In life science and biomedicine much knowledge resides as unstructured information in for instance bibliographic databases. To facilitate searching and categorisation of this information the database entries are annotated with terms or keywords, describing for instance diseases, treatments and anatomy. These annotations are limited to concept level and do not describe relations between terms, for example that a given treatment may be used for a given disease, even if this information is available in both the text and terminologies. In this work we will present a possible approach to extend term annotations with relational information to add another dimension to concept focused annotation schemas. This approach could also be used to highlight implicit information and to structure knowledge.
In life science and biomedicine much knowledge resides as unstructured information in for instance bibliographic repositories or electronic health record databases. To facilitate searching and categorisation of this information the database entries are annotated with terms or keywords, describing for instance diseases, treatments and anatomy. These annotations are limited to concept level and do not describe relations between terms, for example that a given treatment may be used for a given disease, even if this information is available in the paper. This annotation limitation remains in spite of accessible relational information in ontologies and terminologies.
In this work we study the possibility to annotate a term annotated scientific
medical corpus with relations between the terms, for instance relating diseases
to treatments and organ sites. We use a Swedish text corpus of scientific medical
documents [
Current related work has focused on for instance methods for extraction of
specific types of relations [
However, our work focuses not on establishing methods for extraction of relations between terms, but on tying existing term annotations together, reflecting relations in the text. These new relational annotations would allow both medical scientists and search engines to take advantage of highlighted implicit information on for instance diseases, treatments and anatomy. For example a paper regarding prevention of myocardial infarction may be annotated with terms like Aspirin and Myocardial Infarction and our proposal is to also annotate it with the relation may_prevent relating these terms. Not only would this add new useful annotations, but it would also structure existing annotations. 2 2.1
The main resources in this study are an annotated Swedish scientific medical
corpus [
Medical Text Corpus. As a part of the Swedish strategy for e-health, the
clinical terminology SNOMED CT3 has been translated into Swedish. For
validation and quality assessment of the translation, a Swedish medical text corpus
was created from the electronic archives of the Journal of the Swedish Medical
Association 1996-2009 [
Our study focuses on the MeSH annotated sentences in a part of the corpus containing 2021 articles from the domain “Klinik och Vetenskap” (Medical Practice and Science). This part contains 140458 sentences, each with a unique id. For copyright reasons the order of the sentences have been randomised, thereby limiting our study to relational annotations at sentence level. Of this sentence set we used only the 102821 sentences with at least two MeSH annotations. UMLS. The Unified Medical Language System (UMLS)4 connects vocabularies from different biomedical and health-related sources in different languages. It provides, among other things, databases, called Knowledge Sources. One of the databases, Metathesaurus, contains information about more than one million biomedical or health-related concepts. This database is divided into a number 2 www.nlm.nih.gov/research/umls 3 www.ihtsdo.org 4 We have used UMLS version 2010AB, including all source vocabularies of level 0-3. of relational tables. One of the major tables, MRCONSO, contains the structure for each concept, e.g. names, identifiers, languages and source vocabularies. This table is complemented with the MRSAT connecting MeSH identifiers (or identifiers from other source vocabularies) to concept identifiers (CUI).
Metathesaurus also contains relations between concepts, where the table MRREL contains basic relations (REL), e.g. Parent/Child, relating different concepts. Around 25% of the relations have a label (RELA - Relationship Attribute) which comes from the source vocabulary and specifies the relationship, e.g. isa, treated_by, finding_site_of or has_component.
Considering the MeSH part of the UMLS, a term can belong to more than one category and thereby appear in several places in the MeSH hierarchy with different MeSH identifiers. Moreover, by the annotation procedure used in our study corpus, a term like Blood Pressure will be annoted as Blood Pressure, but also as Blood and Pressure. Hence, the MeSH annotations can be nested, which can result in more UMLS concepts than found terms in a sentence. 2.2
Since this work can be seen as a feasibility study of extending term annotation schemas with relational information, the methods have been kept simple and analysis and validation of the approach were done by manual inspection.
For each of the sentences in our corpus containing at least two MeSH annotations, we extracted the sentence identifier and MeSH annotations, not taking into account any nesting of the annotations. The resulting information was stored in a MySQL database, complementing the UMLS one, thereby allowing easy mapping of MeSH identifiers to UMLS concepts via the MRSAT table. These concepts could then be used to derive relations from MRREL, giving relations between the annotated terms in each sentence.
Since our main interest is in the ability to provide relational annotations among annotated terms, the analysis focused on the derived relations with information in the RELA field of MRREL. One such RELA relation is may_prevent, e.g. “Aspirin may_prevent Myocardial Infarction”. These relations were divided into five different categories reflecting disease-treatment, disease-organ, causeeffect, hierarchical and other relations among the terms to reflect relations often found in medical papers. The division of the RELA relations into these classes was based on our subjective interpretation of relations like e.g. may_treat, has_finding_site and cause_of. For each of the relation categories, we randomly picked sentences and manually compared the derived relations to the ones expressed in the sentences to see if and how they were related. Since this work is to be viewed as a feasibility study for future research, we limited our analysis to a handful of sentences per relational category. 3
Our study corpus comprised 140458 sentences, with 102821 (73%) containing at least two MeSH annotations and in 26251 (25%) of these we were able to identify relations between the terms. In these sentences we found 150024 relations of which 44754 (30%) had the specification RELA. There were 188 different RELA. Table 1 shows the percentage of sentences per RELA relation.
Figure 1 shows the 122483 sentences containing MeSH annotated terms divided into number of MeSH terms (left) and number of sentences per number of relations (right)5.
In the rest of this section we analyse each of the defined relation categories and exemplify relations and their corresponding sentences, Table 2, and in Discussion we briefly elaborate on these results.
Disease - Treatment – The relations between diseases and treatments found in the studied part of the corpus are e.g. may_treat and may_prevent. Examples of disease-treatment relations are the relations derived from sentences 77591, 50380 and 105057. From sentence 105057 it is not possible to infer what the relation is between the terms Soft Tissue Infections and Methicillin.
Disease - Organ – There are a number of different relations between diseases and organs, e.g. is_associated_anatomic_site_of and has_finding_site. The relation location_of is often a relation between diseases and treatments, but sometimes refer to organ-organ relations. Disease-organ relations were found in for example sentences 75743, 8039 and 97183. In 75743 and 8039 the relation was not explicitly expressed.
Cause - Effect – Cause and effect relations can be for example cause_of, induces and causative_agent_of. It can be for instance viruses or bacteria which cause diseases, or diseases that cause other diseases. Examples of these relations were found in sentences 133130 and 125988. 5 In the UMLS, for many relations there is also an inverse, e.g. finding_site_of and has_finding_site. s e c 0 tennS 0010 e 0 0 0 0 2 0 0 0 5 1 0 0 0 5 0 0 0 0 4 1 0 0 0 0 se 01 c 0 tenne 800 S 00 6 0 0 0 4 0 0 0 2 0 1 4 7 10 13 16 19 22 25 28 31 34 37 47
MeSH terms 2 4 6 8 10 12 14 16 18
MeSH relations
Hierarchical Relations – The hierarchical relations we have studied are synonymy, hyponymy and sibling relations. Synonyms are for instance concepts which have the relation same_as or has_tradename. The relations mapped_from and primary_mapped_from can be synonym relations, sentence 77314. Hyponymy can be relations like isa, may_be_a and part_of, sentence 43382. Sibling relations found in the sentences are for instance sib_in_isa. In the example of the sibling relation, where Ethanols, Methanols and Ethylene Glycols are all Alcohols (sentence 14668), the relations have no specification in RELA, only the REL abbreviation SIB.
Other Relations – Other relations that were found in the sentences in the corpus are for instance co-occurs_with, associated_with, may_diagnose and occurs_before.
Sentences 15403 and 12826 are examples of the relation co-occurs_with, while sentences 52963, 51136 and 125709 are examples of the relations associated_with, may_diagnose and occurs_before respectively. 4
The examples in this work show that it is possible to derive relations from the annotated terms in a sentence utilising only UMLS. The majority of the examples for the different relation types are existing relations, but sometimes the derived relations are implicit instead of directly expressed in the sentences. However, as exemplified in sentence 75743, the derived relations may not be the ones intended in the sentence.
Since we have studied only concepts and not taken into account the syntax of the sentences, the resulting relations between the terms can be relations not
Opium may_treat Pain Aspirin may_prevent Myocardial Infarction Soft Tissue Infections may_be_treated_by Methicillin Pancreas is_associated_anatomic_site_of Diabetes Mellitus Celiak Disease has_finding_site Intestines Central Nervous System finding_site_of Rabies Bacillus antracis causative_agent_of Anthrax Nitrous Oxide induces Nausea Arthritis primary_mapped_from Joint Diseases Estrogenes isa Hormones SIB: Ethanols/Methanols/Ethylene Glycols Obesity co-occurs_with Diabetes Mellitus Diabetes Mellitus co-occurs_with Hypertension Vaccination associated_with Immunization Triiodothyronine may_diagnose Thyroid Disease
Chickenpox occurs_before Herpes Zoster (shingles) Sentence ID Sentence 77591 50380 105057 75743 8039 97183 133130 125988 77314 43382 14668 15403 12826 52963 51136 125709 [...] opium had a soothing effect on both anxiety and pain. [...] observed that patients with a regular intake of Aspirin had fewer heart attacks than expected. [...] soft tissue infections [...] and infections caused by methicillin resistant staphylococcus.
An alternative for a few patients with diabetes mellitus has been transplantation of pancreas [...] [...] patients with already known celiak disease but who in spite of a strict diet have had gastro-intestinal symptoms. [...] the symptoms of a rabies infection begins when the virus reaches CNS [...] Bacillus antracis causes the disease Anthrax [...] [...] nitrous oxide contributes to post-operative nausea. [...] patients with [...] and arthritis symptoms who were treated with [...] improvement of their joint disease [...] Even though it is well documented that estrogene [...] is an important hormone in [...] [...] have been introduced as an alternative to ethanol in cases of ethylene glycol and methanol poisoning. [...] obesity is associated with a highly increased risk of developing [...] diabetes [...] Other risk factors for stroke are [...] hypertension, diabetes, [...] Participation in the vaccination programs has been very high and sufficient immunization was reached [...] Antibodies targeting Triiodothyronine [...] in up to 10 percent of patients with Thyroid diseases. [...] a connection between chickenpox and herpes zoster. derivable from a sentence. For instance, in sentence 105057 there is no indication of any relation between the terms Soft Tissue Infections and Methicillin, but from the UMLS we get a may_be_treated_by relation. Many of the derived UMLS relations are not explicitly expressed in the sentences, but can be part of the context of a sentence, for instance the relation between Diabetes Mellitus and Pancreas in sentence 75743.
Expressing hierarchical relations like isa and part_of could lead to increased understanding of the context of the sentence. For example in 14668, where, by the sibling relation, we learn that the terms Ethanols, Methanols and Ethylene Glycols have something in common, i.e. they are all Alcohols, thereby framing the concepts in the sentence.
The terms which have the relation co-occurs_with can have slightly different relations to each other in the sentences. For example in 15403 one problem leads to another, but in 12826 the two diseases can both be the cause of a third one.
One of the major reasons for only being able to identify relations in 25% of the sentences with more than two annotations, may be that the annotations are not at the same hierarchical ontological levels in comparison to the defined relations in UMLS. A fundamental challenge with the proposed approach is its dependence on the quality and source of the original term annotations, as in our case using only MeSH in the process to identify relational annotations. 5
In this work we have studied the feasibility of utilising the term annotations of medical text in connection with a collection of terminologies and vocabularies to extend the annotations with relational information. As the examples show, this approach can be used not only for annotation, but also to highlight implicit information and to structure knowledge.
Our work complements existing research on extraction of (semantic) relations in biomedical text, by focusing on identifying and validating relations between conceptual annotations of a text. Hence, instead of the complex process of extracting relations in medical papers, we utilise existing annotations to propose potential relations covered by a paper.
This work is based on only one of the source vocabularies and a limited corpus. Hence, future work will address the ability to make use of combinations of several source vocabularies and more elaborated use of the hierarchical ontological relations to increase the ability to identify relations among annotated terms. However, utilising for instance hyponymy induces challenges like degree of hyponymy/hypernymy to allow in establishing relations among terms, and also how to resolve problems with the complex relational structure of the UMLS with many different types of relations and even cycles. For instance, some vocabularies in the UMLS may treat relations like isa and part_of as synonymous and some as distinct types of relations. Ongoing work also considers other corpora, like parts of MEDLINE6, for relational annotation. 6 www.nlm.nih.gov/databases/databases_medline.html