SNOMED CT is a comprehensive multilingual class hierarchy of medical terms used in clinical records. Few translations are available, but, as new concepts and revisions are continuously being added, the manual translation and revision of the terms will remain a major endeavour. We propose a new approach for translating SNOMED CT terms (or named entities) using ontology mapping methods and various existing multilingual resources with translated concepts. Our purpose is generating initial candidate translations, already close to those proposed by medical experts, to be later used in a curated translation process. Our method for automatically translating SNOMED CT is being developed for Portuguese, using DBPedia, ICD-9 and Google Translate as sources of candidate translations of the clinical terms, which could later be verified. Initial results, using a manually translated Portuguese catalog of allergies and adverse reactions (CPARA) to SNOMED CT as ground truth, show that it has high potential.
SNOMED Clinical Terms1, or SNOMED CT, is a comprehensive multilingual class hierarchy of terms used in clinical records, with extensive overlapping and synonymous descriptions. The primary purpose of SNOMED CT is to encode the meanings of the terminology used in health information, supporting the effective clinical recording of data with the aim of improving patient care. SNOMED CT provides the core general terminology for electronic health records. With about 300,000 active terms, SNOMED CT spans clinical findings, symptoms, diagnoses, procedures, body structures, organisms and other etiologies, substances, pharmaceuticals, devices and specimen.
The need to interchange medical records across states is demanding the development of faster methods to obtain approved, standards-based, translations of medical records, in particular SNOMED CT. The standardisation of clinical terms and their translations to other languages is very important for the unification of the electronic health records worldwide. However, the manual translation and revision of the terms, synonyms and definitions to a new language is a major endeavour. SNOMED CT is presently available in US and UK English, Spanish, Danish and Swedish. It is also being translated to several other languages, but there is no translation of SNOMED CT to Portuguese or an official initiative to develop and maintain that translation. Hence, a tool to automatically translate SNOMED CT to Portuguese would assist in the production of a release to be validated and improved in a subsequent step at a much lower cost than conducting the process manually.
As new trannslations, concepts and revisions are continuously being added, the manual translation and revision of the terms will remain a major endeavour. This paper describes our work on the development of an automatic translator of SNOMED CT to
Portuguese as an assisting tool that could be used for the production
of a future standard translation of SNOMED CT. We take the
approach of using available classifications and automatic translation
services as ontologies that can be aligned and later navigated to
provide the translations of such technical terms. In our method, we
start by identifying existing alignments between SNOMED CT and
other selected ontologies, including the releases of SNOMED CT
in different languages. For the Portuguese translation, given its
proximity to Spanish, many terms in the Spanish release of
SNOMED CT have almost identical spelling. There are other
medical terminologies for which multiple translations have been
published, such as ICD (International Classification of Diseases) 2.
Another major resource is DBPedia, an ontology derived from
Wikipedia, which is very rich in medical terms
To show the feasibility of the above outlined approach for
automatically generating translations of SNOMED CT terms to
Portuguese, we evaluated the translations obtained with the
alignments against the translations of a set of SNOMED CT
terms that have been mapped by medical experts to terms of
the Portuguese catalog of allergies and adverse reactions
In our work, we used the the 01/2014 International (English) distribution of SNOMED CT and the Spanish version dated from April 2014, both provided by the NLM (National Library of Medicine) institutional site 5. The distribution also includes a mapping between ICD-9, a WHO classification of diseases, and SNOMED CT. This mapping can be used to link SNOMED CT
translator-api.aspx
snomed_main.html codes to the ICD-9 Portuguese terms in a translation provided by the Portuguese Ministry of Health 6.
There are no comprehensive medical terminologies for European
Portuguese. In addition to ICD-9 in European Portuguese,
ICD10 has been manually translated to Brazilian Portuguese 7.
There is also an English to Brazilian Portuguese dictionary of
medical terms
In computing, the translation of a terminology, such as the
set of SNOMED CT terms, is an instance of a common task
in Natural Language Processing (NLP) , designated as Named
Entity Translation
Other approaches for generating translations have been attempted.
Algorithms based on linguistic rules are particularly useful for
languages which are poor in language resources, like a recently
proposed Basque semi-automatic translation of SNOMED CT
In this work, we explore DBPedia, an ontology derived from
Wikipedia, as an alternative source of term translations
ICD9CMOut2013.xlsx
reported in this paper only used the first layer algorithms to perform the alignments.
Medical terms, like named entities in general, can matched using
similarity metrics like the Jaro distance, initially proposed for record
linkage systems
Given that SNOMED CT is mostly used to provide terminology
for electronic health records, the risks of using an automatically
generated translation of such large collection of terms without
expert validation are unacceptable. In fact, the SNOMED publisher
provides detailed guidelines for validating the translations made
by medical experts for the official translations available
Our approach for generating the translations of SNOMED CT terms into Portuguese is illustrated in Figure 1. We start by organising two mappings: 1. SNOMED CT to ICD-9: a correspondence between the codes of SNOMED CT and codes and descriptions of ICD-9. 2. SNOMED CT to DBPEDIA: a correspondence between SNOMED CT codes and DBPedia (English and Portuguese) page URIs, and associated page titles.
The first mapping is derived the SNOMED CT to ICD-9 mapping included in the from the UMLS distribution, which includes the correspondence between SNOMED CT and ICD9 codes. For the second mapping, the matching algorithms implemented in AgreementMakerLight can generate an alignment between SNOMED CT terms and English DBPedia labels. Once this alignment is generated, we can map SNOMED CT codes to DBPedia URIs and then obtain the corresponding label for the Portuguese term by a simple lookup.
To obtain candidate translations for SNOMED CT terms, we implemented four translation methods: 1. Google Translate EN: the candidate translation into Portuguese of each English term in SNOMED CT is provided by the GoogleTranslate API service. 2. Google Translate ES: identical to the above, but the translation service uses the Spanish term as input. 3. ICD-9 Mapping: for a given SNOMED CT term in English, we take the corresponding code and lookup the SNOMED CT to ICD-9 mapping in the UMLS distribution to obtain the ICD-9 code and next the term description in the Portuguese version of ICD-9. This description becomes the candidate translation of the SNOMED CT term to Portuguese. 4. DBPedia Mapping: starting with a SNOMED CT term in English, we lookup the code on the SNOMED CT to DBPedia mapping and, from there, obtain the available candidate translation on the Portuguese DBPedia.
DBPedia is too big to be fully mapped in one batch with limited computing power, given the size of the ontologies involved. This would make the time required by AgreementMakerLight to align SNOMED CT with the full DBPedia prohibitive. However, it is unnecessary, given that most of DBPedia is irrelevant to the clinical domain, to use the full DBPedia. We expect that our users, domain experts in clinical specialisations, will select a batch of SNOMED CT terms of their interest at a time and create/revise the translations of the terms in that smaller set. For instance, to identify a set of allergy-related DBPedia pages to be aligned with a set of SNOMED CT terms, we used the UNIX grep tool to filter out of the DBPedia ontology every page with a label not containing any of the words of the SNOMED CT terms. This resulted in a size reduction from 2 GB to 12MB. To obtain the alignment with DBPedia, we parameterized AgreementMakerLight to consider as aligned all pairs of terms with a Jaro Distance 0:5.
The last step in our method involves the application of an
ensemble learning algorithm
CPARA, Cata´logo Portugueˆs de Alergias e Reacc¸o˜es Adversas, is
a list of terms related to Allergies and Adverse Reactions in use
in the Portuguese National Health Service
Evaluation of the translated SNOMED CT terms started with candidate translations for the allergy-related SNOMED CT codes in CPARA generated by application of our method. We then evaluated the resulting set of translations against the ground truth composed by the corresponding CPARA terms as defined by the medical committee that defined the mapping. To assess the accuracy of the evaluated translation methods, we scored each term translation by the Jaro distance between the automatically translated term and the CPARA translation. The Jaro distance (JD) between two strings is 1 if the strings have the exact same number of characters and do not have any transposition-10.
Prior to computing Jaro distances all the translation candidates and CPARA translations were normalised: we removed any qualifiers from the SNOMED CT candidates, deleted quotes from the CPARA translations, and converted all the named entities to lowercase (e.g, “Moderate (severity modifier) (qualifier value)” became “moderate” and “Contact metal agent (substance)” became “contact metal agent”). These preparatory steps are necessary to obtain meaningful similarity metrics, because these qualifiers are common to many terms and can be translated independently. In addition, the Jaro Distance considers the same letter in lowercase and uppercase forms as two distinct characters. The statistics of the translations obtained with each of the four implemented methods described in the previous section are given in Table 1. In these statistics, we considered as valid the translations with J D = 1.
Method LSanoguurcaege Coverage #Method AJVDG STJDDEV GT EN 100% 191 0.78 0.22 GT ES 114% 218 0.58 0.15 ICD 9 EN 10% 20 0.61 0.12 DBPedia EN 37% 70 0.89 0.03 Table 1. Global Results for all translation methods with the respective average Jaro Distance (AVG JD) and Standard Deviation Jaro Distance (STDEV JD). The implemented methods are Google Translate (GT), both from English (EN) and Spanish (ES) to Portuguese, ICD-9 Mapping (ICD 9), and DBPedia Mapping (DBPedia). All translations were attempted with two source languages, English (EN) and Spanish (ES). The number of terms translated by each method is given in the # Method column.
We observe that the SNOMED-DBPedia alignment obtains, for a coverage of 37%, both the highest similarity (0.89) and lowest standard deviation (0.03). This shows that we have been able to accurately translate a set of SNOMED CT terms to Portuguese, using basic alignment techniques, through the SNOMED CT to DBPEDIA alignment. However, the generation of translations 10 The computation of the Jaro distances was made with the Python Jellyfish library https://pypi.python.org/pypi/jellyfish based on ontology alignments as proposed in this paper also has limitations. In particular, only a fraction of the translations can be obtained by this method, while Google Translate always proposed a translation. Our success with Portuguese may not be granted when aligning SNOMED CT with DBPedia in other languages with smaller Wikipedias.
Google Translate EN showed better accuracy than Google Translate ES. This result was not initially expected, because Spanish and Portuguese are close languages. This may result from the CPARA terms being originally derived from the English terminology. The number of translations obtained with Google Translate ES is higher than the number of terms in the CPARA dataset (yielding the 114% coverage). This is the result of how we have obtained the Spanish SNOMED CT candidate terms for translation. We started from the same initial SNOMED CT codes that we used for the English translation and obtained the Spanish codes matching the concept id and type id of the initial English terms. This generated a higher number of ES candidate terms to translate (218) then the initial EN terms (191).
To evaluate which translation method works best for each class of SNOMED CT terms, we measure which translation method performs best in each class. This method is necessary to later model an ensemble learning stage that could pick the best method for each class. To obtain the results, we divided CPARA in classes for translation purposes. These classes were extracted from the qualifiers defined for the SNOMED full specified name terms. We were interested in observing translation performance differences across classes. To measure the differences, we calculated the similarity and standard deviation as above of all the translation candidates in each class. The results are summarised in Table 2.
The SNOMED-DBPedia alignment generates better translations for all classes, except Person and Qualifier Value. The poorer performance could, however, reflect that only a small number of related identified terms in the allergy domain have been identified for both classes.
Google TranslateES has better average similarity for the Person class than Google Translate EN. This shows that the SNOMED CT translation from Spanish could benefit from using the Spanish language distribution for some CPARA translations.
The translations obtained with the ICD-9 mapping translator are worse than obtained by Google Translate (for both languages). This results from ICD-9 being less comprehensive than SNOMED CT. ICD codes mostly diseases, symptoms or causes of death. Therefore, many of the CPARA terms in SNOMED CT were absent in the ICD-9 to SNOMED CT mapping. The results also indicate that, as expected and observed with ICD-9, terminologies of narrower scope are not useful for translating clinical terms through ontology alignment. The ICD-9 mapping is much less successful than other resources, such as DBPedia and Google Translate, which can provide much higher coverage of candidate translations, in many cases while retaining equal or better accuracy. The ICD-9 mapping method generates a high amount of 1-to-many matchings. However, ICD-9 could still be useful in cases where it generates only one matching description, which is usually very accurate and reliable, attending that matchings between ICD and SNOMED CT and the resulting translations are validated by medical experts. 5
SNOMED CT is increasingly prevalent in the heath care sector, resulting from the increasing need to exchange medical records in mobile societies. There is also a growing general interest in accessing standardised machine-readable medical records for improving managed heath care and biomedical research.
We introduced a new methodology for translating SNOMED CT terms, which relies primarily on aligning large ontologies, complementing language-based methods that have been proposed before. We prototyped an initial implementation of this methodology, which obtained high coverage and good accuracy, despite only using string matchers for SNOMED CT and DBPedia alignment along with the domain-independent Google Translator. A translation was considered valid when the expert mapping of an allergy-related SNOMED CT term to Portuguese is identical to the obtained using the SNOMED CT to DBPedia alignment. The accuracy under these settings was 37%. This shows that both the English and Portuguese versions of DBPedia are rich and accurately interlink with medical terms. However, the results for Portuguese may not be indicative of how this method would perform on other languages. Portuguese is one of the top-10 Wikipedia languages in terms of the total number of entries. The coverage of the obtained translations depend on how rich the Wikipedia for a target language is in covering clinical concepts and the extent to which these concepts are mapped to Wikipedia pages in languages for which a SNOMED CT translation exists. In addition, our validation experiment was confined to testing about 200 SNOMED CT Common Terms in the alergies and adverse reactions domain in European Portuguese. It is still unknown how comprehensive and accurate the English and Portuguese Wikipedias are across the full clinical domain, and how this factor affects the accuracy of the SNOMED CT translations.
Some improvements can still be added to the software implementing the presented translation method. For instance, the SNOMED CT to DBPedia alignment should explore the defined semantic relationships between classes and terms in both SNOMED CT and DBPedia. On the other hand, these relationships could be explored to generate accurate translations for untranslated terms in lexical methods to be provided. For this purpose, language resources, such as WordNet, and parallel corpora of named entities, such as previously validated SNOMED CT translations, could be used to learn how words and multi-word expressions should be properly translated.
The measured accuracy of our translation method could still be
significantly increased without sacrificing the quality of translations,
by relaxing the similarity threshold. The negative impacts of such
relaxation are negligible, given that the generated translations
will always need to be validated by experts before used in a
clinical context. The expert-validation step presently relies on
the review of generated translations presented on spreadsheets. A
crowdsourcing platform could speed-up the process of creating and
maintaining a validated translation of SNOMED CT. Moreover,
active learning could also be incorporated in the crowdsourcing
platform, leading to fast improvement of the proposed translations
as the validated translation can also be used as input to generate
good candidates
A complementary assessment of the alignment approach proposed here could be obtained by applying it to the automatic translation with one of the existing released translations, e.g. Spanish. However, given that we rely on alignments bewteen lexical resources we are not certain if the Wikipedia correspondences between clinical term pages in Spanish and English have been created based on SNOMED CT.
We thank Daniel Faria and the other members of the SOMER project for help with running AgreementMakerLight and their feedback. We also thank Dr. Anabela Santos for the help with the CPARA translation of SNOMED CT to validate our tool, and Bruno Martins for the pointers to previous works. This work was partially supported by Fundac¸a˜o para a Cieˆncia e a Tecnologia (FCT), grants PTDC/EIA-EIA/119119/2010 (SOMER), UID/CEC/50021/2013 and EXCL/EEI- ESS/0257/2012 (DataStorm).
AVG JD
Technique
Lang.