There are various types of medical terminology systems to satisfy different needs. To satisfy more needs than what exist today, both Rossi Mori et al. 1 and Cimino 2 ask for an evolution of the medical terminology systems for more flexiblity.

Rossi Mori et al. describe three generations of medical terminology systems 1. The first generation comprises traditional terminology systems 1. This generation includes controlled vocabularies, nomenclatures, taxonomies and coding systems which satisfy most needs in paper-based information systems. In this generation, systems typically consist of a list of phrases, a list of codes, a coding scheme and a hierarchy. The role of the coding scheme is to map between phrases and codes 1. Examples of systems in the first generation are ICD-10, KSH97-P and International Classification of Functioning, Disability and Health (ICF).

The second generation are compositional systems. These systems have a categorical structure, a cross-thesaurus, a structured list of phrases and a knowledge base of dissections 1. The categorical structure gives a high-level description of the content, i.e. what kinds of concepts are included and how they relate to each other. This can be seen as a framework of slots for which the cross-thesaurus provides a set of labels to be inserted when the content is modelled. By means of the cross-thesaurus, each element in the structured list of phrases is represented according to the categorical structure; these descriptions constitute the knowledge base of dissections. Examples of systems in the second generation are Nomenclature, Properties and Units (NPU), Logical Observation Identifiers, Names and Codes (LOINC), and SNOMED International 1.

The third generation consist of formal systems. In this generation, the systems have a set of symbols and a set of formal rules to manipulate the symbols and these sets can be seen as a set of concepts and a set of relations between the concepts 1. It is possible to represent each concept in a unique canonical form and a non-canonical expression may be automatically converted to a unique canonical form using an engine. An example of a third generation system is GALEN-IN-USE's surgical procedures 1. SNOMED CT is evolving towards a third generation system.

One problem in the first generation is that reorganisation of categories in the systems to satisfy different purposes is not supported 1. Reuse of data organised with first generation systems therefore needs human interpretation of the categories and the environment where the data was originally collected. This problem is smaller in the second generation and even smaller in the third generation. In the second generation, categories can be reorganised according to the information in the knowledge base of dissection. In the third generation, formal rules can be used for reorganisation 1.

Cimino enumerates twelve characteristics of the structure and content in medical terminology systems in "Desiderata for controlled medical vocabularies in the twenty-first century" and these characteristics emerge from earlier vocabulary research 2. Four of these characteristics are relevant in this study

• Poly-hierarchy. Systems need to shift from a strict mono-hierarchy (taxonomy) to a poly-hierarchy 2. It is impossible to properly represent the real world in a strict mono-hierarchy where each category has only one parent. The categories in the real world can belong to more than one parents 2345.

• Formal definitions. Systems need formal definitions expressed as collections of different kinds of relationships between the concepts 2. Formal definitions can be used by computers for formal manipulations of the categories which is impossible with unstructured text definitions 2. One manipulation is to help a user locate a specific category in a terminology system 3. A similar manipulation is locating where to include new categories in the system's structure 345. Another manipulation is to test whether a pre-coordinated category is equivalent to a set of post-coordinated categories or whether different sets of post-coordinated categories are equivalent 5.

• Multiple granularity. Systems need to have concepts of different granularity covering the same area 2. Different use cases need systems with different granularities depending on the required level of detail of the categories. A multipurpose system therefore needs multiple granularities 2. One use case is abstracting information in health records to allow compilations of health records' contents. Another use case represents sufficient detail of the information in health records in order to use the information in for example direct patient care, decision support and quality assurance 5.

• Multiple consistent views. Systems need to be able to consistently present their content in different views 2. Some use cases require a simple structure of the system's categories and others need a richer structure. The kind of structure depends on the required level of detail and required type of information of the categories 23. To present equivalent information independent of the view used, the views need to be consistent 23.

Straub et al. have a different opinion than Rossi Mori et al. and Cimino as presented above 6. They argue that the different kinds of medical terminology systems have different purposes and need to co-exist. Medical terminology systems with fewer categories and a semantic model with more restrictions, such as a hierarchical tree, provide useful information reduction or simplification for cases where a richer medical terminology system provides too much information 6.

Hierarchical trees are disjunctive and unidirectional 6. Disjunctive means that all categories on one level are mutually exclusive and unidirectional signifies all hierarchical relations only go in one direction. Unfortunately, diseases are not disjunctive and unidirectionally related to each other and therefore it is not possible to construct a hierarchical tree of diseases based on the diseases' characteristics 6. If a hierarchical tree is still constructed from disease categories, for which Straub et al. think there are good reasons, the hierarchical tree is artificial. The drawback of artificial hierarchical trees is the structure is arbitrary. This means that in the construction of the tree, some information is hidden -- the information on which the hierarchy is not based 6.

ICD-10 7 is primarily intended for statistical reporting and administrative tasks such as disease monitoring and quality assurance 8. Although neither based on nor intended as a model of health problems, but pragmatically developed from the admittedly arbitrary structure proposed by William Farr in 1855 910, the ICD classifications are by far the most used terminology systems in electronic health records 11.

Farr's structure, which is reflected in how diseases are divided into chapters in the ICD-10, groups diseases into five sets 910:

• epidemic diseases

• constitutional or general diseases

• local diseases arranged by site

• developmental diseases

• injuries

The presentation of ICD-10 79 focuses on the role as a member of a family of classifications rather than the internal structure. In ICF, one part of the introduction describes a conceptual framework for the classification 12; that kind of model does not exist for ICD-10.

While ICD classifications are mono-hierarchical, the International Classification of Primary Care (ICPC) 13, originally published in 1987 and later in a second 14 and a revised second edition 15, is bi-axial, consisting of chapters and components. Here, a patient's reason for encounter, health problems to be taken care of and interventions are classified and coded according to a chapter structure. The chapter structure is based on body systems and problem areas and a set of components specifying the nature of the phenomenon coded such as a complaint, procedure or disease.

The move towards the third generation of terminology systems with formal definitions of disorders has proven to be a challenging task 16. This is especially valid if diagnostic criteria are to be taken into account as is the case for psychiatric diagnoses in the Diagnostic and Statistical Manual for Mental Disorders (DSM) as well as ICD 17181920. Version 3 of the Read Codes, a constituent of SNOMED CT, presented a template-based mechanism with attributes and values for basic semantic operations on items 2122. A set of categories describing completeness of definitions was developed as a by-product in the process of disorder definition 21. However, Version 3 of the Read Codes is still a second generation terminology system 1.

Héja et al. have presented work on formal definitions of the ICD-10 based on the GALEN 23 and DOLCE 24 formalisms with the main objective of providing a knowledge-based coding support tool. They found that although lexical processing 23 as well as existing terminology resources 24 may assist formal representation, ICD categories themselves--owing to the historical development rooted in epidemiological considerations--deviate from what is expected in contemporary ontology engineering 23. The result is a need to distinguish the meaning of categories from the structure of the classification, which essentially was the underlying rationale in the early modelling work reported by Petersson et al. 25. Such pitfalls of pragmatic classifications have also been reported in surgery, an area that modelling-wise is usually considered more straightforward than the domain of diseases 26.

Alecu et al. created a grouping of the categories in the World Health Organisation - Adverse Reaction Terminology (WHO-ART) based on mappings between WHO-ART and SNOMED CT in the Unified Medical Language System (UMLS) Metathesaurus 27. More specifically, they used synonym relations between WHO-ART categories and SNOMED CT concepts, creating synonym relations for 85.9% of all categories.

As pointed out by Rossi Mori et al. 1, and demonstrated by Alecu et al. 27, second and third generation systems can augment first generation systems with easier re-organisation and maintenance and with harmonisation and cross-referencing of different first generation systems. The description of the categorical structure could also be used for systematic comparison of terminology systems such as ICD, ICPC and SNOMED CT. Ingenerf and Giere argue along the same line when they explore the different roles of statistical classifications and formal concept representation systems, deducing the need for co-existence and the former being linked to the latter 28. The empirical results described above indicate these merely theoretical assertions require considerable thought before they are realised, which is consistent with the finding that little evidence, other than theoretical, exists on the usefulness of SNOMED in clinical practice 29.

The primary health care terminology system "Klassifikation av sjukdomar och hälsoproblem 1997 Primärvård" (KSH97-P) is based on the International Statistical Classification of Diseases and Related Health Problems, Tenth Revision (ICD-10). The general objective is to explore whether mappings from KSH97-P to SNOMED CT and SNOMED CT's structure can be used to enrich KSH97-P's mono-hierarchical structure. The enrichment thereby hypothetically provides useful multiple views of a disease panorama as coded with a traditional disease classification. The objective contrasts with the related work by Héja et al. 2324 presented above, where the objectives were to develop a new formal concept representation system of ICD categories, but is in line with the intentions of Alecu et al. 27. The results are discussed in relation to improvements of medical terminology systems as presented in the background.

The first specific question is whether SNOMED CT's poly-hierarchical generic structure can be used to add a multiple chapter division to KSH97-P's categories where each category may belong to more than one chapter. The second specific question is whether SNOMED CT's defining attribute relationships can be used to add attributes to KSH97-P's categories.

SNOMED CT is a clinical terminology intended for clinical documentation and reporting 30. In other words, SNOMED CT covers both abstraction and representation 5. It consists of concepts, descriptions and relationships 30.

Here, a concept is a clinical meaning and is identified by a unique number. Associated with each concept are two or more descriptions, which are human readable terms, and information about the terms 30.

Relationships link concepts to each other and are of different relationship types 30. The generic relationship type Is a relates subtypes to supertypes and is always a defining relationship. All concepts, except for the root concept, have at least one Is a relation to a supertype concept 30. The other relationship types that are defining relationships are the defining attribute relationships. The defining relationships logically represent a concept by establishing relationships between the concepts 30.

A concept in SNOMED CT can either be fully defined or primitive 30. A fully defined concept is modelled as described above so it is possible to distinguish the concept from the other concepts through its relationships with other concepts. Primitive concepts lack one or more relationship(s) to be able to fully distinguish from other concepts using the concept's relationships 30. There is also a concept model that controls which types of concepts can be related to which types of relations 30.

Concepts in SNOMED CT can be retired from active to inactive concepts 30. Inactive concepts have historical relationships that relate the inactive concepts to active concepts. The historical relationships can be used to point out active concepts that replace inactive concepts 30.

The Swedish National Board of Health and Welfare has worked out a primary health care version of the International Statistical Classification of Diseases and Related Health Problems (ICD-10) 31 (in Swedish, Klassifikation av sjukdomar och hälsoproblem 1997 Primärvård (KSH97-P)). Codes and rubrics (in both Swedish and English) of KSH97-P together with mappings to ICD-10 can be downloaded from their Web site 32.

KSH97-P contains 972 categories which concern diseases and health-related problems common in primary health care 31. Most categories in KSH97-P correspond to categories in the three or four-character levels in ICD-10. Some categories in KSH97-P correspond to two or more similar categories in ICD-10. Some categories in ICD-10 which are less frequently used in primary health care have been merged with related unspecified categories in ICD-10 to corresponding categories with broader coverage in KSH97-P 31. Rubrics in KSH97-P are as close to the Swedish translation of ICD-10 as possible 31.

Examples of KSH97-P categories and corresponding ICD-10 categories are 31

• KSH97-P category A00- Cholera corresponds to the ICD-10 category A00 Cholera.

• KSH97-P category J45-P Asthma corresponds to the two ICD-10 categories J45 Asthma and J46 Status asthmaticus.

• KSH97-P category H669P Otitis media, unspecified corresponds to the ICD-10 category H66.4 Suppurative otitis media, unspecified, which is more specific than the category in KSH97-P. H669P also corresponds to the ICD-10 category H66.9 Otitis media, unspecified, which is equally specific as the category in KSH97-P.

KSH97-P has the same chapter division as ICD-10. The exceptions are that ICD-10 chapter XX External causes of morbidity and mortality is left out from KSH97-P 31 and chapter XXII Codes for special purposes is left out in both the Swedish version of ICD-10 33 and KSH97-P 31. The rubric and number of categories in each chapter are included in Table 1.

KSH97-P mixes categories related to ICD-10 categories in both three and four-character levels. Therefore, the National Board of Health and Welfare recommends to only compile statistics on the chapter level or to use customised groups of categories 31.

As described above, ICD-10, and thus KSH97-P 31, uses multiple principles for chapter division 7. Some chapters contain categories related to a specific organ system and other chapters contain diseases with some specific aetiology. There are also chapters containing categories related to pregnancy, childbirth and the puerperium; the perinatal period; symptoms and partially specified cases; and important factors for contact with the health care system 7. The preface to KSH97-P states these different kinds of chapter divisions may imply practical problems because it is not evident to which chapter a specific disease or condition belongs 31.

In ICD-10, and thus KSH97-P 31, a category can be only included in one chapter 7. For those categories in which it would be possible to include more than one chapter, a decision has been made about into which chapter to include the category. This is demonstrated in ICD-10 by the excludes remarks on the chapter level. An excludes remark means that the categories in the remark could have been included in the chapter, but are instead included in other specified chapters 7. Table 2 summarises the excludes remarks on the chapter level for three-character level exclusions 7. The excludes remarks for four-character level exclusions on the chapter level are omitted because they only contain six categories 7.

To transform KSH97-P from a first generation system to a second generation system, a three-dimensional additional structure was added to KSH97-P in a previous research project 34. In the three-dimensional structure, each category was categorised according to location, origin and type 25.

A baseline category mapping from KSH97-P's categories to SNOMED CT's concepts is used. The first phase of the mapping process is described in a reliability study where mapping was done by two coders 35. KSH97-P was randomly divided into three sets of categories, used in three mapping sequences. Mapping was done independently by the coders and mapping rules were developed and agreed upon between the sequences. In the last round, mapping was completed through consensus decisions, following the mapping rules and striving to achieve a result with "completely concordant" mappings for each category. In the mapping, disorder and finding concepts were given priority and there was no use of navigational concepts 35. The version used was the releases of SNOMED CT from January and July 2006. In summary, 14 (1%) of the 972 categories in KSH97-P did not have a matched concept in SNOMED CT, 888 (91%) were mapped to one concept, 64 (7%) were mapped to two concepts, and 6 (1%) were mapped to three concepts. Of the 958 mapped categories, 938 (98%) categories were mapped to clinical finding concepts and 20 (2%) categories were mapped to procedural concepts.

Examples of baseline category mappings are

• KSH97-P category A00- Cholera is mapped to the SNOMED CT clinical finding concept Cholera.

• KSH97-P category R252 Cramp and spasm is mapped to the SNOMED CT clinical finding concept Cramp and the clinical finding concept Spasm.

• KSH97-P category D38- Neoplasm of uncertain or unknown behaviour of middle ear and respiratory and intrathoracic organs is mapped to the SNOMED CT clinical finding concept Neoplasm of intrathoracic organs and the clinical finding concept Neoplasm of middle ear and the clinical finding concept Neoplasm of respiratory tract.

• KSH97-P category Z000 General medical examination is mapped to the SNOMED CT procedure concept General examination of patient.

A flow chart of the used methods is presented in Figure 1.

Our study also needs a mapping from KSH97-P's chapters to SNOMED CT's concepts. The initial chapter mapping is therefore constructed during this study by the same persons mentioned above (Vikström et al. 35).

The chapters are mapped to SNOMED CT's concepts based on the meaning of the chapter's rubric and a general assessment of both the chapter's content in ICD-10, using the international WHO-version of ICD-10 7, and the subset of categories present in each chapter in KSH97-P. The same rules used for the category mapping and the excludes remarks in ICD-10 are considered as rules that do not exist in SNOMED CT. An example of an excludes remark is certain localized infections that should not be included in chapter I Certain infectious and parasitic diseases in ICD-10. Adequate mapping demanded good concordance between the rubric's meaning and the concept's meaning in SNOMED CT. For example, Neoplastic disease is considered a good match for chapter II Neoplasms. A concept could be considered as a reasonable match although it does not have relations to all categories from a certain chapter or it has relations to some categories from another chapter. An example is Obesity that is in chapter IV of ICD-10 but is not related to any of the mapped concepts in SNOMED CT as it is located directly under the Disease concept. The mapping is made to the SNOMED CT release January 2007.

Chapters XVIII Symptoms, signs and abnormal clinical and laboratory findings, not elsewhere classified and XXI Factors influencing health status and contact with health services are assessed as unable to map to SNOMED CT's concepts. The combination of different symptoms and abnormal clinical and laboratory findings in chapter XVIII's rubric are not considered to be a clinical concept, but a collection of different phenomena in a rubric that do not map to a concept or post-coordinated expression of manageable size in SNOMED CT. Not elsewhere classified is generally difficult to map to SNOMED CT, because the negation entails that the meaning of all other KSH97-P chapters has to be excluded in the mapped SNOMED CT concepts. Chapter XXI's rubric is likewise considered difficult to interpret as a clinical concept that could be mapped to a concept or post-coordinated expression of manageable size in SNOMED CT. This is especially due to the combination with the many categories in the chapter that describe procedures, not conditions and factors.

In summary, 2 (10%) of the 20 chapters in KSH97-P did not have a matched concept in SNOMED CT, 14 (70%) were mapped to one concept, 1 (5%) was mapped to two concepts, 2 (10%) were mapped to three concepts, and 1 (5%) was mapped to four concepts.

Examples of initial chapter mappings are

• KSH97-P chapter I Certain infectious and parasitic diseases is mapped to the SNOMED CT concept Infectious disease.

• KSH97-P chapter XIII Diseases of the musculoskeletal system and connective tissue is mapped to the SNOMED CT concept Disorder of musculoskeletal system and the concept Disorder of connective tissue.

To improve the baseline category mapping and initial chapter mapping, the mappings are converted to the same SNOMED CT release, the category mappings and the chapter mappings are compared manually and statistical chapter mappings are calculated and compared with the manual mappings. All these steps are described below.

The final mappings, which are the results of the mapping updates described above, are used in the rest of this study. The final category mapping is included as Additional file 1 and the final chapter mapping is included as Additional file 2. A summary of KSH97-P and the final mappings is included in Table 1.

To examine whether the poly-hierarchical Is a relationships of SNOMED CT can be used to replace KSH97-P's mono-hierarchical chapter division with a poly-hierarchical chapter division, KSH97-P's categories are divided into a multiple chapter division using SNOMED CT's Is a relationships. The multiple chapter division is generated using the algorithm described and exemplified below.

For each category, the mapped concepts are extracted together with their ancestors to a mapped set. This creates one mapped set for each mapped KSH97-P concept. If one or more chapters are mapped to any of the concepts in the mapped set, the category related to the mapped set is assumed to belong to these chapter(s)--regardless of what chapter they originally belong to. This means that each category may belong to zero, one or more new chapter(s).

In the example below, the multiple chapter division algorithm is applied to the category A00- Cholera. The algorithm is illustrated in Figure 2.

The algorithm begins by creating a mapped set from an empty set. First, the category A00- Cholera, which is shown as a red ellipse in Figure 2, and its mapping are used to locate the concept(s) the category is mapped to. The algorithm finds that the category A00- Cholera is mapped to the concept Cholera and the concept Cholera is therefore added to the mapped set. All ancestors to the concept Cholera are then also added into the mapped set. The resulting mapped set consists of the concepts shown in black rectangles in Figure 2.

The algorithm then uses the mapped set to evaluate if any chapter(s) maps to any concept(s) in the mapped set. The algorithm finds that chapter I Certain infectious and parasitic diseases maps to the concept Infectious disease in the mapped set, and chapter XI Diseases of the digestive system maps to the concept Disorder of digestive system in the mapped set. The category A00- Cholera is therefore assumed to belong to the chapters I Certain infectious and parasitic diseases and XI Diseases of the digestive system according to the multiple chapter division. These chapters are shown as green shaded ellipses in Figure 2.

To examine whether the defining attribute relationships of SNOMED CT can extend KSH97-P categories with attributes, a list of additional attributes is created. The additional attributes are generated using the algorithm described and exemplified below.

For each category the mapped concepts are extracted together with their ancestors to a mapped set. This creates one mapped set for each mapped KSH97-P concept. (The mapped sets are created in the same way as for the multiple chapter division.) Then all defining attribute relationships from concepts in the mapped set are followed and the target concepts are included in a specific attribute value set for each relationship type. In each attribute value set, the concepts that are ancestors of another concept in the same attribute value set are removed. The remaining concepts in each attribute value set constitute attribute values of the respective attribute types for that category.

In the example below, the additional attributes algorithm is applied to the category A00- Cholera. The algorithm is illustrated in Figure 2.

The algorithm begins by creating a mapped set from an empty set. First, the category A00- Cholera, which is shown as a red ellipse in Figure 2, and its mapping are used to locate the concept(s) the category is mapped to. The algorithm finds that the category A00- Cholera is mapped to the concept Cholera and the concept Cholera is therefore added to the mapped set. All ancestors to the concept Cholera are then also added into the mapped set. The resulting mapped set consists of the concepts shown in black rectangles in Figure 2. (This step is the same as in the multiple chapter division algorithm.)

The additional attributes algorithm uses the mapped set to follow each defining attribute relationship and include all the target concepts in different attribute value set according to which attribute type they are related. The referred concepts shown as blue shaded rectangles in the left of Figure 2 are included in the value set of type Causative agent. The referred concepts shown as blue shaded rectangles in the upper right area of Figure 2 are included in the value set of type Finding site. The referred concept Transudate shown as a blue shaded rectangle in the lower right area of Figure 2 is included in the value set of type Associated morphology. Then, in each value set, the concepts that are supertypes of other concepts in the same value set are removed from the value set. In the value set of type Causative agent, only the concept Vibrio cholerae is left, in the value set of type Finding site only the concept Intestinal structure remains and the value set of type Associated morphology only contains one concept so that value set is unchanged. The category A00- Cholera then is assumed to have an attribute of type Causative agent with value Vibrio cholera, an attribute of type Finding site with value Intestinal structure and an attribute of type Associated morphology with value Transudate.

Even if many categories have attributes of a specific attribute type, the usefulness of these attributes can be of limited value if many attributes share the same attribute value. For example, it is of limited use to know that most categories have attributes of the attribute type Finding site with the attribute value Body structure. We measure the distribution of the attribute values as the proportion of categories that relate attributes of the same attribute type to the same attribute value.

The quality of the multiple chapter division and additional attributes is dependent on how completely modelled the concepts that are mapped from KSH97-P's categories and these concept's ancestors are. (These concepts are the concepts in the mapped sets.) The mapped concepts and their ancestors are therefore extracted and the number of fully defined concepts and primitive concepts are counted. The numbers of outgoing defining relationships from fully defined and primitive concepts are also counted. The proportion of fully defined concepts in SNOMED CT in total are also counted.

Examples of fully defined concepts and primitive concepts are

• The concept Digestive system finding is a fully defined concept and is therefore fully defined by its defining relationships' types and targets listed below

○ Finding site; Structure of digestive system

○ Is a; Finding by site

• The concept Accidental poisoning is a primitive concept and is therefore not fully defined by its defining relationships' types and targets listed below

○ Is a; Poisoning

• The concept Cholera is a fully defined concept and is therefore fully defined by its defining relationships' types and targets listed below

○ Associated morphology; Transudate

○ Causative agent; Vibrio cholerae

○ Finding site; Intestinal structure

○ Is a; Infection due to Vibrio

○ Is a; Intestinal infectious disease due to Gram-negative bacteria

The computational methods described above are performed in a relational database management system (PostgreSQL). SNOMED CT, KSH97-P and the mappings are stored in tables and the computations are executed by SQL queries.

The multiple chapter division is summarised in Table 3, which shows the proportions of categories from each KSH97-P chapter that are divided into new chapters. Some categories are excluded from the multiple chapter division, which can be seen in Table 1. One reason is the lack of mappings from categories and chapters to SNOMED CT, which also can be seen in Table 1. Table 4 shows some examples of how categories are divided in multiple chapters.

Table 5 shows the proportion of categories in each KSH97-P chapter and the total that have additional attributes of different attribute types. Table 6 shows the average number of categories that have the same attribute value for each attribute type. The corresponding median values are also calculated in the study and are in most cases close to the average values. Table 7 demonstrates some examples of categories and their additional attributes.

Among the concepts that are mapped from KSH97-P's categories and these concepts' ancestors, 1,786 (63%) concepts are fully defined and 1,061 (37%) are primitive. There are a total of 10,010 outgoing defining relationships from the concepts and their ancestors. From all concepts in SNOMED CT, 13% are fully defined.

The fact that 37% of the used SNOMED CT concepts are primitive has most likely decreased the number of additional attributes and maybe also decreased the number of multiple chapters, but it is difficult to estimate how much. It does not mean that 37% of the attributes and chapters are missing because even primitive concepts have defining relationships. The effects are also dependant on where the primitive concepts are placed in SNOMED CT's structure. The only thing the numbers reveal is at least 1,061 defining relationships should be added to the existing 10,010 defining relationships used to make the concepts fully defined. This is because all primitive concepts lack at least one defining relationship to be fully defined and the number of used primitive concepts is 1,061.

It is beneficial that 63% of the used concepts are fully defined when the fraction of fully defined concepts of the entire SNOMED CT is 13%. However, the used concepts are central for Swedish primary health care and probably in primary health care in many other countries as well. It would be interesting to explore the possibilities to increase the number of fully defined concepts in this subset.

According to the meta classification by Rossi Mori et al. 1, KSH97-P is a first generation system, but after the additions of the multiple chapter division and additional attributes, it is possible to compare KSH97-P with a second generation system, at least with respect to structure. The categories in KSH97-P constitute a list of categories, and the additional attributes and their structure in SNOMED CT can be seen as a cross-thesaurus. The relationships between KSH97-P categories and their attributes correspond to a knowledge base of dissections. SNOMED CT's concept model together with information about how the mappings are done can be seen as a categorical structure.

One major improvement of a second generation system in comparison with a first generation system is, according to Rossi Mori et al., the possibility of automatically rearranging categories to satisfy different purposes 1. This study shows that KSH97-P's categories can be rearranged using the multiple chapter division and prioritisation to include concepts in other chapters than they normally are. It is also highly possible to rearrange the categories using the additional attributes according to Finding site and to a lesser extent according to Associated morphology. Even more reorganisations would probably have been possible if a higher proportion of the used concepts in SNOMED CT were fully defined.

Cimino 2 advocates similar characteristics as Rossi Mori et al. 1. Four desired characteristics according to Cimino are poly-hierarchy, formal definitions, multiple granularity and multiple consistent views 2. SNOMED CT fulfills these criteria, but none of them is fulfilled by KSH97-P. KSH97-P's "hierarchy" of chapters and categories does not fulfil what Cimino 2 terms multiple granularity; the reason is that only categories are meant to be used for coding. In ICD-10, on the other hand, the possibility of coding on three, four or in some parts five-character level would qualify as multiple granularity. However, this study shows that it is possible to use SNOMED CT to enrich KSH97-P in such a way that it fulfills Cimino's characteristics of poly-hierarchy, formal definitions, multiple granularity and multiple consistent views.

The formal definitions of SNOMED CT concepts make it possible to automatically perform the rearrangements of the categories in this study. The mappings from KSH97-P categories to SNOMED CT concepts, together with SNOMED CT concepts' formal definitions may also, to some degree, be seen as formal definitions of KSH97-P categories. SNOMED CT poly-hierarchy is used for constructing the multiple chapter division. The multiple chapter division also turns KSH97-P into a poly-hierarchical terminology system. SNOMED CT multiple granularity makes it possible to map both KSH97-P's chapters and categories to SNOMED CT concepts, and it also makes new multi-granular aggregations of KSH97-P's categories possible. SNOMED CT's multiple consistent views are used for generating both the multiple chapter division and the additional attributes. The possibility of showing KSH97-P categories according to SNOMED CT structure together with the multiple chapter division or different attributes in the additional attributes makes it possible to show KSH97-P in multiple consistent views.

A result of this study is that it is possible to keep the mono-hierarchical medical terminology system KSH97-P as it is and use it as an aggregation terminology, but at the same time enrich it with a more complete model of health problems. This can be an integrated way for medical terminology systems with different purposes and characteristics to co-exist, which Straub et al. speak in favour of 6. However, this study also shows that since mono-hierarchical medical terminology systems give a limited view of the real world, a better solution in the long run is to use terminology systems with richer structures, but with integrated aggregation structures for statistical calculations.

The methods presented in this study can be used to group KSH97-P categories together in groups according to different criteria than KSH97-P's and ICD-10's inherited design considerations' states. In other words, KSH97-P categories can be seen in multiple consistent views.

For example, according to the design considerations, KSH97-P's original chapter I Certain infectious and parasitic diseases only contains certain, and not all, infectious and parasitic diseases. However, using the multiple chapter division it is possible to include 17 categories that are not included in the original chapter in the group. Examples of these categories are J22-P Acute lower respiratory infection, L01- Impetigo and O86-P Puerperal infections. The method also proposes to exclude 5 categories from the original chapter in the group. Examples of these categories are A09-P Diarrhoea and gastroenteritis of presumed infectious origin and B94-P Sequelae of other specified infectious and parasitic diseases.

The resulting grouping of infectious and parasitic diseases categories is also an example of a possibility to change the chapters' priorities, given by the chapters' excludes remarks, in KSH97-P's mono-hierarchical design. The highest prioritised chapters include all relevant categories for that chapter, but lower prioritised chapters have only the relevant categories not included in a higher prioritised chapter. Collecting all categories related to infectious and parasitic diseases therefore gives the chapter the highest priority.

In the same way KSH97-P's original chapter XI Diseases of the digestive system only contains some and not all digestive system diseases. However, using the multiple attributes and including all categories that have an attribute type of Finding site with an attribute value of any subtype of Structure of digestive system it is possible to include 65 categories that are not included in the original chapter in the group. Examples of these categories are A00- Cholera, C16- Malignant neoplasm of stomach and R13- Dysphagia. The method also proposes to exclude 7 categories from the original chapter in the group. Examples of these categories are K42- Umbilical hernia and K65- Peritonitis.

It is also possible to use the methods described above to make groups according to other criteria than emulating an original chapter in KSH97-P. For example, suitable categories to include in a group related to the upper respiratory tract can be found by collecting all categories that have an attribute type of Finding site with an attribute value of any subtype of Upper respiratory tract or the concept itself. 26 categories fulfil this criterion and examples are C32- Malignant neoplasm of larynx, J32- Chronic sinusitis and J342 Deviated nasal septum. A similar method can be used to find categories related to the lower respiratory tract and 16 categories fulfil this criterion. Examples are C34-P Malignant neoplasm of trachea, bronchus and lung, J18-P Pneumonia and J42-P Chronic bronchitis. This can be seen as introducing multiple granularities into KSH97-P.

One way of looking at the difference between second and third generation terminologies is the discrepancy between mapping and modelling, that is, a matter of descriptions and definitions. Information reuse, which is the underlying topic of this paper, is always restrained by the limitations of the organising principles behind the vocabulary in which it was originally stored. Semantically interoperable information systems, e.g. clinical decision support tools connected to electronic health records, require a higher degree of the above: modelling in the sense of definitions, but a reasonable assumption is that a great deal of the vocabulary underlying all currently available information will never be modelled in this fashion.

Instead of using mapping methods in this study it would have been possible to extend KSH97-P with multiple chapters and additional attributes using entirely manual modelling as in the ICD-10 projects mentioned in the Background. It is possible the results may have been of higher quality than that of the mapped ones, but the modelling would have required more resources to perform than the mapping. The difference in used resources would have been even higher if larger medical terminology systems, like ICD-10, were extended with multiple chapters and additional attributes. Furthermore, as opposed to providing describing attributes, the definition of ICD categories seems to be troubled with difficulties regarding pragmatic, implicit and even unclear semantics. Depending on the purpose and the intended use and reuse of coded information, we believe enrichment through mapping might be an alternative to reconstruction when it comes to meeting requirements on new versions of classifications such as ICD and DSM.

Besides relying on descriptions rather than definitions, mapping imposes an extra layer of uncertainty and source of error. In this study, findings such that some categories did not make their way to their original chapters and that not all categories received additional attributes as expected could be explained by

• KSH97-P (source) issues such as counter-intuitive and overly pragmatic chapter inclusion

• Mapping problems (which are bound to occur when diverse contexts meet)

• SNOMED CT (target) problems such as incomplete or even faulty relationships

The algorithm for the creation of statistical chapter mapping was designed during this study. The premise for the algorithm is to prefer chapter mappings to concepts where the concept's descendants are mapping targets of many categories in the same chapter but few categories from other chapters. However, due to the design of KSH97-P, represented by the inherited excludes remarks from ICD-10, some exceptions from this principle must be allowed.

During the algorithm design, different alternatives were considered for the formula that transforms the variables categories current chapter (c) and categories other chapters (o) to a ranking score in line with the design assumptions. The initial considerations result in two main alternatives. The alternatives are the unweighted formula and the weighted formula .

The unweighted formula gives by definition the score 1, the maximum, for all concepts where o = 0 and a score < 1 for all concepts where o > 0. This characteristic interferes with the design assumption that some categories from other chapters must be tolerated. A test also shows that the formula gives a maximum score to many categories on lower levels in the hierarchy that subsume only one of a few category mapping targets and lower scores to categories on higher levels in the hierarchy that subsume many category mapping targets from a chapter. These characteristics make the unweighted formula unsuitable for the statistical chapter mapping algorithm and the weighted formula is therefore selected.

The results in this study are dependent on the quality of the mappings from KSH97-P to SNOMED CT and the quality of the modelling of the relationships in SNOMED CT. However, actions have been taken to use mappings of good quality. The baseline category mapping and initial chapter mapping are constructed using quality-assurance methods described by Vikström et al. 35. The study described in this paper also starts with two mapping validation methods where a manual comparison of category and chapter mappings and a comparison of manual and statistical chapter mappings are performed.

A limitation in the manual comparison of category and chapter mappings is the focus on minimising the number of categories that will not belong to any chapter in the multiple chapter division. Other incorrectly mapped categories could therefore have been disregarded.

The methods used during comparison of manual and statistical chapter mappings for finding potential incorrect mappings have one component trying to minimise the number of categories included in other chapters than the original KSH97-P chapter. Thus there is a risk the updates during this comparison remove correct additional chapters in the multiple chapter division. However, the method is only used for finding potentially incorrect mappings and all potentially incorrect mappings are manually inspected before any changes are made to the mappings. The risk of removing correct additional chapters in the multiple chapter division is therefore low. Another limitation is that the statistical chapter mapping's scoring formula was designed during this study and are not well established in this research area. However, the scores for a possible mapping are only used for comparisons with scores for other chapter mappings and the results of the comparisons are only used as a base for the manual inspection.

Besides continuing this study with a more detailed look at the semantics on category level, an interesting extension is to analyse the complete ICD-10 and not only a primary health care version. The reason why ICD-10 is not analysed in this study is the lack of mappings from ICD-10 to SNOMED CT, but this may change in the future. The result of a study analysing ICD-10 using these methods could be useful input to the work of revising ICD-10 into ICD-11. For example, the study may provide suggestions of possibly new structures in ICD-11. It would also be interesting to analyse other medical terminology systems using the same methods as in this study.

The mapping provided by the National Health Service in the United Kingdom is not suitable for this kind of study, because it is only a mapping from SNOMED CT to ICD-10 39. The focus of that mapping is thereby to correctly map as many SNOMED CT concepts, independent of the concepts' granularity, as possible to ICD-10 categories instead of mapping each ICD-10 category to one or a few SNOMED CT concepts with the same granularity.