A well-known alternative to identify hidden standards is the use of data mining techniques. In order to obtain more efficiency in data mining, ontologies have been used to improve the representation in specialized knowledge domains. Here, we apply ontologies in a dataset of a diagnostic medicine company, which concerns to viral human hepatitis, with the aim of obtaining the best correlations between the laboratory tests prescribed by physicians and the real occurrences of diseases. Our preliminary findings show that the use of ontologies provides reduction in the number of attributes in the pre-processing phase, then improving the performance of data mining process as a whole.
generalization and the consequent reduction of the number of attributes to be mined via the identification of similarities between laboratories tests, considering the relations mapped in the viral hepatitis ontology called HVO.
The next sections will be organized as follows: section 2.1 provides a brief description on the use of ontologies in data mining; section 2.2, describes the construction of prototype of a viral hepatitis ontology; section 2.3 explains how generalization of terms will be applied in the data mining pre-processing phase as per association rules; section 3 details the results obtained with the proposed model. Finally, section 4 showcases final considerations.
In some fields, such as Biomedicine, specialized communities have been developing and
publishing, since the 1990s, a series of ontologies to aid in representing and retrieving
informational [
Knowledge extraction, generally referenced in literature as Knowledge Discovery in Database (KDD) should be grouped into three phases: pre-processing, DM and postprocessing. Pre-processing, which is relevant for our goals in this paper, comprises the collection, organization and treatment of data, while DM involves algorithms and techniques to search for knowledge.
Ontologies have been used to increase the relevance of the patterns discovered through the mining techniques. One of the techniques in which ontologies are being utilized is mining through association rules, which display the correlation between sets of items in series of data or transactions. [Ferraz 2008].
The advantage of pruning restrictions is to exclude information in which users are not interested in since the beginning. Every general rule should be able to replace a number of specific rules by means of generalization processes. Whenever this approach is feasible, a semantic improvement of the mined association rules and a reduction in the cardinality of the set rules will simultaneously take place.
The development of the viral hepatitis ontology was based mapped clinical analysis
laboratory tests for diagnosing human viral hepatitis considering LOINC, OGMS
ontology [
We associate the laboratory tests with the viral infectious disease to enable the generalization of the attributes to be mined, as proposed in Figure 1. In the triage phase of Hepatitis C, for example, four specific tests may be requested for the virus identification and another eight unspecific tests may be ordered for monitoring liver functions. This situation may be generalized without having denominated each laboratory tests as an attribute for data mining.
The ontology for hepatitis was created at this moment of our experiment so that we could test its application in our computational architecture. We are aware that some improvements in modeling are in order, for example: i) "has symptom" and "is observed" are not ontological relations; ii) "An axioms like Hepatitis C subclass Of hasSymptom some Jaundice can be falsified by one single patient who has Hepatitis C but no jaundice"; iii) instead of "no symptom, and following OGMS, we should think in use "healthy organism". Such improvements which will be part of our future work in the following of the research.
Our study makes use of ontologies, reasoners and Jena software to promote pruning and filtering (generalization) of data from the list of laboratory tests collected from the diagnostic medicine company’s database. When analyzing the relationships shared by the terms, one might identify which laboratory tests are related to which disease and stage. Therefore, the similarity between terms is considered as a means to generalize the attributes in the pre-processing phase. Figure 2 depicts the proposed model.
Based on HVO ontology, we consider relationships with the disease and with its features and also utilizing the Jena6 tool, as well as inference rules. So, we were able to obtain more general terms to represent laboratory test groups associated with viral hepatitis diagnosis.
The development of ontologies along with the use of inference mechanisms during the pre-mining phase has reduced the number of attributes to be mined by the association rules algorithm, namely, the Apriori algorithm. It reduced the amount of laboratory tests related to the direct diagnosis of the hepatitis virus, and also the number of unspecific tests for assessment over the course of diseases.
Based on knowledge obtained from the development of HVO, a list of laboratory
test orders was collected from the company’s database containing at least one test
directly related to a hypothetical hepatitis virus diagnosis. Laboratory test orders that
complied with the previously described rule were selected during three months, January,
February and March 2015, totaling 34440 occurrences (Table 1). Test applications,
which are complementary to the diagnosis, are distributed in collections made on the
organization's service units, conveyed by laboratory partners throughout Brazil. In this
sample, the occurrences featured 465 different laboratory tests. Considering the data of
service units (support = 0.2), laboratory patterns (support = 0.02) and confidence 0.75
was executed the ARules package [
Considering the same database obtained by reduction of the number of attributes, with the use of ontology, applied to 439 different laboratory tests and with the same support value and confidence was executed again the algorithm Apriori . For base units were obtained 258 and 115 rules for laboratory partners, we reached a reduction of 50% of the resulting association rules, as shown in table 2.
The attributes were categorized considering the relationship between the modeled tests through equivalence axioms as showed in the example below: hvo:laboratory_diagnostic_process_ hepatitis_A equivalent to: hvo:laboratory_testing_encounter And (is_composed_of some (‘Laboratory test' and (diagnoses only 'hepatitis A'))) and (is_composed_of min 0 ('laboratory test' and (diagnostic_evaluation some Liver)))
In this equivalence axiom (described by existential restrictions), a part of the detailed diagnosis of the disease process is comprised of at least one medical application (in this case Class HVO: laboratory_testing_encounter), which is composed of complementary examinations (OGMS : laboratory test) for disease diagnostic (doid: hepatitis a) and can also be a laboratory test for evaluating the state of the health (HVO: diagnostic evaluation) of a liver (FMA: liver).
Considering the limitation of further tests and the disease is possible to identify relationships between them and, thus, promote the generalization and its representation in single attribute, in this case, a type of viral hepatitis, as shown in Table 3.
Table 3 shows the laboratory tests (LT) prescribed to patients by the doctor and sent to the medical diagnostic laboratory. With the generalization attributes through the method was represented axiom disease in which some complementary tests are associated, in this case, a type of hepatitis. The other complementary tests were maintained and makes up the list of attributes analyzed by mining technique Apriori which extracted association rules related to viral hepatitis.
Therefore, our findings suggested that it is possible to reduce the rules resulting from data mining by reducing the possibilities of combining attributes. As a second, we found that the generalization of the terms enables results with a greater significance, since it can guide the post-mining phase analysis process.
The application ontology developed here strongly represents the LOINC tests for viral hepatitis, as we understand that this classification is sufficient for assessing the relationship of association rules. The extension of OGMS, DOID, FMA and IDO brings 7Identification of the laboratory test 8LOINC 1834-1 - Alpha-1 Fetoprotein – Laboratory test unspecified viral hepatitis 9 LOINC 61151-7 - Albumin – Laboratory test unspecified viral hepatitis 10 LOINC 5196-1 - Hepatitis B virus surface Ag 11 LOINC 1742-6 - Alanine aminotransferase – Laboratory test unspecified viral hepatitis 12 LOINC 5179-7 - Hepatitis A virus Ab.IgG 13 LOINC 13950-1 - Hepatitis A virus Ab.IgM laboratory tests closer to the diagnosis cycle and, consequently, promotes the identification of the correlations between lab tests.
It is relevant to highlight two points. Firstly, the relevance of patterns extracted by means of techniques that identify semantic similarity between terms is highly dependent of the ontology construction and validation. Therefore, it is fundamental that the domain ontologies being used be validated by specialists. Secondly, the KDD approach requires greater reach of the algorithms and cannot be restricted to “is-a” and “part-of" relations, which reinforces the use of formal semantics of ontologies.
In future work, it is intended to promote the enrichment of the ontology with new concepts and equivalence between complementary tests and disease for greater generalization of attributes.