Building a Graph Representation of LOINC^® to Facilitate its Alignment to French Terminologies

Abstract

The aim of our study was to create a graph model for the description of LOINC^® concepts. The main objective of the constructed structure is to facilitate the alignment of French local terminologies to LOINC. The process consisted of automatically incorporating the naming rules of LOINC labels, based on punctuation. We implemented these rules and applied them to the French variants of LOINC and then created attributes and concepts described with synonymous labels. When comparing the created attributes to the stated ones, the multiple mappings led to the identification of errors that must be corrected for improving the translation quality. These mappings are consecutive to semantic errors generated during the translation process. They mainly corresponded to misinterpretations of LOINC concepts and/or LOINC attributes.

Introduction

Reference terminologies are controlled vocabularies whose terms and structures have been validated by the scientific community^1,2. LOINC^® (Logical Observation Identifiers Names and Codes) is an example of such reference terminology that is used in many countries for recording laboratory observations^3,4. Indeed, LOINC^® imposed itself as a reference through the quality and the novelty of its content. Consequently, many works have been focused on aligning local terminologies to LOINC^5–8, positioning it as an international support terminology for sharing information across different health systems.

Like other international reference terminologies, such as MeSH^a and MedDRA^b, LOINC labels have to be translated⁹. LOINC provides guidelines and recommendations for the translation process of its labels, as it does not directly supply translations. The translation is realized by other institutions (e.g., Canada Health Infoway Inc. (Canada), Consiglio Nazionale delle Ricerche (Italy), LOINC Turkish Translation Group and the Turkish Ministry of Health (Turkey)). When the translation is performed by different institutions for a given language, LOINC provides different storage formats (i.e., comma-separated value tables in the release) as linguistic variants (e.g., the Canadian and Belgian variants of the French language). This particularity makes it possible to acquire synonyms in each specific language⁹.

Despite the richness of its content, the LOINC release does not contain a formal representation combining the different languages (and related variants). Moreover, to our knowledge, the LOINC release provides translations of the main core of LOINC but not of each definitional element. In order to facilitate the alignment of French local terminologies to LOINC, this work describes the construction of a graph structure of LOINC combining all its French variants. In addition to its practical usability, such a graph structure of LOINC may indirectly help to assess the quality of different (French) translations. Indeed, previous work clearly stated that the translation process of concepts from a terminology must not be limited to the translation of their terms¹⁰. This translation must associate a term in a specific language with a concept while maintaining its relations with the other concepts (and their terms) consistent through this language. The constructed structure may help verify the compliance of each translation to this requirement.

First, we present the LOINC structure and the translation principles of its labels that are stated in the literature. Secondly, we describe the method applied to construct and analyze the graph representation of LOINC. Finally, we expose and discuss our main results.

Background: the LOINC structure and translation principles

The LOINC structure.LOINC is a clinical terminology created and maintained by the Regenstrief Institute. It contains concepts for the description of laboratory tests, clinical observations, and radiology³. Each laboratory observation defined in LOINC corresponds to a unique identifier called “LOINC concept”. The LOINC concepts are defined using six major attributes and four minor attributes. In addition to these attributes, each LOINC concept is attached to a specific class that mainly corresponds to the analysis type of a lab test. In the LOINC users’ guide¹¹, it is indicated that classes are not definitional for LOINC concepts but that they are used for sorting purposes.

The major attributes are the following: (1) the Component that corresponds to the observed analyte, (2) the measured Property of this analyte, (3) the Time of the measurement that represents its frequency, (4) the System corresponding to the biological nature of the analyte or its origins, (5) the Scale that provides a precision of the measurement observation, and (6) the Method that corresponds to the technique used to obtain the observation results.

The minor attributes are composed of: (i) the Challenge that corresponds to the preliminary action carried out before the test, (ii) the Adjustment that describes the closing action after the test, (iii) the Super system that specifies the origin of the sample when it is not a patient, and (iv) the Time modifier that defines a threshold or step of the time measurements.

These characteristics of LOINC concepts can be accessed through their labels. Punctuation plays an important role in the structure conveyed by the labels of LOINC concepts11. The label containing all the required punctuation is called the fully specified name, which consists in the concatenation of the attributes of a given LOINC concept.

In practice, the major attributes are separated using the “:” character in the following order:

<component>:<property>:<time>:<system>:<scale>:<method>.

The caret character “^∧” is used to introduce a minor attribute (e.g., for the description of components, this character is used in the following order <analyte>^∧<challenge>^∧<adjustment>).

Other special characters are used to structure labels of LOINC concepts. These characters are listed in the Methods section along with the explanation about how they have been processed.

The LOINC translation principles.LOINC labels were originally available in English. In the literature, it is said that LOINC uses a “part-based translation principle” to automatically generate the fully specified name of LOINC concepts in other languages. “The atomic elements that make up each LOINC term name are called Parts”³.

LOINC parts thus mainly correspond to LOINC attributes to which an identifier has been assigned. For example, in the fully specified name of the LOINC concept 3665-7-Gentamicin^∧trough:MCnc:Pt:Ser/Plas:Qn, the atomic elements “Gentamicin” and “trough” are LOINC parts that are identified by the codes LP15747-6 and LP20176-1, respectively. Like Gentamicin and trough, all the delimited elements of the fully specified name (i.e., MCnc, Pt, Ser, Plas and Qn) are LOINC parts. Thus, LOINC editors recommend translating LOINC parts first and then using these translations to reconstruct the fully specified name of LOINC concepts¹².

Materials: the LOINC release

Our work is based on the LOINC 2.65 version^c. The release contains:

• the LOINC core table (LCT) which is the full version of LOINC containing labels in English,

• the French variant tables (FVT) which describe the main core of LOINC in French,

• the LOINC part table (LPT) which contains the relations of LOINC concepts to their attributes (called “LOINC Parts” in the release).

The 2.65 release contains four French language variants (French, Belgian, Canadian, and Swiss variants). Like in LCT, each line of FVT is composed of a LOINC concept identifier, the labels of the six major attributes (each attribute and the LOINC concept identifier being separated in distinct columns), and the class. The French variant contains 49,437 LOINC concepts, the Belgian variant 45,779 LOINC concepts, the Canadian variant 45,411 LOINC concepts and the Swiss variant 4,940 LOINC concepts. The Swiss variant has not been used in our process because only short names (i.e., labels without the punctuation structure) of LOINC concepts were available. By pooling the French, Belgian, and Canadian variants, the French version of LOINC finally contains 54,480 LOINC concepts, for which multiple labels may exist (Figure 1).

Figure 1:

Description of the LOINC concept 13505-3-Herpes simplex virus 1+2 Ab pattern [Interpretation] in serum in the pooled FVT. The three rows correspond respectively to the labels provided in the Belgian variant, the Canadian variant and the French variant.

LPT is only provided in the English language. More precisely, LPT contains LOINC concept identifiers and labels (short-form labels in English), the identifiers and labels of the related LOINC attributes, as well as the type of link existing between LOINC concepts and their attributes (e.g., COMPONENT when the relation holds between a LOINC concept and a component attribute). Each LOINC concept may be related to multiple LOINC attributes of the same type. The additional tags “primary” and “search” further specify the link between LOINC concepts and attributes. The explanatory note accompanying LPT gives the following two definitions for these tags:

• “Primary-the primary parts associated with a given LOINC term, including the six major parts”,

• “Search-parts that are only linked to a given term in order to facilitate efficient searching and location of that term”.

Thus, the “search” tag is mainly navigational whereas the “primary” tag is definitional. It is noteworthy that this does not exclude that a “primary” tag can be found between a LOINC concept and more than one LOINC attribute of the same type, as illustrated in Figure 2.

Figure 2:

Description of the LOINC concept 13505-3-Herpes simplex virus 1+2 Ab pattern [Interpretation] in serum in LPT.

Methods

Three steps have been followed to construct and analyze the French structure of LOINC: (i) the definition of the LOINC model, (ii) the instantiation of the model according to the structure of LOINC labels, and (iii) the comparison of the obtained structure with the LPT content.

The construction has been performed according to rules based on the punctuation present in the labels of LOINC concepts. The Simple Knowledge Organization System (SKOS^d) format has been chosen to describe LOINC because it allows the representation of multiple labels for a given identifier (concepts and attributes). In addition, we considered that the generalization hierarchy defined in SKOS was appropriate for representing the LOINC hierarchy that is based on the compositional structure of its labels. Indeed, SKOS has been proven to be well adapted for the description of the hierarchical structure of terminologies¹³, and the English version of LOINC has already been described using SKOS¹⁴.

Description of the proposed LOINC model. Figure 3 displays the proposed model for the description of LOINC concepts.

Figure 3:

Proposed model used for the construction of the LOINC graph structure

It consists in:

• The description of LOINC attributes: in the literature, LOINC attributes are described as major attributesor minor attributes. When used, minor attributes are parts of the description of major attributes. Thus, they correspond to optional sub-parts of major attributes. The sub-attributesdesignate the sub-parts of LOINC major attributes. The sub-attributes used for the description of major attributes that are not minor attributes are called main attributesin our model and in the rest of this article. For example, in the component (major attribute) “leukocytes^∧∧corrected for nucleated erythrocytes”, “corrected for nucleated erythrocytes” is an adjustment (i.e., a minor attribute), while the other part of this component (i.e., “leukocytes”) is the main attribute. Because this component does not involve any challenge, nothing is described between the two caret characters. We also added class as attribute in our model.

• The description of relations in the model: for each major attribute, a semantic link has been created between the LOINC concept and the attribute. The relation has been labelled using the prefix “has_” followed by the name of the major attribute (e.g., an has method relationship has been defined to associate LOINC concepts to their method attribute(s)). Between major attributes and minor attributes, the same strategy has been used (e.g., an has adjustment relationship has been defined to associate component attributes to their related adjustment attribute). As main attributes correspond to the attributes’ description without the refinement of minor attributes, a hierarchical relation (skos:broader) has been created between major attributes and main attributes.

  Each LOINC attribute and concept have been described as a skos:Concept.

Instantiation of the model. The instantiation has been realized by using data from FVT. The process started with the creation of a set of labels of each attribute followed by the generation of an identifier for each of them.

• The creation of sets of labels: from each fully specified name, sets of labels (by type of attributes) have been created from all the French variants using a tokenization process based on the punctuation of LOINC labels. The sets of labels corresponding to main attributes have been included in the sets of labels of major attributes. For example, a unique set of labels for components and analytes has been constituted. According to the punctuation character, the process was as follows:

  • – the caret character “^∧” delimits the minor attributes and the main attributes in the description of major attributes. Using this punctuation, the set of major attributes and the set of minor attributes have been created. For example, from the LOINC label of the component Insulin^∧1.5H post dose tolbutamide, “insulin” has been integrated in the set of components and “1.5h post dose tolbutamide” has been integrated in the set of challenges.
  • – the dot character “.” describes hierarchical relations between attributes. For each dot character in a label, an additional attribute corresponding to the left side of the dot character has been created. For example, from the component label “epithelial cells.renal”, the label “epithelial cells” has been created and included in the set of components like the original label.
  • – the slash character “/” describes quotient relations in the components. Like for the dot character, the left side of the slash character has also been extracted and included in the set of components.
  • – the “+” and “&” signs may be used to create combined measurements or combined results. The labels containing the “+” and/or “&” signs can thus be decomposed. The left side of the related characters has been identified as a prefix (an identifier has been created for that prefix) and the right side as a suffix. The attributes have thus been reconstituted by combining the prefix, each related character and the suffix. For example, from the component label human papilloma virus 16+18 Ag, the related elements are “16” and “18”, the prefix is “human papilloma virus” and the suffix is “Ag”. The following labels have thus been created and included in the set of components (in addition to the suffix and the component label): “human papilloma virus 16 ag” and “human papilloma virus 18 ag”.
  • – the “+” and “-” signs may be used to describe the cluster of differentiation (CD) of cells when they appear at the end of a label. In such cases, the “+” sign indicates the presence of a specific CD and “-” indicates its absence. Thus, the same rule as for the “+” and “&” signs has been applied to identify the composed attributes. For example, from the component label Cells.CD3+CD4+CD27-CD45RO+CD62L-, the following labels have been created and included in the set of components: “cells”, “cells cd3”, “cells cd4”, “cells cd27-”, “cells cd45ro” and “cells cd62l-”.

• The creation of a unique identifier for synonymous labels: as previously stated, the set of French labels cannot be related to a LOINC identifier because LOINC parts are not described in French. Thus, a unique code of attribute has been generated for each label. When a code was assigned to a label related to a LOINC concept, this code was also assigned to all the other labels that have the same relation with this LOINC concept in the different French variants. Indeed, for each LOINC concept related to an attribute, this attribute has been considered as equivalent across the different linguistic variants.

  For example, in FVT, the term “hémostase” is used in the French variant to designate the class attribute of the LOINC concept 3245-8-Clot Retraction [Time] in Blood by Coagulation assay while the term “coagulation” is used in the Canadian variant. Thus, the unique code (CLAS1508) created for “hémostase” has also been assigned to “coagulation”.

Comparison of the constructed structure with the stated structure of LOINC. To highlight the advantages of the constructed structure, we compared it to the stated structure of LOINC, which is commonly computed from LPT.

For the construction of the stated LOINC structure, we described each relation between LOINC concepts and attributes as a simple Resource Description Framework (RDF^e) triple. To avoid the ambiguity in LPT, we decided to integrate the relations between a LOINC concept and its LOINC attributes in the stated structure only if, for each type of attribute, these relations are the unique ones being tagged as “primary”.

For example, we added an has_class relation between the LOINC concept 13505-3-Herpes simplex virus 1+2 Ab pattern [interpretation] in serum and the class LP7819-8-Micro. However, we did not relate this LOINC concept to the two components LP14822-8-Herpes simplex virus 1+2 and LP40415-9-Herpes simplex virus 1+2 Ab pattern because they are both tagged as “primary”. Thus, the relation with the class has been added, whereas no relation has been created for the components (the relation with components being ambiguous in LPT).

The attributes created by our process and those from LPT were considered as equivalent if they shared the same LOINC concept identifier. Thus, we computed the cardinality of these mappings as follows:

• 1-0 mappings corresponded to one created attribute for which no stated attribute existed,

• 1-1 mappings associated one created attribute to one stated attribute,

• 1-N mappings associated one created attribute to more than one stated attribute.

Results

The constructed LOINC structure for the French language.LPT contains the description of 89,271 LOINC concepts corresponding to 44,313 component, 1,791 challenge, 35 adjustment, 205 property, 59 time, 8 time modifier, 2,682 system, 62 super system, 10 scale, 1,907 method, and 389 class attributes.

From the labels of the 54,480 concepts obtained by pooling the French, Belgian, and Canadian LOINC variants, the attributes for which French labels have been generated by our process corresponded to 22,819 component, 28,807 analyte, 819 challenge, 15 adjustment, 140 property, 31 time, 26 time aspect, 3 time modifier, 394 system, 368 main system, 16 super system, 6 scale, 754 method, and 103 class attributes.

The stated structure contained much more component, challenge, system, scale, method, and class attributes than the French version because it covers more concepts and it contains attributes that are not definitional. For example, the LOINC attribute LP7747-1-- (the dash being the label) used as a scale in the stated structure has been ignored during our construction process because the dash indicates the absence of any attribute.

Among the attributes our process defined in French, 3,140 components have been described with a challenge and/or an adjustment. From the hierarchy computed for components, the process created 28,807 analytes. An example of such created analytes from the component label of the LOINC concept 90229-6-Herpes simplex virus 1 and 2 Ab.IgG and IgM panel - Serum or Plasma is illustrated in Figure 4.

Figure 4:

The constructed hierarchy of analytes according to the punctuation in the LOINC component label COMP24270-herpes simplex virus 1 & 2 ab.igg & igm panel

Comparison of the constructed structure with the stated structure of LOINC.Table 1 describes the cardinality of mappings between the constructed and the stated LOINC structures.

Table 1:

Distribution of the constructed LOINC attributes according to the cardinality of their mappings to the stated attributes

LOINC attributes	1-0 mappings	1-1 mappings	1-N mappings	Total
Component (major)	9,710	13,018	91	22,819
Analyte (main)	28,807	0	0	28,807
Challenge (minor)	3	804	12	819
Adjustment (minor)	0	15	0	15
Property (major)	0	136	4	140
Time (major)	0	31	0	31
Time aspect (main)	26	0	0	26
Time modifier (minor)	0	3	0	3
System (major)	43	344	7	394
Main system (main)	368	0	0	368
Super system (minor)	0	15	1	16
Scale (major)	0	5	1	6
Method (major)	0	744	10	754
Class	0	100	3	103

Many 1-0 mappings have been found for the component attributes and a few mappings for the system and challenge attributes. This result is due to the disambiguation step applied when a LOINC concept was related to multiple attributes of the same type through a “primary” tag in LPT. For example, the LOINC concept 63309-9-Proteinase 3 Ab [Presence] in Body fluid by Immunoassay was described with the following two components (both through “primary” tags): LP17259-0-Proteinase 3 and LP39491-3-Proteinase 3 Ab. Thus, no mapping could be found because these relations have not been computed in the stated structure. On the contrary, our process described the concept with the component attribute COMP2120-Protéinase 3 anticorps, this component being itself related to the analyte AN61492-Protéinase 3 through a skos:broader relation. Regarding the main attributes (i.e., Analyte, Time aspect, and Main system), only 1-0 mappings were found because these entities are specific to the implemented model.

Predominantly, the created attributes have 1-1 mappings with the stated attributes. The 1-1 mappings corresponded to the non-ambiguous correspondences identified between our constructed structure and the stated one.

The mapping process generated a few 1-N mappings. For example, the scale attribute SCALE1503-quantitatif has been mapped to LP7753-9-Qn and to LP7751-3-Ord because the LOINC concept 3245-8-Clot Retraction [Time] in Blood by Coagulation assay was described with the scale attribute “qualitatif” in the French variant while the other variants used the scale “quantitatif”. This led to erroneously consider that “quantitatif” and “qualitatif” are synonymous labels and to the creation of a unique identifier for them. Another example is the mapping between the component attribute COMP4206-Sulopenem and the LOINC attributes LP94456-8-Linopristin+Flopristin and LP94455-0-Sulopenem because the LOINC concept 55289-3-Sulopenem [Susceptibility] has been erroneously described in the Belgian variant using “linopristin+flopristin” as a component while the other variants used appropriately “sulopenem”.

Discussion

Our work consisted of constructing a graph representation of LOINC to facilitate the alignment of local terminologies to LOINC. Local terminologies are characterized by the low expressiveness of their labels and structures. They mainly contain user-friendly terms that facilitate their usability1 but limit the application of morphosyntactic approaches to map these local terms to the complex terms of LOINC. Thanks to the proposed graph structure containing synonymous labels in French for concepts and attributes, the creation of mappings between LOINC and French local terminologies will be more efficient. Indeed, a mapping to a LOINC attribute (having a simpler label) can be performed first, and then more complex mappings can be inferred to find the appropriate LOINC concept. For example, a named-entity recognition strategy may be applied to the labels of local terminologies by looking for the component/analyte terms, method terms, class terms, etc. Then, according to the attribute terms found in the labels, a query (e.g., SPARQL query) may be made on the structure of LOINC to recover the LOINC concepts containing these characteristics.

We made this construction based on the labels of LOINC rather than on the structure described in the LOINC multi-axial hierarchies (available in the release) and LPT (used only for comparison) for four reasons. First, the multi-axial table of LOINC concepts is manually maintained and the hierarchy is not meant to describe LOINC as a pure ontology but according to the different domains of laboratory analyses. Secondly, the description of parts is ambiguous. As illustrated in Figure 2, multiple LOINC attributes of the same type may be used to describe a LOINC concept. In this example and for a formal description, LP40415-9-Herpes simplex virus 1+2 Ab pattern is the appropriate component but LOINC does not prioritize it in the description of the LOINC concept 13505-3-Herpes simplex virus 1+2 Ab pattern [interpretation] in Serum. Thirdly, in the multi-axial table, hierarchies are not available for each attribute. Again in the example of Figure 2, no hierarchical link exists between LP40415-9 and the other components. Especially between LP40415-9 and LP14822-8-Herpes simplex virus 1+2, a hierarchical relation would clearly be expected (not to say that the “antibodies of herpes virus” are a kind of “herpes virus” but rather to highlight that an analysis on “antibodies of herpes virus” is an analysis on the “herpes virus”). Finally, the identifiers of LOINC attributes were not related to their French labels in the release. The proposed process illustrates that a French version of LOINC can be automatically built with attributes containing synonymous labels from different variants. More interestingly, this process may be used to create a multilingual graph structure of LOINC including all the languages (with their variants) contained in the release.

Terminology translation is an important task allowing the usability of reference terminologies through different countries. The accuracy of the used terms must be maintained across different languages10, 15. Whether the translation is realized manually or automatically, it must be curated16, 17. When comparing the obtained structure with the stated structure of LOINC (only for LOINC concepts related to a unique attribute of each type), we noticed some inconsistencies due to translation errors. The highlighted limitations do not question the quality of the constructed structure but give the possibility to indirectly audit LOINC translations. Indeed, when comparing the created attributes to the stated ones, we found that many 1-N mappings were consecutive to semantic errors in the translation process, mainly due to misinterpretations of LOINC concepts and/or LOINC attributes. For example, in the description of the LOINC concept 3245-8-Clot Retraction [Time] in Blood by Coagulation assay, if the test of “coagulum retraction” consists in measuring the bleeding time, the result is sometimes given in practice according to the interpretation of this time (e.g., normal, prolonged) which is more clinically relevant. However, the LOINC concept does not describe this type of results. Using the “qualitatif” attribute to describe the scale attribute of this LOINC concept changes the purpose of this concept, which runs counter to the translation principles. Indeed, changing the purpose of LOINC concepts according to local requirements defeats the goal of LOINC to share the same knowledge through different health systems. These translation errors are not specific to a given variant. Indeed, the French, Canadian, and Belgian variants all contain errors. For example, the Canadian variant erroneously uses the term “Agglutinine froide” to describe “Cold agglutinin panel” as a component of the LOINC concept 79160-8-Cold agglutinin panel - Serum, while the French variant confusedly uses the plural “Agglutinines froides” to describe the component “cold agglutinin”. For now, these 1-N mappings have to be manually explored for determining which variant(s) contain the error(s). A perspective of this work is to automate this step by adding English labels. Indeed, the following simple steps may be computed for identifying the erroneous translation: 1) for each variant, collecting the set of LOINC concepts whose label contains the specific term corresponding to the attribute’s label, 2) calculating the similarity between the resulting sets of LOINC concepts, and 3) identifying the variant whose set of LOINC concepts is declared different from that of the English version (being a priori the one that contains the error).

Some limitations were highlighted within the proposed process, such as the misinterpretation of characters in labels. Indeed, for some components, the slash “/” character describes a quotient in the challenge or the adjustment but for other components, it describes something else. When splitting the labels, the expected result can be found by deter-mining a priority between the slash “/” and the caret ^∧ characters. For example, the caret character must be prioritized when processing the label of the component “streptococcus pneumoniae Danish serotype 9N Ab^∧1st specimen/2nd specimen” of the LOINC concept 86181-5-Streptococcus pneumoniae Danish serotype 9N Ab [Ratio] in Serum by Immunoassay – 1st specimen/2nd specimen. In this case, the slash delineates the quotient of the challenge and not that of the component. Conversely, the slash character must be prioritized when processing the label of the component “volume^∧at 1.0 s post forced inspiration/Volume.forced inspiration.total” of the LOINC concept 43264-1-FIV1/FIV Predicted because it delineates the quotient of the component and not the quotient of the challenge. On the other hand, the dot character “.” is sometimes a part of the analytes’ names and not an indication of a hierarchical relation. For example, in the label “t(X;11)(q13.1;q23)(FOXO4,MLL) fusion transcript”, the dot character is actually part of the genes’ names. These limitations were also emphasized by the multiple mappings found between the created attributes and the stated LOINC parts. However, it should be noted that they corresponded to a few number of attributes and could be manually corrected.

The constructed French structure was limited to the construction of RDF triples for describing LOINC concepts. A collaboration between the Regenstrief and SNOMED International aimed at providing mappings between the LOINC parts and LOINC concepts to concepts of the Systematized Nomenclature of Medicine and Clinical Terms (SNOMED CT^®f)¹⁸. If the mapping allowed to have a projection of the SNOMED CT structure on LOINC so that a formal structure of LOINC could be defined, it did not propose a disambiguation of LOINC attributes. A perspective of our work could be to use a formal definition and a more formal language like the Web Ontology Language (OWL^g) for expressing LOINC based on the disambiguation of its concepts that can be identified in our process. In this way, other characteristics of LOINC labels could be used (e.g., the structure of challenges) and the LOINC parts corresponding to a combination of attributes from different types (e.g., time and system, system and method) could be easily computed. Indeed, Mary et al. previously showed that a formal structure enhances the classification of laboratory tests¹⁹. As shown in Figure 2, it is certainly the absence of such a formal structure that led the Regenstrief to describe LOINC concepts with multiple attributes of the same type. Admittedly, LOINC concepts can be retrieved from general queries in a limited way. Nevertheless, these queries are more limited than those carried out on a formal structure. Furthermore, this representation introduces ambiguity in the description of concepts.

Conclusion

To facilitate the alignment to local French terminologies, our process aimed at constructing a graph representation of LOINC through the structure of its labels. We computed a structure of LOINC with attributes and concepts having multiple synonymous in French. We found that the majority of LOINC parts’ translation met the requirement that, in a translation, terms used for a given concept must be in accordance with the relation of this concept to other concepts. However, our process highlighted some semantic errors in the translation process that change the purpose of LOINC concepts and thus require to be corrected.