In this study, we integrated 89 items of terminology using RDF, and identified inconsistencies in the hierarchical structure, the relationship between synonyms and preferred terms, and the definition statements using SPARQL. About 60% of the total number of duplicated terms were found. This is because industry associations that created multiple terminologies adopted the same terms in terminologies of similar medical device groups. In the case that all terms with the same spelling have the same concept, efficient integration can be achieved automatically using RDF. Furthermore, we evaluated six matters of inconsistency in this study, terms that need to be reviewed terms accounted for about 10% or less than 10% in each item. It may take a lot of effort to detect these from thousands of words if it is done manually. Since SPARQL can do this automatically, it has the advantage of reducing the physical and time burden.
Inconsistency of two-layer hierarchies
In inconsistency hierarchical structure, 1457 patterns were found among medical device problems and 630 patterns among patient problems using SPARQL. Inconsistencies of two-level hierarchies can be approximately divided into four kinds: (1) category terms which indicate the schema and preferred terms which indicate the details such as the relationships between “faulty device” and “alarm abnormality” (Table 2), (2) the relationship between a category term and a preferred term as a cause-effect relationship, such as the relationships between “battery problem” and “early discharge” (Table 2) among medical device problems, and between “respiratory insufficiency” and “falling arterial oxygen saturation degree,” (3) the hierarchy levels of category terms and preferred terms are reversed (inverted) depending on the terminologies [For example, in one terminology “trauma” is listed as a category term (hypernym) and “injury” as a preferred term (hyponym), but in another terminology they are reversed], and (4) the same words are listed as category terms and preferred terms, such as for “erroneous puncturing” and “uncertain”.
In terminology development, the relationship between hypernym and hyponym is “is-a relation.” The relationship is reflexive and transitive, but not symmetric [17]. (1) is applicable to this rule. The reason why more than three levels are structured is that the granularity of the terms differs depending on industry associations. It is necessary to consider unifying the granularity among the industry associations. In (2), depending on the medical device manufacturer, it may be easier for users to describe the relationship between the category term and preferred term as a causal relationship rather than an inclusive relationship. However, when integrating multiple terminologies, if the inclusion and causal relationships are mixed, the preferred term belonging to the category term becomes inconsistent, which may cause difficulties for the user searching for a term. Accurate aggregation may also be hindered. It is necessary to request the industry association that created the term in which the relationships between the category terms and the preferred terms have a causal relationship to reconfirm the hierarchical structures they are basing it on, or to correct the term using a tool that classifies mechanically. In (3), the inversion of hypernyms and hyponyms means that they are homonymy. If it is correct, it is desirable to consider unifying the notation, and if not, consider unifying the order of hypernyms and hyponyms, or describing another notation. In (4), when developing the terminology using the bottom up method, it is considered that the category terms and the preferred terms became the same because there may have been no appropriate category terms. In addition, these may be autohyponyms [18]. Autohyponym indicates that the hyponym is a subset of the hypernym. In this case, unless exploring a different more appropriate notation or considering managing terms by ID number, there will be discrepancies in the hierarchical structure because the same notation is regarded as the same term by RDF.
Polyhierarchy
Some terminologies have adopted polyhierarchy. In SNOMED-CT, a subtype hierarchy is a directed acyclic graph [19]. Cimino described that general consensus seems to favor allowing multiple hierarchies to coexist in a vocabulary and one could be so designated with the others treated as nonhierarchical with directed and acyclic relationships [20]. It would be possible if JFMDA also became a valid directed acyclic graph when integrated consistently. However, some parts of this integrated terminology were cycle graph due to the inconsistency of the hierarchical structure and the cycle graph between two words as shown in (7).
One of the features of the preferred terms having a number of category terms among medical device problems expressed the cause of the category terms. For example, “battery defect” as the preferred term expressed the cause of “charging defect,” “malfunction,” and “defect” (Table 4). Among patient problems, the tendency of the relationship between category terms and the preferred terms is a pattern of cause-effect relation, such as “trauma” in the category term and “bone fracture” in the preferred term (Table 4). The relationship between category terms and the preferred terms is the opposite compared with that of medical device problems. Since there is a possibility that two terms that each industry association considers to be related are set as hypernyms and hyponyms, it will be necessary to request the industry association to make corrections. Although the JFMDA terminology has been developed based on a monohierarchy, it is necessary to consider allowing a multi-layered structure with directed acyclic graph while accepting differences in industry ideas.
Preferred term and synonym
There were 32 terms among medical device problems which appeared in both the category term and preferred term, and there were 26 terms in patient problem. The following three kinds were found: (1) according to the terminology, preferred terms and synonyms are opposite, such as “exfoliation” (剥離) and “peeling” (はがれ), (2) preferred terms and synonyms are connected by causal relationships, such as “deformation” (変形) and “cratering” (へこみ), and (3) preferred terms and synonyms seem to be connected by “is-a relation,” such as “aberrance” (迷入) and “subretinal migration” (網膜下迷入).
In (1), it should be unified to one or the other. In (2), “cratering”, “fold”, and “bend,” which are synonyms associated with “deformation,” and these terms mean the causes of the “deformation.” The term “deformation” includes various concepts from the views of different industry associations, therefore it may be interpreted differently depending on the industry association. It is preferred that spellings which describe the details are used as preferred terms and the others are adopted synonyms. “Deformation” should not be used as the preferred name, and “cratering”, “fold”, and “bend” are preferred as preferred terms instead. “Collapse,” “breakage,” and “curvature” should be used as synonyms. In (3), it is considered that the terms were used by omitting the part of spellings in some terminology fields. It is necessary to use the spellings expressed in detail as preferred terms in order to carry out accurate statistical analysis as above.
Definition statements
There were some patterns of definition statements. First of all, as patterns common to medical device problem and patient problem, whether there are punctuation marks or not. Secondly, various expressions were used to describe the same concept, such as “a state in which the ventricle of the heart trembles” and “convulsions” in “ventricular fibrillation.” Additionally, as to the specific pattern in medical device problem, some definitions included the cause of the adverse event and the others did not include it, such as “blown fuse.” As to the specific pattern in patient problem, one is whether the cause of the patient problem is included or not as well as the medical adverse event, another is the cause of the patient problem is different, such as “extended operation time,” the other is whether to include a countermeasure for patient problem or not, such as “iridocele.” Those who use the JFMDA terminology may be confused if multiple definitions are given to one term. Therefore, if plural definition statements express the same concept, they should unify the description. If not, constructors should use different terms.
Multilingual mapping
There are international terminologies for medical devices related to adverse events, especially IMDRF, a voluntary group of medical device regulators from around the world that has constructed the terminology to accelerate international medical device regulatory harmonization and convergence [2]. Multilingual terminology mapping is an important process of finding correspondences between terminologies in different languages to allow these to be mutually understandable.
In JFMDA terminology 1st edition, mapping between the preferred terms and CDRH-NCIt terms was conducted manually by the industry associations. The result of an inconsistency detection survey using SPARQL showed that there were 160 preferred terms in JFMDA which appear in two or more CDRH-NCIt terms in medical device problems, and there were 95 among patient problems. This may be due to differences in the interpretation of CDRH-NCIt terms among industry associations. However it is also possible that the Japanese concept and the English concept do not exactly match. In ontology, there are two main strategies for alignment: direct and indirect alignment. The direct alignment is translation-based and uses external resources to help with the translation, while the indirect alignment uses intermediary mapping between the source and target ontologies. In addition, mapping two ontologies can be an automated or manual process [21]. Manual mapping is still the most common choice, but it necessitates a large team of experts; it is time consuming and prone to errors. Meanwhile, automated methods use publicly available terminology resources, but the sources of these are largely incomplete outside of the English speaking world [21]. To improve inconsistencies, it is being considered to reduce the human and time resources required and ensure accuracy using the following process: translating the international (not available in Japanese) terminology into Japanese, performing automatic mapping using machine learning, and confirming the results manually.
Whether a term among medical device problems is included in patient problems
There were 8 terms which are present among both medical device problems and patient problems. These terms share the concept, but depending on the situation where they are used, the subject can be the device or the patient. Therefore, this term was included among both medical device problems and patient problems. However, if the same notation is used, there is a problem that the hierarchies of terms in medical device problem, and those in patient problem are exchanged in mapping by RDF. Therefore, it is necessary to change the notation such as “damage (medical device problem)” and “damage (patient problem)” to distinguish both.