Using SNOMED CT Codes for Coding Information in Electronic Health Records for Stroke Patients Judith van der Kooij a,1, William T.F. Goossen a, Anneke T.M. Goossen-Baremans a, Marinka de Jong-Fintelman a and Lisanne van Beek a a Acquest, Koudekerk aan den Rijn, the Netherlands Abstract. For a project on development of an Electronic Health Record (EHR) for stroke patients, medical information was organised in care information models (templates). All (medical) concepts in these templates need a unique code to make electronic information exchange between different EHR systems possible. When no unique code could be found in an existing coding system, a code was made up. In the study presented in this article we describe our search for unique codes in SNOMED CT to replace the self made codes. This to enhance interoperability by using standardized codes. We wanted to know for how many of the (self made) codes we could find a SNOMED CT code. Next to that we were interested in a possible difference between templates with individual concepts and concepts being part of (scientific) scales. Results of this study were that we could find a SNOMED CT code for 58% of the concepts. When we look at the concepts with a self made code, 54,9% of these codes could be replaced with a SNOMED CT code. A difference could be detected between templates with individual concepts and templates that represent a scientific scale or measurement instrument. For 68% of the individual concepts a SNOMED CT could be found. However, for the scientific scales only 26% of the concepts could get a SNOMED CT code. Although the percentage of SNOMED CT codes found is lower than expected, we still think SNOMED CT could be a useful coding system for the concepts necessary for the continuity of care for stroke patients, and the inclusion in Electronic Health Records. Partly this is due to the fact that SNOMED CT has the option to request unique codes for new concepts, and is currently working on scale representation. Keywords: Controlled Vocabulary, Coding System, Medical Records Systems, Electronic Health record, Medical Informatics, SNOMED CT, Electronic Messages, HL7 v3 1. Introduction Health care is a domain for which information plays an important role. This goes not only for registration and declarations, but even more for the direct care for patients. Especially for the documentation of patient records there is no univocal universal language. At this moment classifications that are meant for classifying diseases or treatments are used for various purposes, both administrative and clinical. Most of the time, these classifications are not developed for the registration of direct care for patients, but more for statistics and billing. For that reason, those classifications are not always suitable for the registration of daily care observations and activities. Under the authority of NICTIZ, the abbreviation of the Dutch name for the National IT Institute for Healthcare [1], a study has been conducted with the main goal to determine the usefulness of a broad introduction of a clinical terminology in The Netherlands. In the report of this study a clinical terminology has been described as: the collection of standard terms with their synonyms, which can be used in direct patient care to record all symptoms, circumstances, interventions, diagnosis, results and the decision making [1]. To make information technology work in health care one needs to accept standardisation of the data, as well as the vocabulary and the electronic messages, next to security of medical information. In the Netherlands, NICTIZ takes responsibility for this [2]. According to NICTIZ standardisation is necessary to realize clinical data exchange between different Electronic Health Record (EHR) systems, all with their own characteristics. NICTIZ chose Health Level Seven version 3 (HL7 v3) [3] to be the standardisation methodology for messaging. In several projects in which HL7 v3 was used, we identified the need for an additional coding system, next to existing coding systems, because unique codes were needed. At first we decided to invent new codes as a temporary solution to this problem; however, this is an undesirable situation in the long run, due to maintenance and interoperability issues. For this reason an explorative study was carried out for using SNOMED CT codes as standardized clinical terminology. SNOMED Clinical Terms (SNOMED CT) is a dynamic, scientifically validated clinical health care terminology and infrastructure [4]. By applying SNOMED CT coding, data can be captured, shared and aggregated in a consistent way across specialities and domains of care. Due to the SNOMED CT Core clinical terminology, SNOMED CT can be used for electronic medical records, ICU monitoring, clinical decision support, medical research studies, clinical 1 Corresponding Author: Judith van der Kooij, Dorpsstraat 50, 2396 HC Koudekerk aan den Rijn, the Netherlands, acquest@acquest.nl.
trials, computerized physician order entry, disease surveillance, image indexing and consumer health information services [4]. In this study we tried to replace the present codes, self made or from existing coding systems, with SNOMED CT codes in the care information models we created for an EHR for the continuity of care of stroke patients. A care information model is often also referred to as a template [3, 5]. The care information models, or templates, that were developed for the EHR for stroke patients, integrate knowledge, terminology and coding, and information models, with HL7 v3 as the underlying standardisation method [5, 6]. These models were derived from (paper) records given to us by the care professionals that were involved in the project for the EHR for stroke patients. After creating these templates they were shown to the care professionals for validation and feedback. If necessary they were corrected so they would correspond to the domain of stroke. The goal of this study was to explore the usefulness of SNOMED CT for uniquely coding the medical information in the templates for the stroke domain. This is part of the ongoing development of standards, messages and the EHR for stroke patients [7] The following research questions were formulated: 1. For how many of the codes of the clinical concepts in the care information models can we find unique SNOMED CT codes? 2. Is there a difference for individual concepts and concepts representing scientific scales or measurement instruments with clinimetric characteristics? The results of the study were not only quantitative results but also qualitative results and recommendations. An important recommendation is to perform real life experiments with SNOMED CT in an EHR and in HL7 v3 messages. As a result of this, one can underpin decisions about the introduction of SNOMED CT as the national standard terminology for coding in EHR and HL7 v3 messages. 2. Method 2.1. Background We were involved in a project on creating an Electronic Health Record (EHR) for the complete chain of care for stroke patients. We gathered all the information that is recorded for the stroke patients in the present situation. Then this information was organised. This organisation resulted in 84 templates. In such templates (validated) scales or instruments, observations or actions are described in detail [5]. These represent best practice, are Health Level 7 compliant, support the uptake of standardized terminologies and facilitate technical implementation in both electronic messages and clinical information systems. One of the paragraphs of the templates describes the mapping table from the domain to the HL7 Reference Information Model and message models. In this mapping table all items that are recorded in an EHR receive a unique code. A unique code is needed to exchange the information with other systems; this is called semantic interoperability [3]. Preferably these unique codes would be adopted from existing coding systems, so first we tried to look for an appropriate code in IDC10, ICF and ICNP, in order of preference. Next to that, LOINC was also searched for unique codes. However, most items could not get a unique code from these existing coding systems. This resulted in self made codes, which are unique but do not correspond to an existing coding system, are difficult to maintain, and, due to lack of standardization, will only support partial interoperability. White and Hauan [8] discuss the way the LOINC coding system can adequately represent instruments and scales. They argue that a particular important aspect of coding is to maintain the psychometric or clinimetric properties of instruments and scales Although this work was carried out with another coding system, their criterion refers to the domain content and therefore, we believe, is relevant for any coding system applied: the meaning of the concept in the scale should precisely be represented in the wording and in the coding. After creating the 84 templates, NICTIZ wanted to test if SNOMED CT codes could replace the self made codes with standardized codes. There have been other studies carried out to test the breadth of SNOMED CT [9, 10, 11]. In the study of Campbell et al. [9] three potential sources of controlled clinical terminology were compared (READ codes version 3.1, SNOMED International, and Unified Medical Language System (UMLS) version 1.6) relative to attributes of completeness, clinical taxonomy, administrative mapping, term definitions and clarity (duplicate coding rate). The authors assembled 1929 source concept records from a variety of clinical information taken from four medical centres across the United States. The source data included medical as well as sample nursing terminology. The study showed that SNOMED was more complete in coding the source material than the other schemes, because SNOMED covered 70% compared to READ covering 57% and UMLS 50%. From this study it could be concluded that SNOMED was more complete, had a compositional nature and a richer taxonomy.
Chute et al. [10] reported a similar result when evaluating major classifications for their content coverage. For their study the clinical text from four medical centres was sampled from inpatient and outpatient settings. This resulted in 3,061 distinct concepts. These concepts were grouped into Diagnoses, Modifiers, Findings, Treatments and Procedures, and Other. Each concept was coded into ICD-9-CM, ICD-10, CPT, SNOMED III, Read V2, UMLS 1.3, and NANDA. When coding the concept, the reviewers also scored the concepts: 0 = no match, 1 = fair match, 2 = complete match. Result of this study was that SNOMED had a broader coverage than any of the other coding systems used in this study. SNOMED received the highest score in every category, including Diagnoses (1.90), and had an overall score of 1.74. Wasserman and Wang [11] found a concept coverage of 88,4% when evaluating the breadth of SNOMED CT terms and concepts for the coding of diagnosis and problem lists within a computerized physician order entry (CPOE) system. When they took the relevance of the 145 terms that could not be coded with SNOMED CT into account, they could even conclude that the concept coverage of SNOMED CT was 98,5%. Although these three studies [9, 10, 11] showed that SNOMED CT has a rather high concept coverage, we expected to find SNOMED CT codes for about half (50%) of our self made codes. This 50% assumption was based on our experiences with looking at other existing coding systems used for these templates and the vast amount of self made codes that where necessary for all clinical details for stroke patients during their full episode of care [7]. 2.2. Description of the study From the American College of Pathologists we received a licence for the project, material about the structure, the contents of SNOMED CT and instructions on how to search for concepts, terms and their corresponding codes. After studying this material, all items from the mapping tables of the templates have been systematically searched for in SNOMED CT. We decided not only to search for codes for items that had a self made code, but also for items that had a code from another coding system, like ICD 10. We did this to find out how complete SNOMED CT is for our purposes. To make sure we used SNOMED CT in the right way, we used the knowledge of experts. Like, for example, the document on how to search in SNOMED CT written by Casey [12]. Based on this document we developed a search strategy; this strategy was as follows: 1. translate the existing Dutch concepts in the mapping table in the templates into English (this was done, as a requirement from NICTIZ, during the construction of the templates); 2. start searching with the translated concept, as mentioned in the mapping table of the templates; 3. when there is no perfect match, search for a concept on the SNOMED CT hierarchical levels above; 4. when there is no perfect concept, search with synonyms; 5. when still no perfect match can be found, search using the SNOMED hierarchy from top down. When we look at the care information model for body temperature, for example, we reported the following search strategy. The model for body temperature consists of two concepts: body temperature and method of measuring. For the first concept the term body temperature was entered in SNOMED. This resulted in a hit: 386725007: body temperature. This resulting term fully represented the concept, so the SNOMED CT code was accepted. Then the second concept, method of measuring, was entered in SNOMED. This generated the following result: 371911009: measurement of blood pressure using cuff method. This was not the right concept. The hierarchical levels that lied above were also about blood pressure, so no result could be found via this strategy. Then we entered a kind of synonym: measurement of body temperature. This only generated a concept related to ovulation, which was not what we were looking for. The next step in our strategy was to search top down. For this we entered the following terms respectively: body temperature, measurement and method. All three remained without a result. So, no SNOMED CT code could be found for method of measuring of body temperature. For every template a short report has been made. In this report, for every item, it was reported what the search results were and what result was chosen as the right one for the concept put in the search, together with a motivation for choosing this SNOMED CT concept. Next to that, other remarks on the search or the search results were reported. When in a search a synonym was needed, this synonym and its accompanying search results were reported. An expert on medical terminology reviewed the complete search report and indicated which search results should be accepted and which should be rejected. Next to this, the expert also corrected codes by advising to use another SNOMED CT code. Sometimes neurologists were approached to come up with synonyms for a certain item, which apparently was also not clearly formulated in the original Dutch wording. The final SNOMED CT codes were added to the mapping table in the templates, so the original codes, self made or from another coding system and the SNOMED CT codes can be used next to each other. The adjusted templates have been sent to Portavita, a company that works on the SNOMED CT implementation of the EHR for stroke patients for DWO in Delft, The Netherlands. This information system is an EHR for all health professionals involved
in the care of a stroke patient. This means that all (medical) information that comes from the GP, the hospital, the rehabilitation centre, the nursing home, and home health care is put in one record. When this system is implemented, the messages, containing patient related information, will be sent from system to system, while using SNOMED CT codes next to the original codes. This way the applicability of SNOMED CT in the EHR and in HL7 v3 messages will be tested. This will also lead to another report in the near future. 3. Results For this research 84 templates need to be coded. From these 84 models, up to today, 32 have been coded. These 32 care information models that have been coded contained 207 concepts or items. From these 207 items 87 could not be coded with SNOMED CT (42%). For 120 (58%) a SNOMED concept ID code has been found. For 178 (88,1%) items there was agreement either for the code that was found or for the fact that no code could be found. For this result a remark must be made. The items in the care information models are grouped in a HL7 v3 way: by using Organizers and Batteries. For example: the nursing assessment contains 2 Organizers, nursing record and decubitus. The nursing record contains 7 templates, each with its own items and Batteries to group these items. SNOMED CT does not support this grouping principle, so none of the Organizers and Batteries can be coded with SNOMED CT. And although these Organizers and Batteries all need a unique code we can not expect SNOMED CT to add these codes because they do not represent medical information, just grouping of medical information. For this study we decided not to code Organizers and Batteries and to leave them out of the calculation for the results. When we make a distinction between self made codes and codes from an existing coding system we see the following result. For the 164 self made codes 90 could be coded with SNOMED CT (54,9%). For the 43 codes from an existing coding system, like ICD 10, 32 could be coded with SNOMED CT (74,4%). To answer our second research question we also looked at the difference between SNOMED CT codes found for templates with individual items and templates representing scientific scales or measurement instruments. The results are presented in Table 1. Table 1: Difference between SNOMED CT codes found for care information models with single items and care information models representing scientific scales. Amount of care information models Amount of concepts Amount of SNOMED CT codes Percentage of SNOMED CT codes 6 scientific scales 50 13 26% 26 models with individual 157 107 68% concepts 32 in total 207 120 58% Next to these quantitative results we would also like to report some relevant qualitative findings. First, items of the care information models which exist of two (or more) combined concepts do not always correspond with one concept ID within SNOMED CT. Our solution was to report all the separate codes. For example, when examining the family history of a stroke patient one wants to know if stroke at an early age runs in the family. For this concept we needed three SNOMED CT codes: one for stroke, one for age and one for young. So sometimes concepts correspond to two (or more) SNOMED CT concept ID s; the question is whether it is allowed to combine two (or more) separate ID s. Second, the terminology used in the clinical area can not always be found in SNOMED CT. In this case we tried to find a concept ID that represents the concept that lies behind the terminology used by the care professional. Third, the English translation of the Dutch items did not always result in a SNOMED CT code. Then we tried synonyms. Fourth, items with a left or right indication, like fingers extensors left, can not always be found in SNOMED CT. We solved this problem by using the attribute in an HL7 v3 class in which, for instance, location can be entered. However, this does not solve the terminological issue. On the other hand, this makes the addendum left and right superfluous in the terminology. For example, finger extensors left and finger extensors right can be replaced by just one item, namely finger extensors, and the left and right are covered in the information model. Fifth, as mentioned earlier, the Organizer and Battery concepts are not supported by SNOMED. We decided not to search for SNOMED codes for the Organizer and Battery concepts anymore. Sixth, the degree of detail within SNOMED CT is very different. Some concepts are coded within detail and other concepts just have one code for the concept itself and no codes for the underlying details of the concept. For example, regularity and constancy of breathing are common observations for stroke patients, however no concept ID s and codes can be found in SNOMED CT. Yet, frequency and profound breathing can be found. Seventh, the items in the care information models are mainly observations, in SNOMED these need to be observable entities. Though, most items can only be found as clinical findings instead of observable entities. Eighth, in SNOMED CT the use of stimulants, like alcohol,
marihuana, and cigarettes is defined as abuse of these stimulants although the use of stimulants is not always considered as abuse, if used with care. Using SNOMED CT here would imply a wrong meaning: use is intended to document, and abuse is implied by the terminology. For now we decided not to use the SNOMED codes for these items. 4. Discussion In our study we could find 58% of the items in SNOMED CT. Until now we coded items for only 32 templates; 52 still need to be finished. The percentage of items found might be higher when all models are coded, although we expect the result of these 32 care information models to be representative for all models. It might even be that the percentage of SNOMED CT codes found will be lower after coding the items of all templates because 17 of the 52 models that still need to be coded are scientific scales or measurement instruments. For these scales or instruments, the concepts need to be exactly the same as the scoring items of the tests or scale [8], including the answering possibilities. Hardly any tests are included in SNOMED CT As was shown above only 26% of the concepts coming from scientific scales or instruments could be found in SNOMED CT against 68% of the individual concepts. The current 58% is slightly disappointing compared to similar studies [9, 10, 11], however, we believe that we have dealt with difficult concepts in our study. These concepts are difficult in two ways: first is the very large granularity of the clinical concepts necessary for the care of stroke patients. These are usually very fine grained details of muscle functions, body position, thought processes etcetera, and in some instances expressed in quite awkward wording. Second, the templates are often especially made to have an accurate representation in HL7 v3 messages of scales, which have specific clinimetric characteristics [5, 8]. These clinimetric characteristics require an accurate equivalence between the concepts as used in practice and in the clinical terminology used for the unique coding [5, 8]. In addition, translation errors and or cultural differences might be a reason for this lower percentage. SNOMED CT has predominantly been built for the English language realm. We worked with clinical concepts in Dutch, given to us by clinicians, often in their local wording. They might use slightly different terms that could not be translated one to one from Dutch to English. For a more reliable result we could ask the care professionals, who gave us the (medical) information on which we based the concepts in the templates, to review the results of our search in SNOMED CT. They are the best to check if the concepts they use in daily practice are well represented by the concepts we found in SNOMED CT. Next to that they might also be able to formulate synonyms for concepts without a SNOMED CT code, although the translation from Dutch to English might still cause a bias. 5. Conclusion We could answer the question as to how many of the concepts from the care information models, that where initially developed for the EHR and HL7 v3 messages for stroke patients continuity of care, could receive a SNOMED CT code. Currently, with a quite difficult set of concepts in the area of stroke care, we could find an overall coverage of 58%. It is important to have standardized clinical terminology such as SNOMED CT applied in the EHR and HL7 v3 messages instead of reinventing the wheel by using self made codes. The motivation for this is to truly achieve semantic interoperability in the exchange of patient information [2, 3]. Based on this study we could replace these self made codes for a bit over half of the clinical concepts. Despite the result of this study we can conclude that SNOMED CT has several characteristics that make it useful to continue its application in stroke care, and after further testing, in the Dutch national infrastructure. First, SNOMED CT has the possibility to request inclusion and coding of concepts that could not be found in SNOMED CT. This means that expansion is an option. Second, SNOMED CT has an ongoing project for scale representation that takes the clinimetric aspects of scales and concepts into account. Third, SNOMED CT is working on further internationalisation in order to meet European requirements. The clinical use in the Electronic Health Record system, which is built at the moment, and in HL7 v3 messages will reveal additional information about the usability of SNOMED CT in the clinical care for stroke patients and is therefore recommended. Of course it is also recommended to have more clinical areas than just the area of stroke researched in such a way. A final recommendation would be to tackle the translation issue.
Acknowledgments We thank the care professionals for providing the information for the care information models; a special thanks to the neurologists Dr. Weinstein (Amsterdam), Dr. Swen (Delft) and Dr. Dippel (Rotterdam). We also thank Portavita, developer of the EHR for stroke patients and implementer of the HL7 v3 messages, for working together on this project. The American College of Pathologists are acknowledged for offering a SNOMED CT licence to carry out different tests for the stroke patients EHR and message development. Anne Casey, Davera Gabriel, Deb Konicek and David Markwell for teaching us how to search SNOMED, for answering our questions and for considering adding concepts to SNOMED CT. This study could be carried out through funding from the Dutch National IT Institute for Healthcare (NICTIZ). References [1] Nationaal ICT Instituut in de Zorg, NICTIZ (2003). Introduction of a clinical terminology in the Netherlands, Needs, Constraints, Opportunities. Leidschendam, NICTIZ. [2] Nationaal ICT Instituut in de Zorg. Web documents [Online]. [cited January 10, 2006]; Available from: URL: www.nictiz.nl. [3] Health Level 7. Web documents [Online].[cited January 10, 2006]; Available from: URL: www.hl7.org. [4] SNOMED Clinical Terms. Web documents [Online]. [cited December 28, 2005]. Available from: URL: www.snomed.org. [5] Kooij van der J, Goossen WTF, Goossen-Baremans ATM, Plaisier N. Evaluation of documents that integrate knowledge, terminology and information models. Article submitted for publication. [6] Goossen WTF. Templates: an organizing framework to link evidence, terminology and information models in the nursing profession. In: de Fatima Marin H, Pereira Margues E, Hovenga E, Goossen W, editors. E-Health for all: designing a nursing agenda for the future. Proceedings 8th International Congress in Nursing Informatics NI 2003. Rio de Janeiro, Brazil, E-papers Serviços Editoriais Ltd, pp. 461-465. [7] Reuser L, Goossen WTF, van der Heijden H. Call for ICT within stroke service care. In: Runnenberg J, et al, editors. Health Information Developments in the Netherlands by the year 2003. 6 th ed, pp. 52-56. [8] White TM, Hauan MJ. Extending the LOINC conceptual schema to support standardized assessment instruments. Journal of the American Medical Informatics Association 2002; 9 (6):586-599. [9] Campbell JR, Carpenter P, Sneiderman C, Cohn S, Chute CG, Warren J. Phase II evaluation of clinical coding schemes: completeness, taxonomy, mapping, definitions, and clarify. Journal of the American Medical Informatics Association 1997; 4:238-51. [10] Chute C, Cohn S, Campbell K, Oliver D, Campbell JR. The content coverage of clinical classifications. Journal of the American Medical Informatics Association 1996 May/June; 3:224-233. [11] Wasserman H, Wang J. An applied evaluation of SNOMED CT as a clinical vocabulary for the computerized diagnosis and problem list. AMIA 2003: Annual Symposium Procedure; 2003. p.699-703. [12] Casey A. Finding terms in SNOMED CT - supplement to browser guidance 2005 June. Internal report SNOMED CT.