycle we had 1,296 (88.23%) exact-matches where the PL terms are exactly the same as the SNOMED terms found. There were 51 (3.47%) match-all where all the words in the PL terms are present in the SNOMED terms but not necessarily in the same sequence. There were 120 (8.17%) partial-matches where one or more words matched the SNOMED terms. Another 20 (1.42%) SNOMED terms were found with semantic matches. Between the two cycles partially-matched terms were revised to tease out qualifiers and secondary concepts if present in order to explore post-coordination. A summary of the PL terms and the SNOMED matches found is shown in Table 3.
In Table 4 we have examples of the frequently used PL terms with their SNOMED terms found by exact, all and semantic matches. Also shown are the matches after revision and post-coordination of the original and partially-matched PL terms. For most exact-matches we selected the preferred terms from SNOMED CT as they are identical or closest to the original PL terms, such as Atrial fibrillation. In some cases we chose the synonym terms, such as Hypertension instead of the preferred term which is Hypertensive disorder. For match-all and some partial-matches we selected the SNOMED terms that were closest to the PL concept involved, such as GERD gastro-esophageal reflux disease. For semantic matches we looked up the current concepts of the matched but inactive SNOMED terms through their historical relationships, such as Cirrhosis. For post-coordination we added qualifier and refinement terms to SNOMED concepts or combined those that are lexically closest to the original PL terms, such as Atrial fibrillation+Chronic, Kidney disease+Chronic, and Headache+Migraine. After the second cycle any remaining partial-matches were treated as unmatched. Initially there were eight PL terms not found in SNOMED CT. Five were spelling errors and were revised for the second cycle (e.g. hepatomegally o hepatomegaly); three were legitimate missing terms – vasculopath, pyocystitis and hypotestosteronemia, where we had to modify the PL term or tag as local extensions. Using these outputs we created an index table to link the PL terms to their matched SNOMED terms, shown in Table 5. Each row contains the PL-termId, conceptId, descriptionId, relationship-typeId match-type, and post-coordination-sequenceId.
Description Frequency No. of patients 2,894 Total PL entries 7,833 Total words in PL terms 16,455 Unique words 1,764 Longest word Hypercholesterolemia, 20 characters Median length 8 characters Most common word Hypertension, 585 times Matching Initial Cycle 2nd Cycle Algorithm Frequency (%) Frequency (%) Exact-Match 905 (52.83%) 1,296 (88.23%) Match-All 167 (9.75%) 52 (3.47%) Partial-Match 633 (36.95%) 120 (8.17%) Semantic-Match 49 (2.86%) 20 (1.42%) Unmatched 8 (0.47%) 2 (0.14%) Post-coordination Not done In-progress Total unique PL terms 1,713 1,468 Table 3. Summary of PL terms and matches. For frequency %, once a match has been found it is not included as part of the next matching algorithm
Manual revisions were done on the 1,713 unique PL
terms after the initial cycle. By selecting the PL terms
that were not matched in SNOMED CT, we were
able to identify entries that were misspelled,
idiosyncratic local terms or ambiguous concepts. A
number of spelling mistakes were corrected. The
CliniClue Browser [
As part of this study, we explored ways to navigate the SNOMED hierarchy to determine if it could improve one’s ability to retrieve related concepts. Of the 1,296 exact matches found for the 1,468 unique PL terms present, we selected a subset of 32 PL terms related to cardiovascular disorders for this analysis. First, we did frequency counts of these PL terms to show how often they were present in the EMR system. For each PL term present, we navigated up the hierarchy until we reached the super-type “49601007|Disorder of cardiovascular system.” We then pruned the tree to include only those concepts with a positive frequency count, but left their immediate super-types intact. This partlyinstantiated cardiovascular disorder hierarchy is shown in Figure 1. The value of this tree is that it shows the SNOMED concepts that are actually present in the EMR and how often they occur via the frequency counts based on the PL terms recorded. This tree can aide in the retrieval of relevant concepts recorded using different PL terms. For instance, by specifying the concept “56265001|Heart disease” in the query, one should expect to retrieve all sub-types under “5754005|Acute myocardial infarction” and “12026006|Paroxysmal tachycardia.” On the other hand, by specifying the concept “57054005|Acute myocardial infarction” in the query, the sibling concept “12026006|Paroxysmal tachycardia” should automatically be excluded.
Drawing on the lessons learned from this study, we propose the following steps for general practitioners to encode problem lists from their EMR in SNOMED as a first step for review before full-scale conversion: 1. Extract all PL entries from the EMR and tabulate the frequency of the PL terms present; 2. Catalogue all unique words across the PL terms; 3. Examine all unique words and PL terms to identify and revise for acronyms, abbreviations, spelling variants and errors; 4. Match the PL terms to SNOMED concepts using the matching algorithms outlined in this paper (contact authors for copies of the algorithms); 5. Create detailed and summary outputs to show the exact, all, partial and semantic matches found; 6. Review matched SNOMED terms for accuracy; remove successful exact-match and match-all terms from further matching cycles; 7. Repeat steps 3 through 6 for remaining partial matches until no further matches found; 8. Post-coordinate remaining PL terms with qualifier, refinement and combined concepts; 9. Create a pruned PL hierarchy tree showing all concepts with positive frequency counts and immediate super-type concepts; 10. Create index table containing unique identifiers for the PL and matched SNOMED terms.
Post-coordination is thought to be a feature that is difficult to implement. Yet based on the small number of SNOMED concepts used in this study to post-coordinate our PL terms, it seems feasible to achieve. We did note the use of pre-coordination in SNOMED CT is unpredictable, and it seems common to include acronyms within SNOMED descriptions. Careful use of modifiers such as laterality, chronicity and severity should be considered. Further studies are needed.
Critics often balk at the unwieldy size and complexity of SNOMED CT as too impractical for local use. In Canada the vendor and general practice communities, which are often small in size, are reluctant to adopt SNOMED CT, questioning their return on value for the effort required. From this study, we have shown it is feasible to incorporate SNOMED CT into EMR in the general practice setting. The methodology we have outlined is practical even for small medical offices with an EMR in place. We have also shown the potential use of SNOMED CT to improve the quality of recall from its hierarchy. The ability to demonstrate return on value, as in our encoding of problem lists with SNOMED CT to improve recall, is an important first step for practitioners to consider before full-scale conversion of their EMR.
There are several limitations to this study. First, the PL terms used have been established over the years mainly by one practitioner from a single setting, which are likely to vary between practices. Second, our current matching algorithms do not take into account subtype hierarchy to limit searches, which could otherwise restrict unlikely choices such as Physical Object and Substance. Third, the evaluation of this methodology is incomplete to date; the full extent of the post-coordination effort required to encode the entire set of PL terms in this EMR should be further examined and reported. Fourth, the use of our partly instantiated hierarchy tree to improve recall quality, while promising, requires more thorough investigation into its utility with more complex real-life cases. Its design should also be aligned with the existing SNOMED navigation hierarchy feature that is already in place as part of the new RefSet release.
We are developing a Web-based mapping tool made up of the matching algorithms described earlier to allow the matching of clinical terms to SNOMED CT in an interactive or batch mode. With our focus continued to be on general practice EMR systems, there are several steps ahead to be considered. For instance, we need to expand the use of SNOMED terms to other parts of the EMR such as procedures, medications and billing. We also need to refine our encoding methodology to take into account specific contexts such as past/family history and health risks, and to use subtype hierarchy to improve search precision. The inclusion of frequency statistics on the distribution of matched SNOMED CT terms across the hierarchies would be useful to validate the results. These efforts should aid in the eventual creation of a primary care SNOMED subset, and eventually a concept model in the primary care domain. But most important, we should continue to exploit ways by which the use of SNOMED CT in the EMR can actually enhance patient care.
Funding support for this project has been provided by the Canadian Institutes for Health Research Strategic Training Initiative. Two sets of post-coordinated terms shown above