=Paper=
{{Paper
|id=None
|storemode=property
|title=Panacea, a Semantic-enabled Drug Recommendations Discovery Framework
|pdfUrl=https://ceur-ws.org/Vol-1061/Paper2_vdos2013.pdf
|volume=Vol-1061
|dblpUrl=https://dblp.org/rec/conf/icbo/DoulaverakisNKK13
}}
==Panacea, a Semantic-enabled Drug Recommendations Discovery Framework==
https://ceur-ws.org/Vol-1061/Paper2_vdos2013.pdf
Panacea, a Semantic-enabled Drug Recommendations Discovery
Framework
Charalampos Doulaverakis 1∗, George Nikolaidis 2 , Athanasios Kleontas MD 2,3 and
Ioannis Kompatsiaris 1
1
Centre for Research and Technology Hellas, Information Technologies Institute, Thessaloniki, Greece
2
Ergobyte S.A., Thessaloniki, Greece
3
Theagenio Cancer Hospital, Thessaloniki, Greece
ABSTRACT works such as (Adnan et al., 2010), but they don’t fully address
The paper presents Panacea, a semantic framework capable of the problem of automated drug prescription using drug-drug and
offering drug-drug and drug-diseases interaction discovery. For enab- drug-disease interactions.
ling this kind of service, medical information and terminology had to Rule-based approaches have been proposed for addressing issues
be translated to ontological terms and be appropriately coupled with relating to biomedical ontologies research. It is common for onto-
medical knowledge of the field. International standards, such as the logies written in expressive Semantic Web languages such as OWL,
ICD-10 and ATC classifications, provide the backbone of the com- not be able to handle all requirements for capturing the knowledge
mon representation of medical data while the medical knowledge of in several biomedical and medicine domains. As a method for enri-
drug interactions is represented by a rule base which makes use of ching the expressiveness of ontology languages, researchers have
the aforementioned standards. Representation is based on the light- proposed the use of rules which act upon the defined ontologi-
weight SKOS ontology. A layered reasoning approach is implemented cal knowledge. According to (Golbreich, 2004) rules are helpful
where at the first layer ontological inference is used in order to disco- in the following situations relating to biomedical ontologies: defi-
ver underlying knowledge, while at the second layer a two-step rule ning “standard rules” for chaining ontology properties, “bridging
selection strategy is followed resulting in a computationally efficient rules” for reasoning across different domains, “mapping rules” for
reasoning approach. Details of the system architecture are presented defining mappings between ontologies entities and “querying rules”
while also giving an outline of the difficulties that had to be overcome. for expressing complex queries upon ontologies. The author gives
The paper compares the current approach to a previous published a thorough review of RuleML and SWRL, the two major ontology
work by the authors, a service for drug recommendations named rule languages, the available rule formation tools and the reasoners.
GalenOWL, and presents their differences in modelling and approach (Golbreich et al., 2005) make use of the outcomes of the previous
to the problem, while also pinpointing the advantages of Panacea. paper to showcase the need for rules in biomedical applications with
a use case of a brain anatomy definition, where a brain structure
1 INTRODUCTION ontology is defined in OWL but rules describing the relationships
between the properties and entities are needed for correct annota-
One of the health sectors where intelligent information manage- tion of MRI images. Another work citing the need for semantically
ment and information sharing compose valuable preconditions for enriched rules, where an ontology coupled with SWRL rules for
the delivery of top quality services is personalized drug prescription. annotating pseudogenes and answering research questions has been
This is more evident in cases where more than one drug is required proposed in (Holford et al., 2010). All the above papers present
to be prescribed, a situation which is not uncommon, as drug inter- the need for extending ontologies with rules in order capture the
actions may appear. The problem is magnified by the wide range of knowledge of complex biomedical domains.
available drug substances in combination with the various excipients The paper presents Panacea, a semantic-enabled system for dis-
in which the former are present. covering drug recommendations and interactions. Panacea is based
If one takes into account that there exist more than 18,000 phar- on experiences and lessons drawn from the development of Gale-
maceutical substances, including their excipients, then it is clear nOWL (Doulaverakis et al., 2012), a similar system which had
that the continuous update of health care professionals is remarkably Semantic Web technologies in its core. As such, Panacea can be
hard. Over this, the extensive literature makes discovery of relevant considered the evolution of GalenOWL in terms of design and sca-
information a time consuming and difficult process, while the diffe- lability. Panacea makes use of established and standardized medical
rent terminologies that appear between sources add more burden on terminologies together with a rich knowledge base of drug-drug
the efforts of medical professionals to study available information. and drug-diseases interactions expressed as rules. Panacea is imple-
Semantic Web technologies can play an important role in the mented having in mind scalability, completeness of results and
structural organization of the available medical information in a responsiveness in query answering.
manner which will enable efficient discovery and access. Rese- The paper is organized as follows: Section 2 gives details about
arch projects funded for enabling Semantic Web technologies in the architecture and usage of the framework. In Section 3 the data
the diagnosis and therapeutic procedures exist such as PSIP (Beus- modelling approach is presented and in Section 4 the core ontology
cart et al., 2009) and Active Semantic Documents (Sheth, 2005) or and the layered reasoning process is described while also outlying
∗ To whom correspondence should be addressed: doulaver@iti.gr
1
Doulaverakis et al
two approaches for rule-based reasoning. Section 5 gives an evalua-
tion of the framework it terms of scalability ans performance and
the paper concludes with Section 6.
2 ARCHITECTURE AND FUNCTIONAL DESIGN
The purpose of Panacea is to provide drug prescription recommen-
dations based on a patient’s medical record, i.e. advise physicians to
prescribe medications according to the drugs active substance indi-
cations and contraindications. For details regarding the initiative that
triggered development of Panacea and the initial medical and phar-
maceutical data that were available, the reader is encouraged to read
(Doulaverakis et al., 2012). Panacea has been developed in the Java
programming language and is built using Apache Jena. Jena pro-
vides the API and methods to translate the medical knowledge of
terminologies and pharmaceutical rules to semantic entities, while
also providing the reasoning engine to enrich the knowledge base.
Panacea follows a layered reasoning process which is depicted
in Figure 1. During the start-up of the system, the medical termi- Fig. 1. Panacea framework architecture and data flow
nologies, namely ATC, UNII, ICD -10 and custom encodings, are
transformed to semantic entities, using an appropriate vocabulary,
and the initial ontology is constructed. The ontology binds to a rea-
diabetes mellitus for which he receives metformin (ATC: A10BA02)
soner to infer relations such as inheritance and unions. This process
and sitagliptin (ATC: A10BH01). For the new condition of pye-
is performed once offline during initialization and the knowledge
lonephritis that was diagnosed, the treating doctor must decide a
base is available to the system for further utilization. In order to
number of things. Regarding the prescription for treating this new
get recommendations in Panacea, a patient instance with the appro-
priate medical record data is created and fed to the knowledge base. disease, the doctor has to decide which active substances to pres-
cribe in order to treat the resulting inflammation, the cause of the
The reasoning process enriches the patient instance with inferred
inflammation, the back and abdominal pain and the resulting fever.
knowledge, thus making it explicit. On this enriched instance, and
However, before a decision is made the following factors regarding
by utilizing a different reasoning process, the set of medical rules is
the patient’s medical history should also be considered:
applied upon. The result of this final stage of rule-based reasoning is
• There should be a check for drug-drug interaction that the
the recommendations list which can be retrieved through SPARQL
patient is taking, before the onset of the new condition (the
querying.
pyelonephritis)
A key characteristic of the suggested architecture is that, regar-
ding second level reasoning, the framework can utilize any rule- • There should be a check for drug-disease interaction that the
based reasoner or rule engine. Since all the inferred knowledge of patient is taking, with the new condition
the medical definitions and patient data is materialized in the know- • The new prescription has to be verified that it will not have
ledge base, the medical rules can be expressed and loaded in an adverse effects or interactions with the previously prescribed
appropriate rule engine. The rule engine could be an ontology rea- medication and with the patient’s medical history
soner, such as Jena’s reasoning engine, or a business rule manager It is clear that the task for the doctor can be hard and a misjud-
such as Drools or even CLIPS with appropriate customizations in gement could lead to wrong prescriptions. Using an automated drug
the data structures. This approach helps in bringing together the recommendation system can minimize this risk. The recommenda-
best of both worlds: semantic and meaningful representation of data tion system will use the input data and the pharmaceutical rules in
using Semantic Web technologies and the maturity of traditional order to infer a treatment that will be safe for the patient.
rule engines in efficiently handling complex and large amounts of
rules.
3 SEMANTIC TRANSFORMATIONS
2.1 Use case scenario Panacea is built on top of international standards of medical termi-
In order to demonstrate the benefits of the proposed semantic nology in order to represent medical and pharmaceutical informa-
recommendation system, a use case regarding a possible scenario tion. The following standard terminologies are used:
is described: An elder man visits his family doctor complaining
for pain in his right lower back and abdominal region which is ICD-10: The World Health Organization classification of
accompanied with fever. After appropriate clinical examination, diseases. It is used in Panacea for unique identification of
he is diagnosed with right pyelonephritis (ICD-10 code: N11.0). diseases thus uniquely identifying drug indications and con-
According to the patient’s medical history, he is suffering from traindications related to diseases.
chronic atrial fibrillation for which he receives clopidogrel (ATC UNII: Unique Ingredient Identifier. Used for the identifica-
code: B01AC04), vertigo for which he receives cinnarizine (ATC: tion of active ingredients found in drugs. In Panacea it is used
N07CA02), high arterial blood pressure for which he receives can- for uniquely identifying drug indications and contraindications
desartan (ATC: C09CA06) and amlodipine (ATC: C08CA01), and related to ingredients.
2
Panacea
ATC: The Anatomical Therapeutic Chemical Classification is of using the language’s semantics to model the available informa-
used for the classification of drugs. In Panacea it is used in tion, there were problems resulting from this design decision. One
similar fashion to UNII. of the major issues was the difficulty in scaling the system. Until
IVT: The International Virus Taxonomy is used for the classi- currently, very few reasoners are available that can efficiently handle
fication of viruses. In Panacea it is used in order to uniquely the amount of class definitions and reasoning required to run the
drug indications and contraindications related to viruses. system, both in terms of memory consumption and speed.
In Panacea, a different approach was adopted. The SKOS (SKOS:
Apart from these international standards, a number of domain Simple Knowledge Organization System, 2009) vocabulary is a
classifications have been declared and used in order to enhance the W3C recommendation, it’s built using RDF(S) semantics and has
usability of the system or to represent data that are not included been developed as a low-cost migration path for porting existing
in the standards. These classifications act as supplementary to the knowledge organization systems, such as thesauri, taxonomies, clas-
standards. sification schemes and subject heading systems, to the Semantic
Substance: As the use of encodings for drug ingredients is not Web. It enables a “lightweight” semantic representation of such
convenient for humans, the identification of active substances is knowledge systems and is a good match for the medical standards
done using its common name references in medical bibliography. that are used in Panacea. As such, all the terminologies which are
These names come from international standards such as the Inter- mentioned in the previous section have been transformed using the
national Nonproprietary Names (INN) and others such as USAN SKOS vocabulary automatically using a parser.
(United States Adopted Name) or BAN (British Approved Name). Comparing SKOS to the approach followed in (Doulaverakis
Members of this identification list are substances such as acetazola- et al., 2012), instead of representing the ATC, ICD-10 and UNII
mide or isradipine. In addition, substances correspond to ATC codes classifications as top-level classes, they are now represented as
such that for example acetazolamide ≡ S01EC01. The substances instances of the skos:ConceptScheme class. “skos:” stands for the
are the actual recommendations of Panacea. SKOS namespace. Each entry in these classifications is represented
Custom Concepts: While the ATC, ICD-10, UNII and IVT stan- as an instance of the skos:Concept class. The OWL class hierar-
dards are complete, they are designed for use in contexts different chy of (Doulaverakis et al., 2012) is represented in Panacea using
from Panacea and drug recommendations, e.g. for annotation, the properties skos:broaderTransitive and skos:narrowerTransitive,
search or information retrieval. As such, it is often desirable to while the unions of classes for Custom Concepts Collections are
enrich the knowledge base with information that, while not standard, represented using the skos:member property. Correspondence bet-
will aid in the usability and overall efficiency of the system. Especi- ween the semantic transformation methodologies that were followed
ally for medical/pharmaceutical rules formulation, it was found out in the current work and in (Doulaverakis et al., 2012) is presented
that there were occasions that the definition of diseases, drugs or in Table 1.
other was either absent, incomplete or too general to be useful for It is interesting to note that the SKOS vocabulary offers exactly
a rule definition. An example for the lack of a definition in ICD-10 what is needed in order to capture the semantics of the medical
is the absence of a precise and specific code for “Chronic obstruc- classifications without making sacrifices in expressiveness. One can
tive pulmonary disease” or for “Hypertrophy (benign) of prostate”. argue that it can be considered more precise than the OWL expres-
For this reason, a number of custom concepts have been defined. sions, as in the case of the similarity of Substances and ATC codes.
Examples of such concepts is disease definition such as “Narco- This similarity is better represented by the skos:closeMatch relation
lepsy”, microorganisms such as “clostridium clostridiiformis” or than owl:equivalentClass. For Panacea a total of 64, 658 definitions
medical acts such as “upper extremity arteriography”. of classification codes have been expressed using SKOS.
Custom Concept Collections: Certain “groups” of substances
and/or diseases are frequently present in drug interactions and these
groups are not recorded explicitly in any standardized classifica-
4 PANACEA ONTOLOGY AND REASONING
tion, so it’s more convenient for medical use to specify these custom The core ontology of Panacea is visualized in Figure 2. The afo-
groups. These often used groups are termed “conditions” in Panacea rementioned SKOS ontologies were imported to the Panacea core
and are defined by medical experts. A condition can appear as a ontology under the MedicalDefinitions class. The Patient class holds
premise in other condition definitions, as in the Custom Concept the patient instances and is connected to the MedicalDefinitions
Collection cardiac-rhythm-abnormalities below, thus enabling their class with the hasData properties. The patient recommendations,
recursive definition: indications and contraindications, regarding substances that should
and should not be prescribed are expressed with the canTake and
cardiac-rhythm-abnormalities = cc:bradycardia | icd:R00 | cannotTake properties, respectively. The patients age group and sex
cc:tachycardia | icd:O68.0 | icd:O68.2 group are expressed through the hasAgeGroup and hasSexGroup
properties.
where cc:bradycardia is defined as (icd:I49.5 | icd:R00.1 |
icd:O68.0) and cc:tachycardia as (icd:R00.0 | icd:I49.5 | icd:I47 4.1 Medical reasoning
| icd:O68.0). “icd:” stands for the ICD-10 namespace. When querying the system for recommendations, a patient instance
is created with the initial patient data (through the hasData,
3.1 SKOS vocabulary hasAgeGroup and hasSexGroup properties) and is loaded in the
In the approach followed in (Doulaverakis et al., 2012), the medi- knowledge base. The reasoner, using RDFS inference and a
cal standards and the custom definitions were translated to OWL small number of additional rules, infers all the implicit pati-
classes, primitive and defined. While this approach had the benefit ent data. As an example consider a patient who suffers from
3
Doulaverakis et al
Table 1. Correspondence between the semantic transformation in the early GalenOWL system and the proposed Panacea framework
GalenOWL Panacea
Annotation rdfs:label skos:prefLabel
Equivalence owl:equivalentClass skos:closeMatch
Custom collections owl:unionOf skos:member
Hierarchy rdfs:subClassOf skos:broaderTransitive
in Figure 1 it can be seen that due to the layered reasoning approach,
the knowledge base (medical definitions + reasoner) is actually used
for producing the enriched patient instance. This means that the
! instance can be fed to a rule reasoner which has appropriately loaded
the medical-pharmaceutical rules, without the reasoner having to
communicate with the knowledge base for further utilization. Using
this approach and with proper modifications, any rule engine can
# $ be used to produce the drug recommendations. To demonstrate this
"
! ability, two separate rule engine integrations have been developed
and are presented below. The medical rule base consists of 1, 342
rules which were extracted and encoded directly from official docu-
Fig. 2. Panacea ontology ments, such as Summary of Product Characteristics (SPC) Patient
Information Leaflets (PIL), regarding drug indications, contraindi-
cations, interactions and dosage. The validity of the rule base has
a form of thrombocytopenia. An instance is created with the already been assessed in (Doulaverakis et al., 2012).
property . The reasoner It should be noted that work is under way in order to add more
through the skos:broaderTransitive relation will infer the triples functionalities in the drug proposed recommendation system. One
, , . Addi- posed dosage for a recommended substance. In order to accomplish
tionally, the custom collection definition of pnc-cc:deficiency- such a task, the pharmaceutical rules are being enriched with clini-
bone-marrow has icd:D69.6 as one of its members so the triplet cal variables that are important, other than sex and age group. These
will variables include somatometric characteristics such as height and
also be inferred. At the end, the patient instance will be enriched body weight, creatinin clearance (useful for calculating the dosage
with all the underlying implicit information. for antineoplasmic drugs) and the disease itself as a substance could
be indicated at a specific dosage to treat a certain disease, but a
4.2 Rule-based reasoning different dosage is recommended for another disease.
Drug recommendations in Panacea are generated using a rule-based
4.2.1 Jena rule engine For using the rule engine of the Apache
approach. The rules express the indications and contraindications of
Jena API (Apache Jena, 2012) the rules had to be translated to the
drug substances while their premises are the medical definitions and
Jena rule language. An automated parser was developed for this
the patients’ age and sex group. The rules use the logical operators
purpose. As for most semantic rule reasoners, OR clauses are not
and (&) and or (|) and parentheses. An example of a rule is for the
allowed in a rule definition so separate rules had to be expressed for
substance “ lisuride” which is expressed as
every premise that was OR’ed in the original rule base. For example,
the rule for “lisuride” was expressed by 3 different rules:
lisuride = icd:E22.0 | (icd:E22.1 & (icd:N91.0 | icd:N97)),
ageGroup=adult or elder 1. (?patient pan:hasData icd:E22.0) →
(?patient pan:canTake sub:lisuride)
The above rule reads that the substance “lisuride” is recommended 2. (?patient pan:hasData icd:E22.1)
for adult and elder patients who suffer from E22.0, OR suffer from (?patient pan:hasData icd:N91.0) →
E22.1 AND one of the N91.0 OR N97. For using these rules, they (?patient pan:canTake sub:lisuride)
have to be properly parsed and transformed in order to match the
3. (?patient pan:hasData icd:E22.1)
knowledge base and the enriched, with implicit knowledge, patient
(?patient pan:hasData icd:N97) →
instance. The proposed rule structure allows modifications to spe- (?patient pan:canTake sub:lisuride)
cific rules without the changes affecting the rest of the rule base.
This enables the rule base to be up-to-date with the latest clinical This rule expansion resulted in a total of 6, 451 rules to be expres-
advancements, which is a requirement as clinical pharmacology and sed in the Jena language. Trying to load the whole rule base and
medicine are constantly evolving. Analysing Panacea’s architecture performing inference for recommendations proved inefficient for
4
Panacea
real time use, requiring on average as much as 8 seconds. In order to high efficiency and optimizations of these engines with the semantic
tackle this issue a coarse rule selection phase was introduced. The description and interpretation of data.
selection was executed in 2 iterations. During the first iteration, a
subset A of candidate rules is created from the initial rule base, that
match the patient’s sex and age group. This subset is selected for fur- 5 EVALUATION AND DISCUSSION
ther processing. In the second iteration, rules from A that contain at For evaluating the framework, a comparison was made between the
least one of the patient’s data, i.e. a skos term, in their premises are two approaches for the final stage reasoning and with GalenOWL
singled out and a final set R ⊆ A is created from them. Remem- (with values taken from (Doulaverakis et al., 2012)). The compari-
bering that the implicit knowledge extraction was performed during sons were focused on the usability of the framework in a production
the introduction of the patient instance to the reasoning framework, environment as the rule base has been validated in (Doulaverakis
creation of R is actually a simple and fast process. It merely requires et al., 2012). Three parameters were measured. These were initia-
string matching and all the whole processing is executed in memory. lization time, the time to get the system up and running, memory
As a result the overall burden that is added to the whole reasoning consumption after initialization, and query response time, i.e. the
process is minimal. From the initial rule base of 6, 451 rules it is time that is needed to have the rule base executed and the results
common for R to contain as less as 50 rules, whose evaluation is retrieved. Results are shown in Table 2.
much more efficient. Rule execution is performed with the Jena There are some points to discuss in the table results. Initializa-
rule engine and the patient instance is modified and now contains tion involves loading the ontology in memory, performing inference,
the drug recommendations. These recommendations are retrieved and preparing the medical rule base for patient data reasoning. In
through SPARQL querying, using Jena’s query engine. The advan- the Jena implementation, the rule base is processed and loaded
tage of the Jena engine is that it can readily consume the patient only after the patient instance has been introduced to the system,
instance for producing the recommendations. while the Drools implementation loads the whole rule base on the
engine before any patient data are introduced. As a result, Drools
4.2.2 Drools rule engine As an alternative approach, the Drools appears slower than the Jena approach regarding initialization. For
(Drools, 2012) business rule engine was used. In contrary to the the same reason, memory consumption appears greater for Drools.
Jena engine, Drools could not directly use the patient instance for This metric corresponds to memory consumption from initialization
performing reasoning. For this purpose, the instance was transfor- to recommendations retrieval. While in Drools the whole rule base
med to a Java bean, where the properties of the ontology Patient is loaded on memory, in Jena the approach was to load a small sub-
class are mapped to Java methods using the JenaBean API (http: set of the rule base that could possibly match the patient data, which
//code.google.com/p/jenabean/). The bean was appro- leads to a smaller memory footprint. Finally, for query response the
priately declared to Drools and was handled for rule execution. A advantage is with Drools, as was expected, mainly due to the fact
similar approach for integrating Jena and Drools was used in (Bra- that Drools is a dedicated rule engine while Jena’s focus is not at
gaglia et al., 2010). The Drools Rule Language (DRL) permits the providing a state of the art reasoner and rule engine, but a versatile
use of OR’ed clauses in the body, so the 1, 342 original medical API for ontology management.
rules were translated to the same amount of rules in DRL, using an Numerically, the Jena approach seems to be more efficient than
automated parser similar to the one used in the Jena approach. For Drools, apart from the query execution time but for which the dif-
example, the rule for “lisuride” from the previous paragraph was ference is not important. However, while for the present knowledge
expressed in DRL as: base Jena seems to perform better, this fact could change as more
and more rules are added. It is estimated that eventually at its final
RULE ‘‘lisuride’’
WHEN stage, Panacea will incorporate more than 9, 000 drug-drug and
p: Patient(data : hasData) drug-disease interactions. As already said, Jena is more focused as
exists( (MedicalDef(uri==icd:E22.0) an ontology API and less as an efficient rule engine which could
from data) || eventually lead to scaling problems. On the other hand, scaling with
(MedicalDef(uri==icd:E22.1 && uri==icd:N91.0)
from data) ||
Drools is not an issue. The value of business rule engines as Seman-
(MedicalDef(uri==icd:E22.1 && uri==icd:N97) tic Web reasoners has been previously exploited using approaches
from data) ) such as (O’Connor and Das, 2012), where the authors implemen-
THEN ted two OWL2-RL Motik et al. (2009) reasoners using the Drools
Substance lisuride = (Substance)JenaBean.
and Jess rule engines respectively. The use of traditional rule engi-
reader().load(sub:lisuride);
modify(p) {p.canTake(lisuride)} nes with the Semantic Web technologies brings together the best of
END both worlds, i.e. increased efficiency coupled with interoperability
and semantic annotation of information.
Execution was straightforward with no preprocessing required. What is also noticeable from Table 2 is the decreased memory
Drools is optimized for handling large rule bases, so no rule pre- requirement of Panacea compared to the previous OWL- based
selection step was required as this would have little impact in GalenOWL system, although the two approaches offer very similar
reasoning efficiency. The result of this reasoning process is a modi- functionality. As a result of this achievement, Panacea can accom-
fied patient Java bean with the drug recommendations. The bean modate a far greater knowledge base thus supporting the claim of
is transformed to Jena model instance and SPARQL querying for increased scalability.
retrieving the recommendations is possible. What this approach Panacea will eventually be offered as a service with potential
demonstrates is that it’s possible to integrate business rule engines customers being health care professionals. Other possible exploi-
as reasoners in the framework, thus being able to make use of the tation routes are being investigated such as integration to patient
5
Doulaverakis et al
Table 2. Evaluation between the 2 Panacea reasoning approaches and GalenOWL
Panacea-Jena Panacea-Drools GalenOWL
Initialization time 32.0 s 34.7 s 148 s
Memory consumption 169 MB 280 MB 649 MB
of which rule base consumes 0 MB 111 MB —
Query response time 47 ms 5 ms 16 ms
management systems in health clinics. The use of personalized drug still in development, it’s actively supported and it is mature enough
prescription systems, as Panacea, in everyday practice will have to be able to use it as a testing framework.
major advantages to the society and the economy. A major bene- ACKNOWLEDGMENTS
fit from the use of such systems is the reduction of medical costs This work has been supported by the national project “Panacea”,
through rational drug prescriptions that personalized drug prescrip- funded by GSRT Hellas under the “Support for SMEs” programme.
tion allows (Fischer et al., 2008). Another benefit is a positive effect
in public health with reduction of outbreaks relating to drug interac-
REFERENCES
tions or adverse effects (Ammenwerth et al., 2008). All knowledge
Adnan, M., Warren, J., and Orr, M. (2010). Ontology based semantic recommendations
regarding drug information is encoded and is available to the experts
for discharge summary medication information for patients. In Computer-Based
in order to aid them during prescriptions thus acting as decision sup- Medical Systems (CBMS), 2010 IEEE 23rd International Symposium on, pages 456–
port systems. It should be stressed out that drug recommendation 461.
systems do not aim to replace medical experts but to support them Ammenwerth, E., Schnell-Inderst, P., Machan, C., and Siebert, U. (2008). The effect of
in their practice. electronic prescribing on medication errors and adverse drug events: A systematic
review. J Am Med Inform Assoc., 15(5), 585–600.
A limitation of the proposed approach is that a rather large Apache Jena (2012). Semantic web framework. http://jena.apache.org/.
amount of manual effort by experts is required in order to populate Beuscart, R., McNair, P., J, J. B., and PSIP consortium (2009). Patient safety through
and enrich the rule base. Although the semantic technologies that intelligent procedures in medication: the PSIP project. Studies in Health Technology
have been employed can make rule authoring simpler, no automa- and Informatics, 148, 6–13.
Bragaglia, S., Chesani, F., Ciampolini, A., Mello, P., Montali, M., and Sottara, D.
ted method for pharmaceutical rule generation has been integrated.
(2010). An hybrid architecture integrating forward rules with fuzzy ontological
However, one would argue that since rule authoring is performed by reasoning. In Proceedings of the 5th international conference on Hybrid Artificial
experts then the rules are verified and guaranteed to be correct. Even Intelligence Systems - Volume Part I, HAIS’10, pages 438–445, Berlin, Heidelberg.
if an automated method, such as rule mining, had been implemen- Springer-Verlag.
ted, the generated rules would still have to be verified be an expert Doulaverakis, C., Nikolaidis, G., Kleontas, A., and Kompatsiaris, I. (2012). Gale-
nOWL: Ontology based drug recommendations discovery. Journal of Biomedical
in the field. Manual verification, although less intensive, would still Semantics, 3(14).
be required. Drools (2012). Business logic integration platform. http://www.jboss.org/drools.
Fischer, M., Vogeli, C., Stedman, M., Ferris, T., Brookhart, M., and Weissman,
J. (2008). Effect of electronic prescribing with formulary decision support on
medication use and cost. Archives of Internal Medicine, 168(22), 2433–2439.
6 CONCLUSIONS Golbreich, C. (2004). Combining rule and ontology reasoners for the semantic web,
invited talk, rules and rule markup languages for the semantic web. In Boley Editors,
The paper presented Panacea, a framework for semantic-enabled
LNCS 3323. Springer.
drug recommendations discovery. The framework utilizes a layered Golbreich, C., Dameron, O., Bierlaire, O., and Gibaud, B. (2005). What reasoning
reasoning approach were the medical ontology and the patient data support for ontology and rules? the brain anatomy case study. In Proceedings of the
instances are fed to an extended RDF(S) reasoner in order to infer KR 2004 Workshop on Formal Biomedical, pages 60–71.
implicit knowledge. Drug recommendations are generated using the Holford, M., Khurana, E., Cheung, K.-H., and Gerstein, M. (2010). Using semantic
web rules to reason on an ontology of pseudogenes. Bioinformatics [ISMB], 26(12),
second reasoning layer where any common rule engine can be used. 71–78.
As a proof of concept implementation, the Jena reasoner and the Motik, B., Grau, B. C., Horrocks, I., Wu, Z., Fokoue, A., and Lutz, C.
Drools rule engine has been integrated and separate tests regarding (2009). OWL 2 Web Ontology Language Profiles, W3C recommendation.
requirements and efficiency were conducted. For the Jena reasoner http://www.w3.org/TR/owl2-profiles/.
O’Connor, M. and Das, A. (2012). A Pair of OWL 2 RL Reasoners. In Proceedings
implementation a 2 step rule selection method was followed which
of OWL: Experiences and Directions Workshop 2012 (OWLED-2012), Heraklion,
resulted in computationally efficient reasoning. The tests concluded Greece.
that both approaches perform significantly better than the earlier Sheth, A. (2005). Semantic web & semantic web services: Applications in healthcare
GalenOWL system, while at the same time maintaining the same and scientific research, keynote talk. In IFIP Working Conference on Industrial
quality of results and improving performance. Concerning future Applications of Semantic Web, Jyvaskyla, Finland.
SKOS: Simple Knowledge Organization System (2009). W3C recommendation.
work, the addition of dosage recommendations in the rules is an http://www.w3.org/2009/08/skos-reference/skos.html.
ongoing work. Additionally, the possibility to add probabilistic rea- Sottara, D., Mello, P., and Proctor, M. (2008). Adding uncertainty to a rete-oo infe-
soning will be investigated. To this end, Drools is being extended rence engine. In Proc. of the International Symposium on Rule Representation,
with a fuzzy reasoning engine (Sottara et al., 2008), which while it’s Interchange and Reasoning on the Web. RuleML ’08, pages 104–118.
6