Inferring Recommendation Interactions in Clinical Guidelines 1 · minimize errors, and generalize...

Undefined 0 (0) 1 1IOS Press

Inferring Recommendation Interactions inClinical Guidelines 1Case-studies on Multimorbidity

Editor(s): Stefan Schlobach, VU University Amsterdam; Krzysztof Janowicz, University of California, Santa BarbaraSolicited review(s): Michel Dumontier, Stanford University; three anonymous reviewers

Veruska Zamborlini a,b,∗,∗∗, Rinke Hoekstra a,c, Marcos Da Silveira b, Cedric Pruski b,Annette ten Teije a and Frank van Harmelen a

a Dept. of Computer Science, VU University Amsterdam, The NetherlandsE-mail: {v.carrettazamborlini, rinke.hoekstra, a.c.m.ten.teije, f.a.h.van.harmelen}@vu.nlb LIST Luxembourg Institute of Science and Technology, LuxembourgE-mail: {marcos.dasilveira, cedric.pruski}@list.luc Faculty of Law, University of Amsterdam, The Netherlands

Abstract. The formal representation of clinical knowledge is still an open research topic. Classical representationlanguages for clinical guidelines are used to produce diagnostic and treatment plans. However, they have importantlimitations, e.g. when looking for ways to re-use, combine, and reason over existing clinical knowledge. Theselimitations are especially problematic in the context of multimorbidity; patients that suffer from multiple diseases.To overcome these limitations, this paper proposes a model for clinical guidelines (TMR4I) that allows the re-useand combination of knowledge from multiple guidelines. Semantic Web technology is applied to implement themodel, allowing us to automatically infer interactions between recommendations, such as recommending the samedrug more than once. It relies on an existing Linked Data set, DrugBank, for identifying drug-drug interactions. Weevaluate the model by applying it to two realistic case studies on multimorbidity that combine guidelines for two(Duodenal Ulcer and Transient Ischemic Attack) and three diseases (Osteoarthritis, Hypertension and Diabetes)and compare the results with existing methods.

Keywords: Clinical Knowledge Representation, Reasoning, Combining Medical Guidelines, Multimorbidity, OWL,SPARQL, Rules, SWRL

1This is an extended version, by invitation, of a paper ac-cepted at the 19th International Conference on KnowledgeEngineering and Knowledge Management (EKAW 2014),in which the TMR4I model was first introduced [31].

*Corresponding author. E-mail:[email protected].

**Funded by CNPq (Brazilian National Council for Sci-entific and Technological Development) within the programScience without Borders.

1. Introduction

Information and communication technologiesare widely adopted in the clinical domain asHealthcare Information Systems. The primary fociof these systems are data management and vi-sualization. Because of this, there is an increas-ing amount of patient data that can be used tostudy disease behavior and the effectiveness of cer-tain treatments. The outcomes of such studies findtheir way into so-called “reference documents” forclinical practices, promoting the use of evidence-

0000-0000/0-1900/$00.00 c© 0 – IOS Press and the authors. All rights reserved

2 V. Zamborlini et al. / Inferring Recommendation Interactions in Clinical Guidelines

based medicine (EBM). Clinical guidelines (CGs)accumulate and reflect knowledge on how to bestdiagnose and treat diseases in the form of a list ofrecommendations. Guidelines have a formal status;they are sanctioned by national and internationalhealth care organizations. Relatively recently, thereference documents of EBM have started to be re-grouped and published as CGs in their own right:evidence-based guidelines.In the last decades, Computer Interpretable

Guidelines (CIGs) emerged as formal representa-tions of CGs. They can have distinct benefits overpaper-based CGs in that they increase flexibility,minimize errors, and generalize the use of guide-lines across institutions. CIGs are expressed indedicated languages such as GLIF [2], Asbru [16]or PROforma [24]. They are mainly designed topromote the execution of guidelines, i.e. to applythem to patient data for supporting diagnosticsor treatment plans. Hereafter we refer to CIGs asguidelines and to CGs as paper-based guidelines.By definition, CGs address specific clinical sit-

uations, i.e. situations around a single disease.Following the same principle, guideline languageswere defined to represent (and execute) one CGper time. They are not flexible enough to supportcases where (parts of) multiple guidelines need tobe combined. This is most problematic in the caseof multimorbidity; situations where a patient suf-fers from multiple diseases at the same time. Withdemographic changes related to aging, and the in-crease of chronic diseases, multimorbidity is be-coming more frequent. A Dutch study shows thatnot only the number of chronic diseases doubledin twenty years (1985-2005), the proportion of pa-tients with four or more chronic diseases tripledin the same period [25]. Applying CGs for dif-ferent diseases independently may lead to con-flicting situations that can be detrimental to thehealth condition of the patient. For instance, As-pirin is recommended as anti-platelets to patientsdiagnosed with Transient Ischemic Attack. On theother hand, Aspirin is not recommended for Duo-denal Ulcer patients because it increases the riskof bleeding.A solution for this problem is considered as an

important challenge for clinical decision supportsystems [23]. Some existing approaches for CIGsare unable to automatically detect conflicts forcombinations of guidelines [11,14]. They also can-not propose alternative measures that would re-

solve the conflicts. Other approaches [26] requireall the possible conflicts and their solutions to beavailable in a knowledge base. In general, the stud-ied approaches do not properly address the com-bination of more than two guidelines.

In earlier work [30], we introduced the Transition-based Medical Recommendations (TMR) model;a conceptual model that uses the concepts of rec-ommendation, transition, care action type and sit-uation type to represent clinical guidelines in amore flexible and expressive way. We had selecteda set of common conflicting situation to evaluatethe capability of detecting and solving of severalapproaches and identify the opportunities to com-plement them.

The goal of the work presented in this paperis to pave the way towards an automatic identi-fication of common potential conflicts or interac-tions that can happen when merging guidelines.We address the interactions that can be tackledindependently of particular contexts of patients,taking into account the general context describedin the guidelines and eventually complemented byMedical Background Knowledge (MBK) (Section1 paragraph 6). This paper extends TMR to al-low for the automatic inference of interactions be-tween recommendations (TMR4I). The model al-lows us to classify different types of interactions(conflicts, duplicates, repetitions), and show possi-ble alternative solutions. We implement the mod-els using OWL, SWRL and SPARQL, and builda web application that allows for (semi)automaticdetection of internal and external interactions be-tween recommendations. This allows us to detectcases that require special attention from expertswhen two or more guidelines are combined for amultimorbidity patient; and provide suggestionsfor improvements.

The use of Semantic Web representation lan-guages allows us not only to infer and detect theseissues, they also allow us to tap into the enor-mous amount of relevant clinical information pub-lished as Linked Data, e.g. biomedical terminolo-gies, drug databases, symptoms and side effectdatabases.

We evaluate TMR4I by applying it to two casestudies of combined guidelines taken from alter-native approaches in the literature [26,11,14]. Thefirst is a merge of two guidelines, one on DuodenalUlcer (DU) and the other on Transient IschemicAttack (TIA). The second merges three guidelines

V. Zamborlini et al. / Inferring Recommendation Interactions in Clinical Guidelines 3

on Osteoarthritis (OA), Hypertension (HT) andDiabetes (DB). These use cases also featured inour earlier work [30,31]. We show the added valueof our model by comparing it to the existing ap-proaches described in [26,11,14].The contributions of this paper are:– A formal representation of TMR4I, an exten-sion of TMR [30] that allows for the detectionof interactions between recommendations inmultiple clinical guidelines.

– Definition of “guideline-independent” rulesfor detecting internal and external interac-tions that allows for a more reusable and scal-able approach in the sense of number of in-teractions that can be semi-automatically de-tected and number of guidelines that can becombined.

– An implementation of TMR4I using Seman-tic Web languages (OWL, SWRL, SPARQL)that integrates with external knowledge fromthe Linked Data cloud (the Linked Data ver-sion of DrugBank [28,12]).

– A web application that allows for brows-ing (combined) guidelines and informs usersabout inferred interactions between recom-mendations, the type and source of each in-teraction, and offers potential solutions.

– An evaluation of the approach for two usecases, that compares the TMR4I model withexisting approaches that target the same com-bination of guidelines [26,11,14].

This paper is structured as follows: Section 2discusses related work. Section 3 introduces somepreliminaries, including the concepts of the TMRmodel that underly TMR4I. Section 4 presents theTMR4I extension of TMR as a formal model. Sec-tion 5 describes our case studies for multimorbid-ity. Section 6 and 7 present respectively the imple-mentation and the obtained results. Finally, Sec-tion 8 discusses the results and outlines futurework and Section 9 summarizes this work.

2. Related Work

This section presents works related to Computer-based Representations of Clinical Guidelines ingeneral, and Semantic Web-based ones in particu-lar. Moreover, we present approaches for address-ing multimorbidity and clinical knowledge avail-able on the Semantic Web.

2.1. Computer-Interpretable Clinical Guidelines

Several guideline description languages existthat are aimed at representing clinical knowledge:PROforma [24], GLIF [2], Asbru [16], etc. Themain focus of these languages is on guideline exe-cution, which makes them highly procedural andtargeted to modeling a specific case. They are lim-ited with respect to interoperability (guidelinescannot be combined), semantics (free text is of-ten used to describe conditions and actions) andreasoning power (e.g. the inference of interactionsand their solutions is not supported) [9,18].

At the same time clinical decision support sys-tems (CDSS) fall short in assisting healthcare pro-fessionals in defining treatment plans for multi-morbid patients [3]. There is a “lack of provision ofintegrated medical guidelines for multiple chronicdiseases within current CDSS” [3]. Combined withthe inflexibility of existing CIG languages, this in-dicates the necessity for new, more powerful andflexible formalisms or for adapting existing ones.

2.2. Semantic Web-based GuidelineRepresentations

Semantic Web technologies have already beeninvestigated for the representation of guidelines.In particular, OWL ontologies were used to en-hance the representation of clinical knowledge ofguidelines making it more comprehensible for com-puters. For instance, Pruski et al. [21] focused onthe description of care actions usually expressed innatural language to better personalize treatmentplans. Peleg et al. [17] proposed to reason withguideline action effects using OWL and abduc-tion. Abidi et al. [1] introduced the COMET (Co-morbidity Ontological Modeling & ExecuTion)system that provides decision support to handleco-morbid chronic heart failure and a trial fibril-lation. This framework relies on the capabilitiesof ontologies to represent guideline content andsemantic alignment techniques to merge clinicalpathways. The work of Isern et al. [10] stresses theuse of ontologies to enable the execution of guide-lines tailored to a well-identified clinical environ-ment. In [29], Yao et al. derive SWRL rules fromguidelines to dynamically adapt clinical pathways,represented in BPMN, according to clinical ac-tivities (contextual data and procedure actions).Hoekstra et al. [7] propose a lightweight ontology


for annotating evidence-based clinical guidelines(adopting the W3C PROV and biomedical-relatedvocabularies), connecting recommendations to un-derlying evidence, and representing recommenda-tions strength and evidence quality.Still, these approaches are not designed for ad-

dressing multimorbidity. In this work we proposeto use the Semantic Web for representing guide-line content to enhance the reasoning capabilitiesof computer systems; allowing a better detectionof interactions between guidelines in case of co-morbid patients. Secondly, it facilitates the inte-gration of external knowledge sources expressed inRDF. This allows us to reuse clinical knowledge;thereby, in the end, improving the quality of med-ical decisions.

2.3. Multimorbidity

In [30], we reviewed existing approaches formerging treatments plans or guidelines and wecategorized them into: (i) guideline-level verifica-tion, (ii) on-prescription verification, (iii) after-prescription verification and (iv) on-treatment-execution verification. As re-usability is one of ourmajor concerns, in this section we discuss threeapproaches in the first level [26,11,14].

Constraint Logic Programming Wilk et al. [26]describe guidelines as an activity graph. They useconstraint logic programming (CLP) to identifyconflicts that result from applying two CGs to thesame patient and propose mitigation alternatives.The temporal aspect is not considered, thus theapproach can only be applied to specific situa-tions (e.g. acute diseases diagnosed during a singlepatient-physician encounter). This approach alsoconsiders that all predicates use the same termi-nology and that they can have only two states(true or false). Although their approach allows theconflicts and solutions to be automatically iden-tified, it relies on the availability of knowledgebases associated with each guideline. It meansthat both conflicts and solutions need to be previ-ously defined in a Medical Background Knowledge(MBK) as guideline-dependent rules/constraints.Moreover, the solutions proposed to the conflictsare introduced using the same formalism of theoriginal recommendations, which means that theirapproach would still be applicable to the resultantguideline (merge plus extra recommendations) in

order to verify eventual new conflicts. However,they do not perform such verification to the pro-posed case study. Indeed, a considerable limitationfor allowing this feature is that detecting new con-flicts would require the introduction of new rules/-constraints in the MBK. Finally, automatically re-trieving clinical knowledge from existent reposito-ries such as drug-drug and drug-disease interac-tions, as well as combining more than two guide-lines are proposed as future work both in [26,27].However, to the best of our knowledge these fea-tures were not yet addressed.

OntoMorph The OntoMorph approach by Jafar-pour [11] defines a set of ontologies to represent:the guidelines (LKO - local knowledge ontology),the general domain (DKO - domain knowledge on-tology), the mappings between LKO and DKO(KMO - knowledge mappings ontology), and thedecision rules for concurrent execution of LKO,provided by domain experts (KPO - knowledgemorphing ontology). The ultimate objective is todefine a framework that can be used to mergeheterogeneous knowledge from the same domain.However, using several ontologies can potentiallyrise some extra challenges, like decidability (hav-ing rules for all decision cases), maintenance (man-aging the consequences of changing the ontolo-gies or mappings), and consistency (verifying con-tradictions between local or domain rules). Thissituation requires a laborious work from domainexperts to map LKO to DKO following a rig-orous and unambiguous process (all LKOs mustbe mapped in the exact same way). These ex-perts are also required to identify all possible in-teractions between LKOs and to provide the cor-respondent consistent decision rules as solutionsto the conflicts. Although these rules are neces-sary to generate conflict-free guidelines, they mustbe general enough to be applied to all combina-tions of LKOs. Some situations are hardly pre-dictable, for instance, pairwise analysis of rec-ommendations from guidelines will not allow de-tecting overdose of medication if it comes fromthe associations of three or more recommenda-tions. Another challenge from this approach comesfrom the ontology alignment technique. The es-tablishment of semantic relations (mappings) be-tween heterogeneous data can require many com-plex techniques, for instance, natural languageprocessing techniques would be necessary if tex-


tual information are provided describe conceptsof the LKO (e.g. in one LKO there are theblood pressure "higvalue:120", "lowvalue:80" and"unity:mmHg", while in the other LKO there is"value: 120/80mmHg").Jafarpour uses OWL and SWRL to represent

and merge the guidelines. The latter defines Con-straints (rules) as entities that relates pairs of in-teracting Tasks (actions). These manually createdrules are supposed to address cases where tasks(a) are identical, (b) should be executed simul-taneously, (c) are conflicting, (d) can reuse eachother‘s results, (e) have a temporal or sequentialconstraint between them, (f) can be combined to anew task and (g) their execution schedule dependson operational constraints for their simultaneousexecution.

Rule-Based Combinations Lopez-Vallverdu et al.[14] focus on Rules and Actions concerned withthe administration of drugs. They use a standardterminology called ATC (Anatomical Therapeu-tic Chemical Classification System for drugs1).Based on knowledge available in clinical guide-lines, they manually built “knowledge units” forpairwise combination of three diseases: Hyperten-sion, Diabetes Mellitus and Heart Failure. Knowl-edge units can be about the co-existence of in-compatible drugs (drug-drug interaction), the ex-istence of a drug incompatible to a disease (drug-disease interaction) and the absence of a drug nec-essary for a combination of diseases. Based onthese units, they manually built a minimal set ofcombination rules in the format pre-condition :condition → action. The pre-condition is a dis-ease, the condition is the presence or absence ofdrug recommendations for each disease, and theaction holds recommendations for adding, remov-ing or replacing drugs. Although it is not clearfrom the knowledge format whether it is limitedto two diseases, the strategy adopted by the au-thors for addressing the three aforementioned dis-eases is by pairwise combining them. Moreover,the manual identification of the interactions andtheir solutions is in itself a limiting factor for com-bining multiple guidelines. Their approach is im-plemented in a proprietary system for combiningtreatments.

1http://www.whocc.no/atc/

In summary, these approaches are not suffi-cient to meet the objective of increasing the abil-ity to handle multimorbidity, since they have im-portant limitations such as manual definition ofguideline-dependent rules/constraints for pairwisedetecting conflicts. Therefore, we pursue an ap-proach that allows us to evaluate a set of rec-ommendations, originating from multiple guide-lines, inferring and identifying the types of inter-actions amongst them; with minimal human inter-action. In this context, the reuse of existing medi-cal sources such as available in Linked Data is veryattractive.

2.4. Ontologies and Linked Data

Since one of the long term goals of our workis to provide means to completely describe careactions in a computer-interpretable formalism, aprimary problem to be considered is how to as-sure the normalization of the terminologies usedto describe the actions and conditions. If a wideuse of formalized care actions is expected, thentheir description need to be based on standard ter-minologies with an explicit semantic behind. Sev-eral standard terminologies exist (e.g. SNOMED-CT2, ATC, etc.) which can be used to describecare actions and conditions. In our approach, weassume that instances of the TMR model will becreated based on existing standard terminologiesor, when necessary, based on other publicly avail-able Ontologies from the biomedical domain. Forinstance, one of the ∼ 390 ontologies (∼ 5 mil-lions classes) published in BioPortal [40] or acces-sible via EHTOP [4]. Another assumption of ourapproach is that care action descriptions refer tothe source of information that generates the re-lated evidence. For instance, one transition canbe associated to evidence from clinical trials (e.g.those published in [41]) which details the condi-tions where it was observed. The potential of Se-mantic Web technologies for making clinical guide-lines and protocols more flexible is also recognisedin [22]. Since a lot of life science information isavailable as Linked Data, it becomes an importantresource for improving the quality of informationgenerated by clinical decision support systems. Ac-cording to Mannheim Linked Data Catalog [33]

2http://www.ihtsdo.org/snomed-ct


there are 85 datasets built based on Semantic Webtechnologies that can be used to provide large net-works of Linked Data for the Life Science domain.Some platforms were created to regroup datasetsand provide an unified way to query across re-sources using the W3C SPARQL query language(e.g. Bio2RDF [34] with more that ∼11 BillionRDF triples available and EBI-RDF [13]). In thiswork we selected one dataset relevant to evaluatethe proposed approach: DrugBank [35,28,12] (∼3milions RDF triples), but our approach is genericenough to use more datasets (e.g. Drugbank +Sider [38] (∼17 millions RDF triples), Clinical tri-als [37] (+25 milions RDF triples), Diseasome [36](+9K RDF triples), etc.).

3. Overview of the TMR Model

This section briefly summarizes a slightly ex-tended version of the TMR model, originally pre-sented in [30] where we investigated the core con-cepts required for representing recommendationswithin CGs. Figure 1 shows a UML diagram of theTMR model. The concepts and relations presentedas they were in the original version are depicted ingray shade to differ from the new concepts and re-lations changed or introduced in this work. Thosethat have a slash sign before their names are fur-ther defined by FOL (First Order Logics) formu-las (e.g. /similarTo). We consider the concepts asbeing atomic, since its compositionality is out ofscope of this work.

– A Guideline is an aggregation of Recommendations,whilst the latter is part of one Guideline. It can bea Single Disease Guideline, or it can be a ComposedGuideline, which is derived from the combination oftwo or more Guidelines.

– A Recommendation either recommends to pursue orrecommends to avoid (originally recommends and non-recommends) one Transition.

– A Transition is promoted by a single Care Action Type,which in turn can promote one or more Transitions.A Transition can be similar to or inverse to anotherone (definition is further provided).

– Situation Types can be Pre or Post-Situation Types inthe context of different Transitions.

– Every Transition have:

∗ one Transformable Situation Type through the rela-tion has transformable situation,

∗ one expected Post-Situation Type through the re-lation has expected post situation, and

∗ zero or more Non-Transformable Situation Typesthrough the relation have as filter condition.

For example, Table 1 presents the recommen-dation “If the patient’s temperature is over 37 de-grees and he/she is over 10 years old then reducethe temperature by administering aspirin” decom-posed into the TMR concepts.

We introduce the binary relations /similarToand /inverseTo between Transitions, which are re-quired for detecting Interactions. In this work weconsider a simple approach of comparing equal-ity among Pre and Post-Situation Types, thoughthese definitions would benefit from a richer def-inition of Situation Types and possible matchesamong them. Therefore, similar transitions (def.1.1) are those whose Pre-Situation Types are thesame, as well as the Post-Situation Types, but thatare promoted by different Care Action Types (oth-erwise they are the same transition). Two tran-sitions are inverse (def. 2.1) if the Pre-SituationType of one is equal to the Post-Situation Type ofthe other. Both relations similarTo and inverseToare symmetric (def. 1.2 and 2.2).(1.1) ∀t1, t2, sa, sb, ca1, ca2( Transition(t1)∧ Transition(t2) ∧ CareActionType(ca1)∧ CareActionType(ca2) ∧ promotedBy(t1,ca1)∧ promotedBy(t2,ca2) ∧ ca1 6= ca2∧ SituationType(sa) ∧ SituationType(sb)∧ hasTransformableSituation(t1,sa)∧ hasTransformableSituation(t2,sa)∧ hasExpectedSituation(t1,sb)∧ hasExpectedSituation(t2,sb) )↔ similarTo(t1,t2)

(1.2) ∀t1, t2 similarTo(t1,t2) ↔ similarTo(t2,t1)(2.1) ∀t1, t2, sa, sb( Transition(t1) ∧ Transition(t2)∧ SituationType(sa) ∧ SituationType(sb)∧ hasTransformableSituation(t1,sa)∧ hasTransformableSituation(t2,sb)∧ hasExpectedSituation(t1,sb)∧ hasExpectedSituation(t2,sa) )↔ inverseTo(t1,t2)

(2.2) ∀t1, t2 inverseTo(t1,t2) ↔ inverseTo(t2,t1)

In [30] we illustrated the applicability of TMRby describing the possible interactions among rec-ommendations. These interactions can be contra-dictory, optimizable or reflect alternative recom-mendations (redefined in Sect. 4). We advocatedthat the TMR concepts favor the detection of suchinteractions, which may require some attentionfrom experts when combining CGs due to comor-bidity. Moreover, we considered not all interac-tions are unwelcome (e.g. the recommendations toinverse transitions may be desirable and the alter-native ones are useful to avoid conflicts) althoughthey could still require attention (e.g. defining


Fig. 1. UML class diagram for the TMR Model.

Table 1TMR Concepts Summary.

Situation Type Represents a property and its admissible valuesTransformableSituation Type

Regards the situation that is expected to be changed Patient’s temperature is over 37 degrees

Non-TransformableSituation Type

Regards the situation that holds as filter condition Patient’s age is over 10 years old

Post-SituationType

Regards the situation that is expected to be achieved Patient’s temperature is below 37 degrees

Care Action Type Represents the action types that can be performed byhealth care agents in order to change a situation.

Administer aspirin

Transition Represents the possibility of changing a situationregarding a patient by performing a care action type.

Administering aspirin in patient over 10 yearsold reduces its temperature below 37 degrees

Recommendation Represents a suggestion to either pursue or avoid atransition promoted by a care action type.

which alternative recommendation is preferred).In the following section we extend the TMR modelfor the specific task of representing and detectinginteractions among recommendations.

4. The TMR4I Model

The TMR4I (Transition-based Medical Recom-mendations for [detecting] Interactions) model ismeant for detecting interactions among recom-mendations when addressing multimobidity. Inthis case more than one disease, originally ad-dressed in different guidelines, need to be takeninto account. Therefore, the recommendationscombined from the different guidelines may inter-act, e.g. presenting inconsistencies or being suscep-tible to optimization. The main concept in TMR4Iis the interaction, which can be internal, amongthe recommendations themselves, or with some ex-

ternal knowledge base holding e.g. clinical knowl-edge (e.g. overdose). In Sect. 4.1 we focus on theinternal interactions whilst the external ones areaddressed in Sect. 4.2. Moreover, since the tempo-ral/sequence aspects are still not addressed in thiswork, the interactions here defined are consideredeither time-independent or simultaneous.

4.1. Internal Interactions

Figure 2 illustrates in a graphical notation theTMR4I representation for the examples discussedin Table 2. Arrows connecting Recommendations(on the left hand side) to Transitions (on the righthand side) mean that the former recommends thelatter to be pursued or avoided. The blue arrowlabeled “do” means the transition is to be pur-sued, while the red one labeled “do not” means theopposite. Each transition regards a property (e.g.gastrointestinal bleeding) and the pre and post sit-


Table 2Interactions Summary.

Contradiction Interactions two recommendations that would lead to a conflict if recommendedtogether (i.e. they cannot be both followed at the same time)

Opposed recommendations to the same careaction

- Do not administer aspirin to avoid increasing the risk of gastrointestinalbleeding- Administer aspirin to handle inflammation

Opposed recommendations to similartransitions

- Do not adm. beta-blockers to avoid lowering blood pressure- Administer ACE inhibitor to lower blood pressure

Recommendations to inverse transitions - Administer ACE inhibitor to lower blood pressure- Administer midodrine to increase blood pressure

Repetition Interactions set of recommendations that are susceptible to optimizationRepeated recommendations to the same careaction

- Administer aspirin to reduce the risk of thrombus- Administer aspirin to relief pain- Administer aspirin to handle inflammation

Alternative Interactions set of recommendations that hold as alternatives.Repeated recommendations to the similartransitions promoted by different care action

- Administer aspirin to handle inflammation- Administer ibuprofen to handle inflammation- Administer naproxen to handle inflammation

Non-recommended transition whose inversetransition is recommended

- Do not administer aspirin to avoid increasing the risk of gastrointestinalbleeding- Adm PPI to decrease risk of gastrointestinal bleeding

uations are presented as small boxes connected byan arrow, where the box in the left hand side isthe transformable pre-stituation and the other oneis the expected post-situation. If there is a dot-ted box in the left hand side, it represents a filtercondition (non-transformable situation). Finally, adotted ellipse on the arrow represents the care ac-tion type that could promote the transition. Wedepict the interactions named by their three maintypes (Repetition, Contradiction and Alternative)connecting the interacting recommendations. Forexample, the third interaction (from top to bot-tom) is a Repetition Interaction among three rec-ommendations for different Transitions promotedby Administer Aspirin.Figure 3 presents an UML class diagram for the

TMR4I model. Elements presented in a gray-shademean they were previously introduced. The con-cept Recommendation is specialized into /Inter-nally Interacting Recommendation (def. 3) to de-note the ones that internally interacts with otherrecommendations.(3) ∀r, ∃i( Recommendation(r)∧ InternalRecommendationInteraction(i) ∧ relates(i,r) )↔ InternallyInteractingRecommendation(r)

The interaction relation is reified as /InternalRecommendation Interaction and /relates two ormore Recommendations in the context of a Guide-line. The latter concept is specialized according tothe classifications discussed in Table 2. They arefurther described and defined according FOL rules

(def. 4 to 9). In particular some interaction typesneed an extra rule to capture their cumulative be-haviour. For example, recommending 4 times Ad-minister Aspirin for different purposes should notbe considered as 6 different interactions betweeneach pair of recommendations, but one interactionamong 4 recommendations.

Repetition due to Same Action: when Transi-tions promoted by a same Care Action Type arerecommended to be pursued (def. 4.1). This inter-action is cumulative within a (merged) guideline,i.e. if a recommendation is related to two interac-tions of this type, they are the same interaction(def. 4.2)

(4.1) ∀g, r1, r2, t1, t2, ca( Guideline(g) ∧ Recommendation(r1)∧ Recommendation(r2) ∧ partOf(r1,g) ∧ partOf(r2,g)∧ Transition(t1) ∧ Transition(t2) ∧ r1 6= r2∧ recommendsToPursue(r1,t1)∧ recommendsToPursue(r2,t2) ∧ CareActionType(ca)∧ promotedBy(t1,ca) ∧ promotedBy(t2,ca) )→ ∃i( RepetitionDueToSameAction(i) ∧ relates(i,r1)∧ relates(i,r2) )

(4.2) ∀i1, i2, r1, r2, r3( RepetitionDueToSameAction(i1)∧ RepetitionDueToSameAction(i2) ∧ Recommendation(r1)∧ Recommendation(r2) ∧ Recommendation(r3) ∧r1 6= r3∧ r1 6= r2 ∧ r2 6= r3 ∧ relates(i1,r1)∧ relates(i1,r2) ∧ relates(i2,r2) ∧ relates(i2,r3) )→ i1 = i2

Contradiction Interaction due to Inverse Transi-tions: when two inverse Transitions are (simulta-neously) recommended to be pursued (def. 5);


Recommendations

Gastrointestinal Bleeding

high risk

AdministerAspirin

lowrisk

Avoid Bleeding do not

Blood Pressure

lowAdminister Beta-blockersnormal

Avoid Low Blood Pressure do not

Pain

no painAdminister Aspirinpain

Reduce Pain do

Blood Pressure

lowAdminister ACE inhibitornormal

Reduce Blood Pressure do

Transitions by Care Actions

Avoid Thrombi do

Risk of Thrombus

low riskAdminister Aspirin

medium risk

Protect Duodenum do


high risk

low risk

Administer PPI

contradiction

Inflammation

healedAdminister Aspirinunhealed

Heal Inflammation do

Inflammation

healedAdminister Ibuprofenunhealed


Inflammation

healedAdminister Naproxenunhealed


Blood Pressure

normalAdminister Midodrinelow

Increase Blood Pressure do

alternative

contradiction

contradiction

alternative

repetition

Fig. 2. Instance-schema for illustrating interactions amongrecommendations.

(5) ∀g, r1, r2, t1, t2( Guideline(g) ∧ Recommendation(r1)∧ Recommendation(r2) ∧ partOf(r1,g) ∧ partOf(r2,g)∧ Transition(t1) ∧ recommendstoPursue(r1,t1)∧ Transition(t2) ∧ recommendstoPursue(r2,t2)∧ inverseTo(t1, t2) )→ ∃i( ContradictionDueToInverseTransition(i)∧ relates(i,r1) ∧ relates(i,r2) )

Contradiction Interaction due to Same Action:when two Transitions promoted by the same CareActions Type are recommended to be pursued andthe other recommended to be avoided (def. 6);(6) ∀g, r1, r2, t1, t2, ca( Guideline(g) ∧ Recommendation(r1)∧ Recommendation(r2) ∧ partOf(r1,g) ∧ partOf(r2,g)∧ r1 6= r2 ∧ Transition(t1) ∧ Transition(t2)∧ recommendsToPursue(r1,t1) ∧ promotedBy(t1,ca)∧ recommendsToAvoid(r2,t2) ∧ promotedBy(t2,ca)∧ CareActionType(ca) )

→ ∃i( ContradictionDueToSameAction(i) ∧ relates(i,r1)∧ relates(i,r2) )

Contradiction Interaction due to similar Transi-tions: when two similar Transitions are one recom-mended to be pursued and the other recommendedto be avoided (def. 7);

(7) ∀g, r1, r2, t1, t2( Guideline(g) ∧ Recommendation(r1)∧ Recommendation(r2) ∧ partOf(r1,g) ∧ partOf(r2,g)∧ Transition(t1) ∧ Transition(t2) ∧ similarTo(t1, t2)∧ recommendsToPursue(r1,t1)∧ recommendsToAvoid(r2,t2) )→ ∃i( ContradictionDueToSimilarTransition(i)∧ relates(i,r1) ∧ relates(i,r2) )

Alternative Interaction due to similar Transitions:when similar Transitions are recommended to bepursued (def. 8.1). This interaction is cumulativewithin a (merged) guideline, i.e. if a recommen-dation is related to two interactions of this type,they are the same interaction (def. 8.2)

(8.1) ∀g, r1, r2, t1, t2( Guideline(g) ∧ Recommendation(r1)∧ Recommendation(r2) ∧ partOf(r1,g) ∧ partOf(r2,g)∧ Transition(t1) ∧ Transition(t2) ∧ similarTo(t1, t2)∧ recommendsToPursue(r1,t1)∧ recommendsToPursue(r2,t2) )→ ∃i( AlternativeDueToSimilarTransition(i)∧ relates(i,r1) ∧ relates(i,r2) )

(8.2) ∀i1, i2, r1, r2, r3( AlternativeDueToSimilarTransition(i1)∧ AlternativeDueToSimilarTransition(i2)∧ Recommendation(r1) ∧ Recommendation(r2)∧ Recommendation(r3) ∧ r1 6= r3 ∧ r1 6= r2∧ r2 6= r3 ∧ relates(i1,r1) ∧ relates(i1,r2)∧ relates(i2,r2) ∧ relates(i2,r3) )→ i1 = i2

Alternative Interaction due to Inverse Transi-tion: when two inverse Transitions are one recom-mended to be pursued and the other recommendedto be avoided (def. 9);

(9) ∀g, r1, r2, t1, t2( Guideline(g) ∧ Recommendation(r1)∧ Recommendation(r2) ∧ partOf(r1,g) ∧ partOf(r2,g)∧ Transition(t1) ∧ Transition(t2) ∧ inverseTo(t1, t2)∧ recommendsToPursue(r1,t1)∧ recommendsToAvoid(r2,t2) )→ ∃i( AlternativeDueToInverseTransition(i)∧ relates(i,r1) ∧ relates(i,r2) )

4.2. External Interactions

In this paper we address two types of Exter-nal Interactions based on the following informa-tion about Drug (types): (i) drugs that are knownto be incompatible to each other (e.g. Aspirin isincompatible with Ibuprofen) and (ii) drug (types)


Fig. 3. UML class diagram for the TMR4I Model- Internal Interactions.

that are grouped in a broader type/category ac-cording to the effects they are expected to promote(Aspirin and Tramadol (types) are grouped/sub-ssumed by Anti-platelets Type/Category).Figure 4 presents an UML class diagram for

the fragment of TMR4I model regarding ExternalInteractions. Elements presented in a gray-shademean they were previously introduced. The con-cept Recommendation is specialized into /Exter-nally Interacting Recommendation (def. 10) to de-note the ones that interacts with external infor-mation and/or other recommendations.(10) ∀r, ∃i( Recommendation(r) ∧

ExternalRecommendationInteraction(i) ∧ relates(i,r) )↔ ExternallyInteractingRecommendation(r)

The external interaction relation is reified as/External Recommendation Interaction and /re-lates one or more Recommendations of a Guidelinewith External Information. The two types of ex-ternal interaction aforementioned are representedas Incompatible Drug Interaction and AlternativeDrug Interaction. They are further described anddefined according FOL rules (def. 11 and 12). Fi-nally a Transition or a Care Action Type mayregard an External Information. In particular Acare action type may involve an External DrugType (e.g. Administer Aspirin involves Aspirin)and a Transition may regard an External DrugType (e.g. Reduce Pain regards Painkiller).

Incompatible Drug Interaction: when two rec-ommended (to be pursued) actions are associ-ated with external information about incompati-ble drugs (def. 11);

(11) ∀g, r1, r2, t1, t2, ca1, ca2, d1, d2( Guideline(g)∧ Recommendation(r1) ∧ Recommendation(r2)∧ partOf(r1,g) ∧ partOf(r2,g) ∧ r1 6= r2∧ Transition(t1) ∧ recommendsToPursue(r1,t1)∧ Transition(t2) ∧ recommendsToPursue(r2,t2)∧ CareActionType(ca1) ∧ CareActionType(ca2)∧ promotedBy(t1, ca1) ∧ promotedBy(t2, ca2)∧ ExternalDrugType(d1) ∧ ExternalDrugType(d2)∧ regards(ca1, d1) ∧ regards(ca2, d2)∧ incompatibleWith(d1, d2) )→ ∃i( IncompatibleDrugInteraction(i) ∧ relates(i,r1)∧ relates(i,r2) ∧ relates(i,d1) ∧ relates(i,d2) )

Alternative Drug Interaction: when a recommen-dation for administering a drug have a contradic-tory or a incompatible-drug interaction, and thetransition is associated with external informationabout the drug category corresponding to the ex-pected effect of the action, then other drugs be-longing to the same drug category can be sug-gested as alternatives, once they are not incom-patible with other recommended drugs (def. 12);

(12.1) ∀g, r1, i1, t1, ca1, d1, dc, d− alt( Guideline(g)∧ Recommendation(r1) ∧ partOf(r1,g) ∧ relates(i1,r1)∧ (ContradictionDueToSameAction(i1)∨ ContradictionDueToSimiliarTransition(i1)∨ ContradictionDueToInverseTransition(i1)∨ IncompatibleDrugInteraction(i1))∧ Transition(t1) ∧ recommendsToPursue(r1,t1)∧ CareActionType(ca1) ∧ promotedBy(t1,ca1)∧ ExternalDrugType(d1) ∧ regards(ca1,d1)∧ ExternalDrugType(dc) ∧ regards(t1,dc)∧ subssumes(dc,d-alt)∧¬∃r2, t2, ca2, d2( Recommendation(r2) ∧ partOf(r2,g)∧ Transition(t2) ∧ recommendsToPursue(r2,t2)∧ CareActionType(ca2) ∧ promotedBy(t2,ca2)∧ ExternalDrugType(d2) ∧ regards(ca2,d2)∧ incompatibleWith(d2,d-alt) ))


Fig. 4. UML class diagram for the TMR4I Model - External Interactions.

→ ∃i( AlternativeDrugInteraction(i) ∧ relates(i,r1)∧ relates(i,ca1) ∧ relates(i,d1) ∧ relates(i,d-alt) )

(12.2) ∀i1, i2, r1( Recommendation(r1)∧ AlternativeDrugInteraction(i1) ∧ relates(i1,r1)∧ AlternativeDrugInteraction(i2) ∧ relates(i2,r1) )→ i1 = i2

In the next section we illustrate the applicabilityof the TMR4I model and rules on two case studiesfor combining sets of guidelines.

5. Case Studies on Multimorbitidy

This section presents two conceptual experi-ments on detecting interactions for merged guide-lines, taken from the literature. They are meantto provide an implementation independent demon-stration of the TMR4I model applied to multimor-bidity; it also illustrates the improvements it offerscompared to the approaches introduced in Sect.2. The implementation of TMR4I using SemanticWeb languages, discussed in Sect. 6, serves as anevaluation of the conceptual model, and corrobo-rates the pen and paper exercise of this section.The two case studies are:

DU+TIA The first experiment, taken from Wilket al. [26], originally presented in [30], com-bines two guidelines, namely Duodenal Ulcer(DU) and Transient Ischemic Attack (TIA),adapted from original guidelines published by

the National Institute for Health and ClinicalExcellence, UK (NICE)3.

OA+HT+DB The second experiment combines 3guidelines for Osteoarthritis4 (OA), Hyper-tension5 (HT) and Diabetes6 (DB) taken from[11].

We make some adaptations to the original casestudies by:

i representing only the recommendations thatare relevant for studying the interactions andillustrating the approach;

ii simplifying the features that are not yet ad-dressed by the TMR/TMR4I models (e.g.time, quantities); and

iii adding some extra recommendations thatserve to better illustrate our approach.

In particular, when the original guidelines takenfrom the related work do not provide all informa-tion that we need in the TMR model, we makesome assumptions based on related paper-basedguidelines or common sense.

We divide the experiment in 3 steps: (1) Acquir-ing guideline data; (2) Combining guidelines anddetecting interactions and (3) Addressing unde-sired interactions and verifying the outcome. For

3http://www.nice.org.uk/4http://www.nhstaysideadtc.scot.nhs.uk/TAPG\%20html/

Section\%2010/osteoarthritis.htm5http://pathways.nice.org.uk/pathways/hypertension6http://pathways.nice.org.uk/pathways/diabetes


each step the results are presented in figures asinstance-schemas of the TMR4I model. They re-gard the results we would expect to obtain froman information system that implements the modeland rules described in the previous sections. Wealso present and compare with the original ap-proaches applied to each case study.

5.1. Case Study 1: DU + TIA

This section repeats a conceptual experimentpresented in [30], considering adaptations to theTMR4I model, with the aim of providing an imple-mentation for it. This experiment is originally pro-posed by Wilk et al. [26] on modeling and mergingthe guidelines for DU and TIA.

STEP 1: Acquiring guideline dataThe original DU guideline, taken from [26], is pre-sented in Fig. 5. This actionable graph is manu-ally created from the original paper-based guide-lines (as well as the one further presented). Thehighlighted recommendations are the relevant onesrepresented according to the TMR model, illus-trated as an instance-schema in Fig. 6. The actionStop Aspirin if used in the original guideline is rep-resented in the TMR DU guideline as a recommen-dation named Avoid Bleeding, which recommendsto avoid the transition promoted by the care ac-tion Administer Aspirin. The undesired transitioncan lead from the situation where gastrointestinalbleeding risk is Low to High. Other two actionsStart Eradication Therapy and Start PPI from theoriginal guideline are represented as the recom-mendations named Heal DU, describing that boththe care action Eradication Therapy can lead fromthe situation where DU is unhealed to healed whenH.Pylori is positive, and when H.Pylori is nega-tive instead the care action Administer PPI canpromote a similar transition.The original TIA guideline, taken from [26], is

presented in Fig. 7. The highlighted recommen-dations are the relevant ones represented accord-ing to the TMR model, illustrated as an instance-schema in Fig. 8. The actions Start Aspirin andStart Dipyridamole from the original guideline arerepresented as the recommendations named Re-duce Medium Risk VE (Vascular Events) and Re-duce High Risk VE. They recommend respectivelythe transition promoted by the care actions Ad-minister Aspirin, which leads from Medium Risk

of VE to Low Risk of VE, and the transitionpromoted by the care action Administer Dipyri-damole, which leads from High Risk of VE to LowRisk of VE.

STEP 2: Combining guidelines and detecting in-teractionsFor combining the two guidelines, Wilk et al. [26]uses CLP to identify and resolve conflicts in twoguidelines by consulting a MBK, locally defined,describing possibly conflicting actions/recommen-dations and their alternative solutions accordingto experts. For example, the MBK states that therecommendations Stop Aspirin if used and StartAspirin cannot coexist. Since this combination in-deed occurs in the combined version of guidelines,a conflict is identified.

Counterwise, the TMR4I model allows to sis-tematically identify interactions among recom-mendations by using the defined rules, as de-picted in Fig. 9. A contradiction interaction be-tween the recommendations Avoid Bleeding andReduce Medium Risk VE is identified since theyrecommend to pursue and to avoid transitions thatare promoted by the same Care Action Type. Analternative interaction is identified between thetwo recommendations named Heal Duodenal Ul-cer, since they recommend similar transitions7.The care actions and transitions are connected toan external drug dataset.

STEP 3: Addressing undesired interactions andverifying the outcomeIn order to address the identified conflict in Wilket al. [26] approach, the MBK provides two solu-tions: (i) substitute Aspirin by Clopidogrel; and(ii) combine Aspirin treatment with PPI. The au-thors choose the second option and introduced itin the merged guideline as Start PPI when therisk of stroke is elevated, and also excluded therecommendation Stop Aspirin if used in order toavoid the conflict. Although their outcome wouldbe verifiable by reapplying their approach, theauthors did not perform such verification. How-ever, in order to be effective, the possible conflictswould have to be included in the MBK. Finally,since their final goal was not to produce a genericcombined version of guidelines, but to prescribe a

7The recommendation Reduce High Risk VE is omittedfor sake of simplicity, since it does not have related inter-actions.


Fig. 5. DU guideline taken from [26].

DU Guideline Transitions by Care ActionsGastrointestinal Bleeding

high riskAdministerAspirinlow riskAvoid Bleeding do not

Duodenal Ulcer

healedEradication TherapyunhealedHeal Duodenal

Ulcer do H.Pyloripositive

Duodenal Ulcer

healedAdminister PPIunhealedHeal Duodenal

Ulcer do H.Pylorinegative

Fig. 6. DU guideline according to TMR Model.

Fig. 7. TIA guideline taken from [26].

TIA Guideline Transitions by Care ActionsVascular Events


mediumrisk

Reduce medium risk VE do

Vascular Events

low riskAdminister Dipyridamole

highrisk

Reduce high risk VE do

Fig. 8. TIA guideline according to TMR Model.

treatment for a specific patient, they proposed asolution that is applicable to a specific patient.In our approach both the proposed solutions

are introduced as recommendations, without ex-cluding the recommendation Avoid Bleeding. Theyare named Protect Duodenum and Reduce MediumRisk VE as depicted in Fig. 10 (the original rec-ommendations are presented in gray shade). Natu-rally, alternative interactions are derived betweenthe original recommendations and the new ones.On the one hand, Protect Duodenum recommends

a transition that reverse the undesired transitionrecommended to be avoided by Avoid Bleeding.On the other hand, the new recommendation Re-duce Medium Risk VE recommend a similar tran-sition. Moreover, a repetition interaction can alsobe identified between the recommendations Pro-tect Duodenum and Heal DU, since they both rec-ommends the same Care Action Administer PPI.

Therefore, the combined DU-TIA guideline thatwe produced does not eliminate the original con-flict but allows it to be avoided by introducing one


DU + TIA Guideline

contradiction

alternative

Vascular Events


mediumrisk


Duodenal Ulcer

healedEradication TherapyunhealedHeal Duodenal

Ulcer do H.Pyloripositive


high riskAdministerAspirinlow riskAvoid Bleeding do not


Duodenal Ulcer

healedAdminister PPIunhealedHeal Duodenal

Ulcer do H.Pylorinegative

Fig. 9. Combined DU+TIA guideline according to TMRmodel with detected interactions.


high riskAdministerAspirinlow riskAvoid Bleeding

Heal Duodenal Ulcer

Heal Duodenal Ulcer

Vascular Events


mediumrisk

Reduce medium risk VE

contradiction

alternative

Protect Duodenum do

Vascular Events

low riskAdminister Clopidogrel

mediumrisk


Duodenal Ulcer

healedAdminister PPIunhealeddo H.Pylori

negative

Duodenal Ulcer

healedEradication Therapyunhealeddo H.Pylori

positive

do

do not



high risk low riskAdminister PPI

alternative

repetition

alternative

DU + TIA Guideline

Fig. 10. Combined DU+TIA guideline according toTMR model, including the new recommendations anddetected interactions.

or more alternative recommendations. Actuallythe recommendation Avoid Bleeding is not elimi-nated since it is a restriction that holds for DU pa-tients regardless what else disease they could have.Indeed, the resultant guideline is both (i) appli-cable to many patients and (ii) liable to furthercombination with other guidelines or treatmentsthat the patient already follows. Finally, the in-ternal interactions are identified by relying on thedetailed semantics for the recommendations andguideline-independent rules without (necessarily)consulting a MBK.

5.2. Case Study 2: OA + HT + DB

This section presents a new experiment takenfrom Jafarpour’s thesis [11], on combining threeguidelines, namely OA, HT and DB. This is meantfor illustrating the ability to (i) identify internalinteractions between more than two recommenda-tions, combining more than two guidelines, and (ii)identify external interactions. We first introducehow the experiment was originally addressed andthen we discuss and compare with our approach.In Jafarpour’s approach the guideline data is

acquired by manually instantiating an ontology.Then constraints are manually created to resolveconflicts between tasks (recommendations) fromtwo guidelines, as illustrated in Fig. 11. The three

tasks regarding the administration of Ibuprofen orNaproxen are considered to be in conflict with thehypertension pathway, since these drugs may in-crease the blood pressure, according to experts.The constraints named conflict 1, 2 and 3 areintroduced suggesting the replacement of thesedrugs by Tramadol or similar. The role of theseconstraints is to interfere in the execution of theguidelines, i.e. when a task that is to be executedhas one of these constraints associated, then in-stead of executing the task, the constraint instruc-tion will be followed. In this example, the instruc-tions are for substituting the task.

The goal in Jafarpour’s approach (further dis-cussed in Sect. 8) is to produce a reusable pair-wise combination of guidelines, such that severalpairwised combined guidelines can be executed to-gether to handle multi-morbidity. Besides the com-bination of OA+HT, the approach is applied tocombine OA+DB and HT+DB such that the threeof them can be executed together. However, thesolution proposed in this example introduces arepetition interaction, according to the definitionpresented in 4.1, since Tramadol is recommendedfour times: in order to address the aforementionedconflicts and also to address another conflict be-tween OA and DB recommendations, where Tra-madol is recommended to replace Aspirin as anti-thrombotic. Since recommending the same drug


Fig. 11. The instantiation of the CPG-KPO and the partsof the OA and HT guidelines that participate in the merge,taken from [11].

more than once may lead to overdose, it requiresattention from the experts. Jafarpour’s approachdoes not detect the interaction we just mentioned.

STEP 1: Acquiring guideline dataHereby we present our solution for merging guide-line for OA+HT+DB (for sake of simplicity figuresare omitted for this step). First, we represent foreach guidelines the recommendations that are in-volved in the conflicts detected by Jafarpour. Thenthe effects that must be avoided for each disease,which are the reason for the conflicts, are explic-itly represented as recommendations within eachguideline (if this information is not yet available).For instance for detecting the aforementioned con-flict the recommendation Avoid High Blood Pres-sure promoted by Administering Aspirin is explic-itly introduced in the HT guideline. Avoid Bleed-ing and Avoid High Blood Sugar Level are intro-duced respectively in OA and DB guidelines. Al-though this resembles the manual identification ofthe interactions, it is actually not the case. Oncethis information is available as part of the guide-line, it can be reused to derive many interactions.The care actions and transitions are connected toan external drug dataset.

STEP 2: Combining guidelines and detecting in-teractionsThe recommendations from the original guidelinesare reused in order to create a merged guideline,

as depicted in Fig. 12 according to the TMR4Imodel, and interactions manually identified by Ja-farpour can be derived: (i) Administer ibuprofento relief pain from OA contradicts Do not admin-ster ibuprofen to avoid increase the blood pres-sure from HT; (ii) Administer thiazide to lower theblood pressure from HT contradicts Do not admin-ster thiazide to avoid increase the level of bloodsugar from DB; and (iii) Administer aspirin tolower the risk of thrombus from DB contradicts Donot adminster aspirin to avoid increase the riskof gastro-intestinal bleeding from OA. Moreover,and alternative interaction is identified among therecommendations Avoid High Blood Pressure andReduce Blood Pressure since the latter reverse theundesired effect according to the former.

STEP 3: Addressing undesired interactions andverifying the outcomeIn order to address the contradiction interac-tions, two solutions proposed by Jafarpour areintroduced as recommendations (one of the so-lutions, regarding reducing the quantity, is notaddressed, since it is out of scope of this work)and more interactions are derived, as depicted inFig. 13 (the original recommendations are pre-sented in gray shade). (i) an alternative interac-tion is derived between recommendations for Re-duce Pain, (ii) another alternative interaction isderived between the three Avoid Thrombi recom-mendations and (iii) a repetition interaction isderived among the introduced recommendationsfor Administering Tramadol. Moreover, an extrarecommendation Avoid Thrombi promoted by Ad-minister Clopidogrel is added in order to betterillustrate the cumulative alternative interaction.Since the resultant guideline can be further veri-fied by applying the same method,

In addition, external interactions can also beidentified besides the internal ones, once the orig-inal care actions and transitions are connectedto an external drug dataset. Particularly in thiscase study, the recommendations Avoid Thrombiand Reduce Pain have an Incompatible Drug In-teraction since the drug types Aspirin and Ibupro-fen, associated to the respective care actions, in-teracts to each other. Moreover, Alternative DrugInteractions can be identified as possible solu-tion for the recommendations involved in an un-desired interaction. For example, the recommen-dation Avoid Thrombi, being involved in contra-


DB

HT

OA

OA+HT+DB Guideline

alternative

contradiction

contradiction

contradiction


high risk

AdministerAspirin

lowrisk


Blood Pressure

highAdminister Ibuprofennormal

Avoid High Blood Pressure do not

Blood Sugar Level

highAdminister Thiazidenormal

Avoid High Blood Sugar Level do not

Pain

no painAdminister Ibuprofenpain

Reduce Pain do

Blood Pressure

lowAdminister Thiazidehigh



Avoid Thrombi do

Risk of Thrombus


medium risk

Fig. 12. Instance-schema illustrating the OA+HT+DBguideline and three contradictory interactions derived.

contradiction

repetition

DB

HT

OA


high risk

AdministerAspirin

lowrisk


Blood Pressure

highAdminister Ibuprofennormal

Avoid High Blood Pressure do not

Blood Sugar Level

highAdminister Thiazidenormal

Avoid High Blood Sugar Level do not

Pain

no painAdminister Ibuprofenpain

Reduce Pain do

Blood Pressure

lowAdminister Thiazidehigh


Risk of Thrombus


medium risk

Avoid Thrombi do


contradiction

alternative

contradiction


no painAdminister Tramadolpain

Reduce Pain do

Risk of Thrombus

low riskAdminister Tramadol

somerisk

Avoid Thrombi do

Risk of Thrombus

low riskAdminister Clopidogrel

somerisk

Avoid Thrombi do

OA+HT+DB Guideline

alternative

alternative

Fig. 13. Instance-schema illustrating new recommenda-tions and derived interactions for OA+HT+DB guide-line.

diction and incompatible drug interactions, has al-ternative drugs (e.g. Epoprostenol) that can beretrieved according to the drug category (anti-platelet) related to the pursued transition (reducerisk of thrombus).By performing this conceptual experiment we

demonstrate that the TMR4I approach allows amore systematic detection of internal interactionsboth after combining guidelines and after intro-ducing new recommendations in order to addressthe interactions. This approach favor the autom-atization of this task, in contrast with Jafarpourapproach that requires ad-hoc interaction rules.

6. Evaluation

This section provides an evaluation for theTMR/TMR4I models and rules by implement-ing them using Semantic Web representation lan-

guages. The goal is to use the Semantic Web toprovide a technical solution that applies our con-ceptual approach into a concrete Web application,available online8. We instantiate the case studiesdiscussed in Sect. 5 as individuals in our knowledgebase. We then use Semantic Web inferencing toautomatically derive the interactions we predictedthe TMR4I model would infer.

The implementation itself is presented in Sect.6.1, and Sect. 6.2 focuses on the instantiation ofthe case studies.

6.1. Implementation of the TMR Models

The TMR and TMR4I models have a verystraightforward implementation in OWL 2. To

8See http://guidelines.hoekstra.ops.few.vu.nl/, andGitHub https://github.com/Data2Semantics/guidelines/

tree/v1.0.


guarantee the content independence of the model(cf. [26]), the schema information (OWL2, rules)and queries are generic and specified indepen-dently from the representation of individual guide-lines. The latter are represented as individuals –instances – in the triple store.All unary predicates in the rules 1.1 through

12.2 in Sect. 3 and 4 are classes, and all bi-nary predicates are either object properties, or anequivalent from the OWL and RDFS vocabulary.The class and property subsumption hierarchy, do-main and range restrictions, as well as cardinalityconstrains follow the structure of the UML classdiagrams in figures 1, 3, and 4.Not all rules can be translated to OWL restric-

tions because of the tree-model property that un-derlies the SROIQ(D) description logic of OWL 2[8]: an OWL class restriction can only fan out toform a tree shape of which the branches do notconnect. This essentially means that two variablesintroduced in the antecedent of a rule, that aretransitively connected through predicates in theantecedent, cannot be connected by a predicate inthe consequent of the rule. For example, one can-not represent the fact that the four legs of a ta-ble must touch the same floor (unless only a singlefloor exists in the universe).There are partial workarounds for this limita-

tion by using property chains [6,5], but for allpractical purposes we implemented the restric-tions that couldn’t be expressed in OWL2 using acombination of SPARQL queries and the StardogRules Syntax,9 an alternative to the more stan-dard SWRL.10 Stardog Rules Syntax is a SWRLimplementation that uses SPARQL syntax [19,20].It has the advantage of being more readable, butalso more expressive in that it allows the use ofSPARQL niceties such as filters, property pathsand blank nodes. Stardog is a triple store that sup-ports OWL2, SWRL and integrity constraints rea-soning at query time.11Another drawback of OWL2 and SWRL-style

inferencing is that these allow only for inferringthe existence of unnamed entities implicitly. Forinstance, if we know that two transitions are re-lated, we cannot assert an interaction-individualthat relates the respective recommendations. To

9See http://docs.stardog.com/owl2/.10See http://www.w3.org/Submission/SWRL/11See http://www.stardog.com.

overcome this limitation, we programmatically in-troduce these individuals in our knowledge baseusing SPARQL select queries.

The following sections illustrate the implemen-tation of our model. We first give an example in-stantiation of a recommendation, followed by adiscussion of the rules for internal interactions,and those for external interactions (ones that in-volve external knowledge, or MBK). For reference,listing 1 presents the prefixes used in the code ex-amples.

6.1.1. Example RecommendationThe listing 2 provides a short example of an in-

stantiation of the TMR4I model for the recommen-dation :RecDB-AntiThrombotic to ‘avoid thrombi’that is part of the Diabetes Mellitus guideline(:CIG-DB). It promotes a transition :T4 that aimsto lower the risk of a thrombus from medium tolow by means of the :ActAdministerAspirin careaction type. This care action is linked to the ex-ternal identifier DB00945 for the drug (Aspirin) inDrugBank, through the tmr4i:involves relation.

6.1.2. Implementation of Internal InteractionsIn our knowledge base, the example recommen-

dation in listing 2 does not exist in isolation:there are several more guidelines, recommenda-tions, transitions and care action types (as dis-cussed previously). In TMR4I, the internal inter-actions between recommendations are derived fol-lowing the FOL rules in Sect. 4.1. The implemen-tation of these rules is by means of a SPARQL se-lect query that retrieves every two recommenda-tions that matches the pattern in those rules:

– there exists a tmr4i:simlarToTransition be-tween the transitions recommended by tworecommendations (see listing 3, the otherrules are similar); or

– there is a tmr4i:inverseToTransition be-tween them; or

– the transitions are tmr4i:promotedBy thesame care action.

Then, a SPARQL insert is executed for each pairof recommendations, creating a new individual oftype tmr4i:InternalInteraction and relating it tothem12.

12This is necessary to work around the limitations ofSWRL and OWL with respect to the assertion of new in-dividuals.


1 PREFIX tmr4i: <http://guidelines.data2semantics.org/vocab/>

2 PREFIX owl: <http://www.w3.org/2002/07/owl#>

3 PREFIX drugbank: <http://wifo5-04.informatik.uni-mannheim.de/drugbank/resource/drugbank/>

4 PREFIX drugCat: <http://wifo5-04.informatik.uni-mannheim.de/drugbank/resource/drugcategory/> .

5 PREFIX drug: <http://wifo5-04.informatik.uni-mannheim.de/drugbank/resource/drugs/>

Listing 1: Prefixes used in SPARQL queries (and Turtle examples).

1 # The Guideline for Diabetes

2 :CIG-DB

3 a tmr4i:Guideline ,

4 owl:NamedIndividual ;

5 rdfs:label "CIG for Diabetes Mellitus"@en .

67 # A Recommendation

8 :RecDB-AntiThrombotic

9 a tmr4i:Recommendation ,


11 rdfs:label "Avoid thrombi"@en ;

12 tmr4i:partOf :CIG-DB ;

13 tmr4i:recommendsToPursue :T4 .

1415 # A transition

16 :T4 a tmr4i:Transition ,


18 tmr4i:promotedBy :ActAdministerAspirin ;

19 tmr4i:hasExpectedPostSituation

20 :SitLowRiskThrombus ;

21 tmr4i:hasTransformableSituation

22 :SitMediumRiskThrombus .

2324 # The care action type that promotes the transition

25 :ActAdministerAspirin

26 a tmr4i:CareActionType ,


28 rdfs:label "Administer Aspirin"@en .

29 # Link to a DrugBank identifier

30 tmr4i:involves drug:DB00945 .

Listing 2: Excerpt of the TMR4I-based represen-tation of a recommendation to ‘Avoid thrombi’.

The classification of the interaction type is againdone via SPARQL Rules. The code on Listing 4presents the implementation of the FOL rule 8.1defined in Sect. 4.1: if a tmr4i:InternalRecommen-

dationInteraction relates two tmr4i:Recommen-

dation instances that both tmr4i:recommendsTo-

Pursue two transitions, which have a tmr4i:simi-

larToTransition relation, then the interaction isof the type tmr4i:AlternativeDueToSimilarTran-sition. The other FOL rules have similar imple-mentations.As one can see, these rules are generic; they only

use vocabulary of the TMR and TMR4I models

1 [] a rule:SPARQLRule;

2 rule:content """

3 IF {

4 ?ca1 a tmr4i:CareActionType,

5 owl:NamedIndividual .

6 ?ca2 a tmr4i:CareActionType,


8 ?sa a tmr4i:SituationType,


10 ?sb a tmr4i:SituationType,


12 ?t1 a tmr4i:Transition,


14 ?t2 a tmr4i:Transition,


16 ?t1 tmr4i:hasExpectedPostSituation ?sb .

17 ?t2 tmr4i:hasExpectedPostSituation ?sb .

18 ?t1 tmr4i:hasTransformableSituation ?sa .

19 ?t2 tmr4i:hasTransformableSituation ?sa .

20 ?t1 tmr4i:promotedBy ?ca1 .


22 ?ca1 a owl:NamedIndividual .

23 ?ca2 a owl:NamedIndividual .

24 FILTER (?ca1 != ?ca2 && ?t1 != ?t2)

25 } THEN {

26 ?t1 tmr4i:similarToTransition ?t2 .

27 }""" .

Listing 3: Stardog rule for inferringtmr4i:similarToTransition relations.

themselves, and do not contain any specifics of theguidelines discussed in the previous sections.

6.1.3. Implementation of the ExternalInteractions

In order to connect to external knowledge basesit is not needed to explicitly instantiate the classExternal Information (from Fig. 4) but we can useexternal identifiers (URIs) directly in our knowl-edge base. For instance, the Turtle code in Listing2 connects a care action to the representation ofAspirin in DrugBank. Of course, these links haveto be added to the care actions by hand, but onlyonce; since care actions are reused across transi-tions, and transitions are reused across recommen-dations, which can be reused across guidelines (aswe will see below).


1 [] a rule:SPARQLRule;

2 rule:content """

3 IF {

4 ?i a tmr4i:InternalRecommendationInteraction;

5 tmr4i:relates ?r1, ?r2 .

6 ?r1 a tmr4i:Recommendation,


8 ?r2 a tmr4i:Recommendation,


10 ?r1 tmr4i:recommendsToPursue ?t1 .

11 ?r2 tmr4i:recommendsToPursue ?t2 .

12 ?t1 a owl:NamedIndividual .

13 ?t2 a owl:NamedIndividual .

14 ?t1 tmr4i:similarToTransition ?t2 .

15 FILTER(?r1 != ?r2)

16 } THEN {

17 ?i a tmr4i:AlternativeDueToSimilarTransition .

18 }""".

Listing 4: SPARQL rule for classifying AlternativeInteractions due similar transitions (FOL 8.1).

In addition to the direct link between a care ac-tion and a drug, we also use the drug category in-formation present in DrugBank: the category cap-tures the type of drug, i.e. what its desired effectis (drugs can belong to multiple categories at thesame time). TMR4I links drug categories to in-stances of the tmr4i:Transition class using thetmr4i:regards property (see listing 5).For scalability reasons, our knowledge base in-

cludes an excerpt of the Linked Data versionof DrugBank as published by the University ofMannheim13 that contains all drug names, drugcategories and the drug-drug interactions. The ex-cerpt is made available through our GitHub repos-itory.14In order to detect external interactions, we pro-

ceeded in a way similar to the internal ones.However, we used two specific queries for imple-menting the FOL rules 11 and 12.1 (defined inSect. 4.2) for incompatible drug interaction andalternative drug interaction. The first uses thedrug-drug interaction information from DrugBankto detect unique pairs of recommendations thatinteract because the care actions that promotethe recommended transitions involve drugs thatform an interaction pair on DrugBank (listing6). As with the internal interactions, the system

13See http://wifo5-03.informatik.uni-mannheim.de/

drugbank/14See https://github.com/Data2Semantics/guidelines/

blob/v1.0/src/data/drugbank_small.nt.

1 :ActAdministerAspirin

2 a tmr4i:CareActionType ;


4 :ActAdministerClopidogrel

5 a tmr4i:CareActionType ;


78 :T2 a tmr4i:Transition ;

9 tmr4i:regards drugCat:anti-bacterialAgents .

10 :T3 a tmr4i:Transition ;

11 tmr4i:regards drugCat:proton-pumpInhibitors .

Listing 5: Example connections to DrugBank.

1 SELECT distinct ?r1 ?r2 ?d1 ?d2

2 WHERE {

3 ?r1 tmr4i:partOf ?g;

4 tmr4i:recommendsToPursue ?t1.

5 ?r2 tmr4i:partOf ?g;

6 tmr4i:recommendsToPursue ?t2.



9 ?ca1 tmr4i:involves ?d1 .


11 ?d1 drugbank:interactsWith ?d2 .

1213 FILTER(?r1 != ?r2 && ?d1 != ?d2 && ?ca1 != ?ca2 )

14 # Avoid duplicates

15 FILTER NOT EXISTS {

16 ?iir a tmr4i:IncompatibleDrugExternalInteraction .

17 ?iir tmr4i:relates ?r1 .

18 ?iir tmr4i:relates ?r2 .

19 }

20 }

Listing 6: SPARQL query for selecting the pairsof recommendation that fit the pattern to be re-lated by an external incompatible drug interaction(FOL 11).

then asserts corresponding new instances of thetmr4i:IncompatibleDrugExternalInteraction classfor each pair.

The second tries to identify non-interactingalternative drugs based on the drug categoryspecified for transitions that cause a contradic-tion or incompatibility between recommendations(listing 7). The result of this SPARQL queryis a reified relation between a recommendation,a care action, a drug and the alternative drugfound. As with the internal interactions, the sys-tem then asserts corresponding new instances ofthe tmr4i:AlternativeDrugExternalInteraction

class for each unique reified relation.


1 SELECT DISTINCT ?rec ?ca ?d ?dALT

2 WHERE {

3 {

4 {?i a tmr4i:ContradictionDueToSameAction .}

5 UNION

6 {?i a tmr4i:ContradictionDueToInverseTransition .}

7 UNION

8 {?i a tmr4i:ContradictionDueToSimiliarTransition .}

9 UNION

10 {?i a tmr4i:IncompatibleDrugExternalInteraction .}

11 }

12 ?i tmr4i:relates ?rec .

13 ?rec tmr4i:recommendsToPursue ?t;

14 tmr4i:partOf ?g .

15 ?t tmr4i:promotedBy ?ca .

16 ?ca tmr4i:involves ?d .

17 ?t tmr4i:regards ?dc .

18 ?dALT drugbank:drugCategory ?dc .

1920 FILTER (?d != ?dALT) .

21 # Filter incompatible drugs


23 ?dALT drugbank:interactsWith ?d2 .



26 ?rec2 tmr4i:recommendsToPursue ?t2 .

27 ?rec2 tmr4i:partOf ?g .

28 FILTER(?dALT != ?d2)

29 }

30 # Avoid duplicates


32 ?iir a tmr4i:AlternativeDrugExternalInteraction .

33 ?iir tmr4i:relates ?rec .

34 ?iir tmr4i:relates ?ca .

35 ?iir tmr4i:drug ?d .

36 ?iir tmr4i:alternative_drug ?dALT .

37 }

38 }

Listing 7: SPARQL query for selecting the recom-mendations that fit the pattern to be related by anexternal alternative drug interaction (FOL 12.1).

On doing so we are able to infer the existenceof external interactions and classify them accord-ing to the FOL definitions we provided, preservingtheir structure. This is a limited but important im-provement on allowing the identification of drug-drug incompatibilities as well as alternative drugswithout human intervention (except for linking thecare actions and transitions to DrugBank).

6.2. Populating the Knowledge Base

This section briefly discusses how the knowledgebase is populated. Because all necessary knowl-edge for inferring the interactions are part of the

schema of the knowledge base, the representationof a guideline is only at the instance level. In list-ing 2 we already provided an example of what aninstantiation of the model for a single recommen-dation looks like. At the moment this is still amanual process: all guidelines, recommendations,transitions and care actions are modeled using aneditor of choice.In order to repeat the experiment described in

the section on the two case studies (Sect. 5) weproceed as follows:STEP 1 - All the data for the original guidelines

is manually entered. The care actions and transi-tions are linked to the corresponding drug typesand categories from DrugBank. We aim to facili-tate this step with a simple web-based user inter-face.STEP 2 - To simulate the merging of the guide-

lines, a new merged guideline is created that du-plicates recommendations from the original guide-lines, and directly reuses the information abouttransitions and care actions contained in them. Asa result, transitions and care actions are not du-plicated across guidelines, but shared. Once theSPARQL-based inferences are run, the internaland external interactions are created and classifiedwithin the merged guidelines.STEP 3 - In order to address contradictory in-

teractions, alternative recommendations can be in-troduced to the merged guidelines. Again, once therules are executed, eventual new interactions arederived.Finally, in order to verify the cumulative inter-

actions, a SPARQL query retrieves recommenda-tions whose interactions relate more than two rec-ommendations. In other words, it retrieves the rec-ommendations whose interactions are related byowl:sameAs, the type of interaction, including pur-sued or avoided transition, pre and post situationsand care action.Given the definitions of which and how inter-

actions can be identified according to the TMR4Imodel, we present in the next section an approachof how to apply this model in a multimorbiditycase study.

7. Results

This section brings the results we obtained byinstantiating the two case studies using the im-


plementation described in the preceding section.The results are presented in tables that presentthe data retrieved after each step proposed in theprevious sections. We also developed a Web-basedapplication that allows to explore both the orig-inal guidelines and the combined ones, and theirassociated recommendations, transitions and careactions.15

Web-based Guideline Explorer The Web-basedguideline explorer is a jQuery + Python Flask ap-plication that uses SPARQL queries to interactwith the Stardog triple store in which the TMR4Imodel and guideline representations reside.16Figure 14 shows a screen-shot of the application

presents on the left hand side the combined guide-line for OA+HT+DB and its interactions. Whenthe recommendation Avoid Thrombi by Adminis-tering Aspirin is selected, its details are presentedon the right hand side, including the internal andexternal interactions inferred and explained in ahuman readable format. By positioning the cursorover an orange interaction, the other recommen-dation involved in that interaction is highlightedin orange on the left.In summary, this recommendation says (at the

bottom) that the transition for Reducing the Riskof Thrombus can be achieved by executing thecare action Administer Aspirin and that there aretwo other, similar transitions recommended in thisguideline (T13 and T7, recommended twice). Thisrecommendation has (i) one contradiction interac-tion since a care action it recommends is also rec-ommended to be avoided; (ii) one cumulative alter-native interaction that involves two other recom-mendations with similar transitions; (iii) one ex-ternal drug interaction because Aspirin interactswith Ibuprofen according to DrugBank, and (iv)one alternative care action, namely Epoprostenol,that may alleviate this interaction according toDrugBank.

Step 1 - Data Acquisition Table 3 presents theoriginal recommendations introduced for DU andTIA guidelines (the other guidelines are omittedfor sake of simplicity).

15The web application is running at http://guidelines.hoekstra.ops.few.vu.nl

16See http://jquery.com, http://flask.pocoo.org, http://www.stardog.com.

Step 2 and 3 - Inferences Tables 4 and 5 presentthe results obtained after steps 2 and 3, for re-spectively the first (DU+TIA) and second use case(OA+HT+DB). The new recommendations andinteractions added in step 3 are highlighted inbold. Recommendations presenting more than oneinteraction appear in the table just once but fol-lowed by “empty” line(s) containing just the de-rived interaction type.

Cumulative Interactions Table 6 presents therecommendations related by cumulative interac-tions.

8. Discussion and Future Work

The outcomes from the two experiments shownthat different types of interactions can be (semi)automatically detected in multimorbidity cases byapplying TMR4I model, enriched with SemanticWeb features. In opposed to the original experi-ments reported in related work, in both case stud-ies the internal interactions are detect withoutneed for specific MBK or rules, but relying on thedetailed semantics provided for the recommenda-tions. Interactions with cumulative behavior (e.g.alternative) are detected among more than tworecommendations. Another important differenceis that the final outcome, including the mergedguidelines and the recommendations introduced toaddress the conflicts, can be verified again usingthe very same approach. Finally, DrugBank is usedfor automatically detecting interactions that couldnot be derived from the data comprised in theguidelines.

In what follows we provide a comparison be-tween our approach and the others that we anal-ysed in Sect. 2, summarized in Table 7:Core concepts. Although the analysed approaches

also use rules/constraints to identify the interac-tions among actions and to propose solutions, theydo not support a guideline-independent detec-tion of interactions either because the formalismadopted (Wilk [26], Lopez-Vallverdu [14]) is notexpressive enough or the model (Jafarpour [11])does not favour the reuse of rules. For instance,to detect the inconsistency between the recom-mendations Administer Aspirin & Avoid Adminis-tering Aspirin a context-specific rule is proposed.However, for other substances (e.g. if there is in-


Fig. 14. Screenshot of the live Web application, running at http://guidelines.hoekstra.ops.few.vu.nl.

Table 3Recommendations for original DU and TIA guidelines.

guide-line

Recommendation Mode pre Situation post Situation Care Action

DU Avoid bleeding Avoid Risk of GastrointestinalBleeding is Low

Risk of GastrointestinalBleeding is High

AdministerAspirin

DU Heal duodenal ulcer Pursue Ulcer is unhealed Ulcer is healed Administer PPIDU Heal duodenal ulcer Pursue Ulcer is unhealed Ulcer is healed Eradication

TherapyTIA Reduce high blood

pressurePursue Risk of Thrombus is High Risk of Thrombus is Low Administer

DipyridamoleTIA Reduce medium

risk VEPursue Risk of Thrombus is Medium Risk of Thrombus is Low Administer

Aspirin


Table 4DU-TIA - Recommendations and respective new interaction types detected after adding new recommendations (highlighted).

Recommenda-tion

Mode pre Situation post Situation Care Action Interaction Type

Avoid bleeding Avoid Risk of GastrointestinalBleeding is Low

Risk of GastrointestinalBleeding is High

AdministerAspirin

Contradiction Due ToSame ActionAlternative Due ToInverse Transition

Heal duodenalulcer

Pursue Ulcer is unhealed Ulcer is healed AdministerPPI

Alternative Due ToSimilar TransitionRepetition Due To SameAction

Heal duodenalulcer

Pursue Ulcer is unhealed Ulcer is healed EradicationTherapy

Alternative Due ToSimilar Transition

Protectduodenum

Pursue Risk of GastrointestinalBleeding is High

Risk of GastrointestinalBleeding is Low

Administer PPI Alternative Due ToInverse TransitionRepetition Due To SameAction

Reduce highrisk VE

Pursue Risk of Thrombus isHigh

Risk of Thrombus is Low AdministerDipyridamole

none

Reducemedium riskVE

Pursue Risk of Thrombus isMedium

Risk of Thrombus is Low AdministerAspirin

Contradiction Due ToSame Action


Reduce mediumrisk VE


Risk of Thrombus is Low AdministerClopidogrel


consistency with insulin administration instead ofaspirin) a new rule is required. In our approachwe rely on the TMR concepts to define guideline-independent rules for detecting some interactiontypes. For the same example, the inconsistencybetween Administer Aspirin and Avoid Adminis-tering Aspirin is detected by a generic rule thatcan be applied to other substances (e.g. Insulin).Representation of recommendations. In all the

studied approaches, the representation of the rec-ommendations occur in a textual form, collapsingboth concepts recommendation and action, andeventually pre and post-conditions. For instance,Administer Aspirin and Avoid Administering As-pirin are an example of opposed recommendationsabout the same action that are represented as twodifferent actions instead. The textual description isused to collapse the positive and negative aspectsof the recommendation and the action itself. Suchcharacteristic render it difficult the automatic de-tection of this type of interaction. We instead pro-pose a more structured representation of the rec-ommendation and its related concepts, the TMRmodel, which favor reasoning about it (althoughit also has textual labels for describing each con-cept). Moreover, the relate work impose the ac-tions to be in a sequence, when it is actually notalways the case. For instance, Avoid Administering

Aspirin is a recommendation that does not need tohappen before or after another one. Therefore, wechose a declarative representation of recommenda-tions in principle, while a sequence of recommen-dations will be an optional feature investigated asfuture work.Languages. As highlighted in this section, the

description languages used to represent guide-lines can introduce limitations to the reasoningprocess. This is the case for Wilk’s and Lopez-Vallverdu’s approaches that can only analysepairwise recommendations. Jafarpour’s approachadopted OWL+SWRL languages that are moreflexible to deal with multimorbidities. We concep-tually describe our approach using FOL and UMLwhat give us more liberty to select the computerlanguage later. We adopt OWL+SPARQL for im-plementation in order to benefit of available Se-mantic Web features.Medical background knowledge. Since we based

our implementation on Semantic Web technologiesand there are many semantically described knowl-edge bases available in the Web, we can evalu-ate our approach without requiring experts frombiomedical domain to create and validate propri-etary databases. The same was not observed in theother approaches, which either define it themselvesor do not use it.


Table 5OA+HT+DB - Recommendations and respective new interaction types detected after adding new recommendations(highlighted).

Recommendation Mode pre Situation post Situation Care Action Interaction TypeAvoid bleeding Avoid Risk of Gastrointestinal

Bleeding is LowRisk of GastrointestinalBleeding is High

AdministerAspirin


Avoid high bloodpressure

Avoid Blood Pressure isNormal

Blood Pressure is High AdministerIbuprofen

Alternative Due ToInverse TransitionContradiction Due ToSame Action

Avoid high bloodsugar level

Avoid Level of Blood Sugar isNormal

Level of Blood Sugar isHigh

AdministerThiazide


Avoid thrombi Pursue Risk of Thrombus isMedium

Risk of Thrombus is Low AdministerAspirin

Contradiction Due ToSame ActionIncompatible DrugExternal InteractionAlternative Drug ExternalInteractionAlternative Due To SimilarTransition


Risk of Thrombus is Low AdministerClopidogrel

Alternative Due To SimilarTransition


Risk of Thrombus is Low AdministerTramadol

Alternative Due To SimilarTransitionRepetition Due To SameAction

Reduce bloodpressure

Pursue Blood Pressure is High Blood Pressure isNormal

AdministerThiazide

Alternative Due ToInverse TransitionContradiction Due ToSame Action

Reduce pain Pursue Patient has Pain Patient has no Pain AdministerIbuprofen

Contradiction Due ToSame ActionIncompatible DrugExternal InteractionAlternative Due To SimilarTransition

Reduce pain Pursue Patient has Pain Patient has no Pain AdministerTramadol

Alternative Due To SimilarTransitionRepetition Due ToSame Action

Table 6Cumulative Interactions.

Recommend. T pre Situation post Situation Care Action Interaction TypeAvoidthrombi


Risk of Thrombusis Low

AdministerAspirin


Avoidthrombi



AdministerClopidogrel


Avoidthrombi



AdministerTramadol


Reuse of guideline knowledge. The laboriouswork of transforming paper-based guidelines intocomputer-interpretable guidelines (CIG) is a com-mon process for all approaches. However, con-cerning the reuse of guideline knowledge, we haveadopted a different strategy. While the analysedapproaches consider the whole CIG as one piece of

knowledge, we chose to split CIGs into two parts:One describing the recommendations and one de-scribing the actions. This choice allows for hav-ing shareable pieces of knowledge (i.e. actions andrules). Reusing the recommendations requires alittle bit more work because they need to be du-plicated from one existing guideline to the other


Table 7Comparison to related works.

Wilk[26] Jafarpour[11] Lopez-Vallverdu[14]

TMR4I Model

CoreConcepts

Actions & Rulesguideline-dependent

Tasks & Constraintsguideline-dependent

Actions &Rulesguideline-dependent

TMR + Interactions & Rulesguideline-INdependent

Repre-sentationof Rec-ommen-dations

TextualProcedural

TextualProcedural

TextualProcedural

StructuredDeclarative

Lan-guage

Workflow & CLP OWL + SWRL ProprietaryRule-basedNotation

Conceptual: UML + FOLImplemented: OWL +SPARQL

Medicalbackgroundknowledge

defined for specificguideline merges

No No DrugBank

Reuse ofCIGknowl-edge

Yes Yes, associatingrules to the originaltasks

Yes, byreusing therules

Yes, duplicatingrecommendations, sharingactions/transitions

Interac-tionsIdentification

Semi-automaticPairwise

ManualPairwise

ManualPairwise

Semi-automaticAmong several guidelines

Type ofInterac-tions

not applicable identical andconflicting actions,reuse of results,temporal/sequentialaspects

Drug-relatedconflicts

Internal: Contradiction,Alternative and Repetition;External: Incompatible Drugand Alternative Drug

SolutionsIdentification

Semi-automaticintroduced as CLPclauses and furtherincorporated toactionable graph.

ManualIntroduced as textin SWRLconstraints

ManualIntroduced asstandard textin a Rule

Manual: Introduced as TMRrecommendationsSemi-Automatic: Alternativedrugs retrieved fromexternal database

OutcomeVerification

Semi-automatic,limited applicabilityVerifiable by the sameapproach

Automated bySWRL rules, limitedapplicability

Manual Semi-automaticVerifiable by the sameapproach

Imple-menta-tion

Proprietary databaseAllow combiningguidelines consideringa patient context

ProprietarydatabaseAllow executingtogether manypairwise combinedguidelines

ProprietarydatabaseAllowcombiningmanytreatments

Web available database(Drugbank)Allow combining two ormore guidelines

and be validated by domain experts. For the otheranalysed approaches all reuse of piece of knowledgerequires the validation of domain experts. Anotheradvantage observed in our approach is that ac-tions can be modified independently of the guide-line (under certain conditions). In the future weintend to better define these conditions and usethem to evaluate when a CIG need to be updated.

Interactions. All approaches provide ways to de-tect interactions between recommendations. Jafar-pour’s approach proposes a number of types ofinteractions that can be identified. It allows ad-dressing identical and conflicting actions, as wellas reuse of results (similar effec/result), which areclose to our proposal, besides addressing tempo-ral/sequential aspects. Lopez-Vallverdu approach


is meant for identifying drug-related interactions.Both approaches propose to have one rule to eachpotential interaction (based on the instances),what can lead to a huge number of rules that be-come complex to maintain up-to-date. In the ap-proach of Wilk et al. a similar situation is ob-served, although the rules as pre-defined CLPclauses allow semi-automatically identification ofconflicts and correspondent mitigation of opera-tors (types are not defined). However, those ap-proaches do not address interactions involvingthree or more recommendations. In our approach,we apply guideline-independent rules for automat-ically identifying some internal and external typesof interactions. These rules can also detect inter-actions holding between two or more recommen-dations within two or more guidelines.Outcome verification. the introduction of new

recommendations requires further verification foreventual new conflicts that could arise. If all poten-tial conflicts need to be solved by adding context-dependent rules, it may lead to a combinatorialexplosion of rules. This, in turn, increases the com-plexity of detecting conflicting rules; especially ifthe verification is done manually by experts. Inparticular, Jafarpour defines SWRL rules that al-low automatically detecting some time/priority-related conflicts between pairs of introduced con-straints, but does not address other types of con-flicts and does not find eventual conflicts with ex-isting tasks. In our approach, since we introducedthe solutions as recommendation in the same for-mat as the original ones, the same rules can beapplied for verifying new conflicts (generated bythe outcome of the reasoning engine).Implementation. All approaches were imple-

mented, however only our approach uses SemanticWeb technology to evaluate the efficacy of the pro-posed method. The other approaches developedtheir own database to evaluate their methods.The proposed approach is still not mature

enough to be applied to handle the whole com-plexity of real guidelines. However, since it is acomplex problem, we are applying an incremen-tal approach to address it. The evaluation of thework done until now uses (simplified) realistic casestudies to demonstrate the current contribution.More features will be incrementally added to theapproach and evaluated in more complec scenar-ios. The proposed prototype will also be incremen-tally improved for demonstrating the feasibility of

automatic detection of interactions. The potentialend-users of an information system with such fea-ture can be (i) guideline developers when handlingdiseases that commonly happen together or (ii)physicians in clinical practice setting, if they wouldneed to combine guidelines (when they were notpreviously addressed by the guideline developers).

An extension of this work addressing hierarchiesof Care Action Types (e.g. Administer Aspirinspecializes Administer NSAID) and side effects isprovided in [32]. Future work includes addressingalso sequencing, composition, time and quantities.Further improvements concerning more detailedrepresentation for transitions, situation types andrecommendations, besides including goals, evi-dence and recommendation strength. We plan tofurther investigate (i) the use of Semantic WebTechnologies to implement future versions of theTMR/TMR4I models, as well as (ii) links to otheravailable databases (e.g. Sider) and terminologies(e.g. SNOMED-CT). In particular, the use of stan-dard terminologies mapped to each other (e.g. viaUMLS) would allow combining guidelines anno-tated with different terminologies. We also intendto pursue compatibility with existing CIG lan-guages meant for execution of guidelines and weposition our work as complementary to them.

Finally, this work focus on the clinical knowl-edge modelling and its applicability for multimor-bidity issue rather than the process of acquir-ing/extracting such knowledge. In other words, itis about which knowledge should be acquired inorder to be able to account for detecting interac-tions among the recommendations, as well as sug-gesting some solution. Therefore, the process of ac-quiring this knowledge from natural language clin-ical guidelines or other sources is not the focus ofthis work.

9. Conclusion

With the ever aging of the population, multi-morbidity is becoming a huge problem and re-quire appropriate tools for supporting the physi-cians to design adapted treatment plans. To thisend, we introduce in this paper TMR4I as amodel for detecting interactions among recom-mendations within several guidelines using Se-mantic Web technologies. We demonstrate thatour approach goes beyond state-of-the-art in ad-


dressing multimorbidity with respect to some im-portant features. This is possible by (i) rely-ing on a more detailed semantics for represent-ing the recommendations and actions, (ii) defin-ing guideline-independent rules for detecting in-teractions and (iii) making the most of Web-basedexternal knowledge sources that allow to detectdrug-drug interactions and alternative drugs withfew or no human-intervention. Our approach fa-vor combining several guidelines since it providesmeans for (semi) automatically identifying interac-tions among many recommendations within manyguidelines, and effectively verifying the resultantmerged guideline by reapplying the approach.

References

[1] S. Abidi and J. Cox and S.S.R. Abidi and M. ShepherdUsing OWL Ontologies for Clinical Guidelines BasedComorbid Decision Support. In: International Confer-ence on System Science (HICSS), EUA, 2012, pp. 3030–3038.

[2] A.A. Boxwala and M. Peleg and S.W. Tu and O. Ogun-yemi and Q.T. Zeng and D. Wang and V.L. Patel andR.A. Greenes and E.H. Shortliffe, GLIF3: a representa-tion format for sharable computer-interpretable clinicalpractice guidelines. Journal of Biomedical Informatics37:3 (2004), 147–161

[3] A. Grace and C. Mahony and J. O’Donoghue and T.Heffernan and D. Molony and T. Carroll, A vision forenhancing multimorbid care using clinical decision sup-port systems. Studies Health Technology and Informat-ics. 192 (2011), 1117–1117.

[4] J. Grosjean and T. Merabti and N. Griffon and B. Da-hamna and S. Darmoni Multilingual multiterminologymodel to create the European Health Terminology/On-tology Portal, in: Proceedings of the 9th InternationalConference on Terminology and Artificial Intelligence,France, 2011, pp. 119–122.

[5] R. Hoekstra, Ontology Representation - Design Pat-terns and Ontologies that Make Sense, Frontiers of Ar-tificial Intelligence and Applications 197 (2009), 1–236

[6] R. Hoekstra and J. Breuker Polishing Diamonds inOWL2 , in: Knowledge Engineering and Knowledge Ac-quisition (EKAW), Italy, 2008, LNCS 5268, pp. 64–73.

[7] R. Hoekstra and A. de Waard and R. Vdovjak, Anno-tating evidence based clinical guidelines - A lightweightontology, in: Workshop on Semantic Web Applicationsand Tools for Life Sciences, France, 2012.

[8] I. Horrocks and O. Kutz and U. Sattler, The Even MoreIrresistible SROIQ, in: Principles of Knowledge Repre-sentation and Reasoning (KR), UK, 2006, pp. 57–67.

[9] D. Isern and A. Moreno, Computer-based execution ofclinical guidelines: a review, International journal ofmedical informatics 77:12 (2008), 787–808.

[10] D. Isern and D. Sanchez and A. Moreno, Ontology-driven execution of clinical guidelines, Computer Meth-

ods and Programs in Biomedicine 107:2 (2012), pp. 122–139.

[11] B. Jafarpour, Ontology Merging using Semantically-defined Merge Criteria and OWL Reasoning Services:Towards Execution-time Merging of Multiple ClinicalWorkflows to Handle Comorbidity, Ph.D. thesis, Dal-housie University, 2013.

[12] A. Jentzsch and B. Andersson and O. Hassanzadehand S. Stephens and C. Bizer, Enabling Tailored Thera-peutics with Linked Data, in: Workshop on Linked Dataon the Web (LDOW), Spain, 2009.

[13] S. Jupp and J. Malone and J. Bolleman and M. Bran-dizi and M. Davies and L. Garcia and A. Gaulton andS. Gehant and C. Laibe and N. Redaschi and S.M.Wimalaratne and M. Martin and N. Le Novere and H.Parkinson and E. Birney and A.M. Jenkinson, The EBIRDF platform: linked open data for the life sciences,Bioinformatics 30:9 (2014), 1338–1339.

[14] J.A. López-Vallverdú and D. Riaño and A. Collado,Rule-based combination of comorbid treatments forchronic diseases applied to hypertension, diabetes mel-litus and heart failure, in Process Support and Knowl-edge Representation in Healthcare, LNAI 7738, Estonia,2012, pp. 30–41.

[15] I. Merelli and H. Perez-Sanchez and S. Gesing andD. D‘Agostino, Managing, Analysing, and IntegratingBig Data in Medical Bioinformatics: Open Problemsand Future Perspectives, BioMed Research Interna-tional (2014).

[16] S. Miksch and Y. Shahar and P. Johnson, Asbru:a task-specific, intention-based, and time-oriented lan-guage for representing skeletal plans, in: Workshop onKnowledge Engineering: Methods & Languages, UK,1997, pp. 1–25.

[17] M. Peleg and S. Tu and G. Leonardi and S. Quagliniand P. Russo and G. Palladini and G. Merlini, Rea-soning with Effects of Clinical Guideline Actions UsingOWL: AL Amyloidosis as a Case Study, in: KnowledgeRepresentation for Health-Care (KR4HC), LNCS 6924,Estonia, 2012, pp. 65–79.

[18] M. Peleg, Computer-interpretable clinical guidelines:a methodological review, Journal of biomedical infor-matics 46:4 (2013), pp. 744–63.

[19] H. Perez-Urbina and B. Motik and I. Horrocks,Tractable query answering and rewriting under descrip-tion logic constraints, Journal of Applied Logic 8:2(2010), pp. 186–209.

[20] H. Perez-Urbina and E. Rodriguez-Diaz and M. Groveand G. Konstantinidis and E. Sirin, Evaluation of queryrewriting approaches for owl 2 , in: Joint Workshop onScalable and High-Performance Semantic Web Systems,CEUR-WS (vol. 943), 2012, pp. 32–44.

[21] C. Pruski and R. Bonacin and M. Da Silveira, Towardsthe formalization of guidelines care actions using pat-terns and semantic web technologies, in: Artificial In-telligence in Medicine, Slovenia, 2011, pp. 302–306.

[22] A. Ruttenberg and T. Clark and W. Bug and M.Samwald and O. Bodenreider and H. Chen and D. Do-herty and K. Forsberg and Y. Gao and V. KashyapJ. Kinoshita and J. Luciano and M.S. Marshall and C.Ogbuji and J. Rees and S. Stephens and G.T. Wong


and E. Wu and D. Zaccagnini and T. Hongsermeier andE. Neumann and I. Herman and K.H. Cheung, Advanc-ing translational research with the Semantic Web, BMCbioinformatics 8:3 (2007).

[23] D. Sittig and A. Wright and J. Osheroff and B. Mid-dleton and J. Teich and J. Ash and E. Campbell andD. Bates, Grand challenges in clinical decision support,Journal of biomedical informatics 41:2 (2008), 387–392.

[24] D.R. Sutton and J. Fox, The Syntax and Semanticsof the PROforma Guideline Modeling Language, Jour-nal of the American Medical Informatics Association10 (2003), 433–443.

[25] A.A. Uijen and E.H. van de Lisdonk, Multimorbid-ity in primary care: prevalence and trend over the last20 years., European Journal of General Practice 14:1(2008), 28–31.

[26] S. Wilk and M. Michalowski and W. Michalowski andM.M. Hing and K. Farion, Reconciling pairs of concur-rently used clinical practice guidelines using ConstraintLogic Programming, in: AMIA Annual Symposium Pro-ceedings, EUA, 2011, p. 944.

[27] S. Wilk and M. Michalowski and W. Michalowski andK. Farion and M.M. Hing and S. Mohapatra, Mitiga-tion of adverse interactions in pairs of clinical practiceguidelines using constraint logic programming, Journalof biomedical informatics 46:2 (2013), pp. 341–53.

[28] D.S. Wishart and C. Knox and A.C. Guo and S. Shri-vastava and M. Hassanali and P. Stothard and Z. Changand J. Woolsey DrugBank: a comprehensive resource forin silico drug discovery and exploration, Nucleic AcidsResearch 1:34 (2006), pp. 668-72.

[29] W. Yao and A. Kumar, CONFlexFlow: Integratingflexible clinical pathways into clinical decision supportsystems using context and rules, Decision Support Sys-

tems 55:2 (2013), pp. 499–515.[30] V. Zamborlini and M. Da Silveira and C. Pruski and

A. ten Teije and F. van Harmelen, Towards a Con-ceptual Model for Enhancing Reasoning about ClinicalGuidelines: A case-study on Comorbidity, in: Knowl-edge Representation for Health-Care (KR4HC). LNCSvol. 8903, Austria, 2014, pp. 29–44.

[31] V. Zamborlini and R. Hoekstra and M. Da Silveiraand C. Pruski and A. ten Teije and F. van Harmelen,AConceptual Model for Detecting Interactions amongMedical Recommendations in Clinical Guidelines, in:Knowledge Engineering and Knowledge Management(EKAW), Sweden, 2014, pp. 591–606.

[32] V. Zamborlini and M. Da Silveira and C. Pruski andA. ten Teije and F. van Harmelen, Analyzing Recom-mendations Interactions in Clinical Guidelines: Impactof action type hierarchies and causation beliefs, in Ar-tificial Intelligence in Medicine (AIME), LNCS 9105,Italy, 2015, pp. 317–326.

[33] http://download.bio2rdf.org/release/3/release.html,Accessed October 2014.

[34] https://github.com/bio2rdf/bio2rdf-scripts/wiki, Ac-cessed October 2014.

[35] http://www.drugbank.ca, Accessed October 2014.[36] http://wifo5-03.informatik.uni-

mannheim.de/diseasome, Accessed October 2014.[37] http://linkedct.org, Accessed October 2014.[38] http://sideeffects.embl.de, Accessed October 2014.[39] https://uts.nlm.nih.gov, Accessed October 2014.[40] http://bioportal.bioontology.org, Accessed October

2014.[41] http://clinicaltrials.gov, Accessed October 2014.

Inferring Recommendation Interactions in Clinical Guidelines 1 · minimize errors, and generalize...

Documents

Transcript of Inferring Recommendation Interactions in Clinical Guidelines 1 · minimize errors, and generalize...