Improving Adherence to Heart Failure Management Guidelines via Abductive Reasoning

15 Pages·2017·0.16 MB·English

UnderconsiderationforpublicationinTheoryandPracticeofLogicProgramming 1
Improving Adherence to Heart Failure Management
Guidelines via Abductive Reasoning
7
1
ZhuoChen,ElmerSalazar,KyleMarple,GopalGupta,LakshmanTamil
0
2 UniversityofTexasatDallas,Texas,USA.
(e-mail:{zxc130130,ees101020,kbm072000,gupta,laxman}@utdallas.edu)
l
u
SandeepDas,AlpeshAmin
J
6 CardiologyDivision,DepartmentofInternalMedicine,UniversityofTexasSouthwesternMedicalCenter,Texas,USA.
1 (e-mail:{Sandeep.Das,Alpesh.Amin}@UTSouthwestern.edu)

] submitted1January2003;revised1January2003;accepted1January2003
I
A
.
s Abstract
c
[ Management ofchronicdiseasessuchasheartfailure(HF)isamajorpublichealthproblem.Astandard

approachtomanagingchronicdiseasesbymedicalcommunityistohaveacommitteeofexpertsdevelop
1
guidelinesthatallphysiciansshouldfollow.Duetotheircomplexity,theseguidelinesaredifficulttoimple-
v
7 mentandareadoptedslowlybythemedicalcommunityatlarge.Wehavedevelopedaphysicianadvisory
5 systemthatcodestheentiresetofclinicalpracticeguidelinesformanagingHFusinganswersetprogram-
9 ming(ASP).Inthispaperweshowhowabductivereasoningcanbedeployedtofindmissingsymptomsand
4 conditionsthatthepatientmustexhibitinorderforatreatmentprescribedbyaphysiciantoworkeffectively.
0 Thus,ifaphysiciandoesnotmakeanappropriaterecommendation ormakesanon-adherent recommen-
7. dation,oursystemwilladvisethephysicianaboutsymptomsandconditionsthatmustbeineffectforthat
0 recommendationtoapply.ItisunderconsiderationforacceptanceinTPLP.
7
KEYWORDS:chronicdiseasemanagement,abduction,answersetprogramming,knowledgerepresentation
1
:
v
i
X
r
a
1 Introduction
Heartfailure is theinability ofthe heartto keepup with thedemandsof the body.As a result,
aninadequateamountofbloodissuppliedtothehumanbodyand/orpressuresintheheartcan
rise. This can cause congestionof bloodin the lungs, abdomen,legs. All of this culminatesin
symptomsofexerciseintolerance.HalfofthepeoplediagnosedwithHFdiewithinfiveyears.In
U.S.alonethereare5.7millionpeoplecurrentlylivingwithHF(Goetal.2013).
OptimalmanagementofHFrequiresadherencetoevidence-basedclinicalpracticeguidelines.
TheAmericanCollegeofCardiologyFoundation(ACCF)/AmericanHeartAssociation(AHA)
GuidelinesfortheManagementofHeartFailurehavebeencreatedbyamulti-disciplinarycom-
mittee of experts and are based on thorough review of best available clinical evidence on the
managementofheartfailure.Theyrepresentaconsensusamongexpertsontheappropriatetreat-
mentandmanagementofHF(Jacobsetal.2013).
Althoughevidenced-basedguidelinesshouldbethebasisforalldiseasemanagement,physi-
cians’adherencetothoseguidelinesispoor(Cabanaetal.1999).Oneofthereasonsforthisis
thatguidelinescanbequitecomplex,asisthecaseforHFmanagement.Forexample,morethan
2 Z.Chen,E.Salazar,K.Marple,L.Tamil,G.Gupta,S.Das,A.Amin
100variableshavebeenassociatedwithmortalityandre-hospitalizationinheartfailurepatients.
Inthe2013ACCF/AHAGuidelinefortheManagementofHeartFailure(Yancyetal.2013),the
variablesrangefromstraightforwardinformationlikeageandsextosophisticateddatalikeelec-
trocardiogramresultsandthehistoryofHF-relatedsymptomsanddiseases.Therearemorethan
60complexrulestointegrateallthesedatatogether.Suchcomplexitycanleadtodelaysinadop-
tionandmissedopportunitiesatgivingeffectiverecommendationsforeventhemostexperienced
healthcareproviders(Group2006).
We have developed a physician advisory system that assists physicians in adhering to the
guidelines for managing HF (Chenetal.2016) based on ASP (MarekandTruszczyn´ski1999;
Niemela¨1999). This system automates the 2013 ACCF/AHA Guideline for the Management
of Heart Failure and gives recommendationssimilar to human physicians that strictly follows
the guidelines, even if the information about the patient is incomplete. This system has been
implemented using the the s(ASP) system (Marpleetal.2016b), a goal-directed ASP system
(Marpleetal.2012).Moreover,knowledgepatternsin theguidelinewereidentifiedandmodu-
larized using ASP. An example of the application of knowledge patterns in the system can be
found in Sect. 3. The outputs the physician advisory system produces have been found to be
consistent with our interpretation of the guidelines. The system has also been tested with the
helpofourcardiologistco-authorsonrepresentativepatientdataprovidedbythem.Thesetests
showed that the system makes treatment recommendationsconsistent with physicians’ recom-
mendations.Infact,insomecasesthesystemmadetreatmentrecommendationsthatresultedin
physiciansrevisingtheiroriginaltreatmentrecommendations.
In this paper we build on our previouswork and show how abductivereasoningcan be em-
ployedin oursystem to help physiciansvalidatetheir treatmentrecommendations.In practice,
physiciansoftenneedtogivetreatmentplansbasedonincompleteinformationaboutapatient,
especially in underdevelopedareas where complete information about a patient is not always
available.Weshowhowourphysicianadvisorysystemhelpstoensurethecompliancewiththe
2013ACCF/AHAGuidelinefortheManagementofHeartFailureviaabductivereasoning.The
coreideaisthatthe(non-adherent)physician-prescribedtreatmentistreatedasanobservation
andfedasaquerytoourphysicianadvisorysystem,whichisthenrunintheabductivemodeto
discoverthenecessaryevidencethatmustbeavailableforthegiventreatmenttobeapplicable.
Thisevidencemaybesymptomsthatthepatientmustexhibitortheconditionsthatmustapply
tothepatientforthe(non-adherent)physician-prescribedtreatmenttobeapplicable.Itislikely
thatthephysicianoverlookedestablishingthisevidence,whichisnowproducedviaabduction.
Inlightofthisabducedevidence,thephysicianmaywanttore-evaluatehis/hertreatmentrecom-
mendation(s).
Casestudiesbasedonthismethodologyanditsresultsarealsopresented.Giventhatcompli-
ancewiththeguidelinesispoor(Cabanaetal.1999),weshowhowabductioncanbeusedtoin-
formaphysicianaboutsymptomsandconditionsthatmustexistinapatientforanon-guideline-
complianttreatmentrecommendationtobeapplicable.Ourabduction-poweredsystemcanalso
beusedasatoolfortrainingphysiciansformanagingHF.Thesystemhasbeenevaluatedonten
representativepatients, data for which was providedby our cardiologistco-authors. In four of
thesetencasesoursystempointedoutsymptoms/conditionsthatmayhavebeenoverlooked.In
threeofthetencases,oursystemcouldnotabduceanythingtovalidatetheoriginalrecommen-
dation,indicatingthatthe recommendationshouldbe rejectedorclosely examined.Inthreeof
thecases,thephysician’srecommendationwerecorrect,i.e.,compliantwiththeguidelines.
Ourresearchmakesthefollowingcontribution:
TheoryandPracticeofLogicProgramming 3
• We show how abductive reasoning within ASP can be useful for medical diagnosis, es-
pecially, for heart failure management. To the best of our knowledge, our work is the
firstreal-lifeapplicationofabductionwithASPtechnologyintheveryimportantareaof
medicine.
• Ourresearchillustratesamajoradvantageofgoal-directedASPsystemssuchass(ASP),
namely, their ability to perform effective abductive reasoning. As explained later, it is
difficulttodeterminetheminimumsetofabducibleswithSAT-solverbasedASPsystems
suchasCLASP.In(Chenetal.2016),webrieflymentioneds(ASP)’spotentialtoproduce
abducibles with respect to a query. This paper demonstrates the technical details of the
implementationofabductivereasoningins(ASP).Tenverifiedexperimentalresultsusing
abductionarealsopresented.
2 AbductiveReasoningInGoal-directedASPSystems
Ourphysicianadvisorysystemhasbeendevelopedonthes(ASP)PredicateAnswerSetProgram-
mingsystem.Thes(ASP)systemalsoperformsabductionunderthestablemodelsemantics.We
givea briefintroductionto bothgoal-directedexecutionofanswer setprogramsandabductive
reasoningforanswersetprogramming.WeassumethatthereaderisfamiliarwithASP.
Goal-directedExecution of ASP: With the ability to compute stable models of normal logic
programs,goal-directedsystems forexecutinganswer set programsadopta differentapproach
thantraditionalASPsystemsthatarealmostallbasedontheuseofSATsolvers.TraditionalASP
systemsfindtheentiremodel(answerset)ofagivenanswersetprogram.CheckingifaqueryQ
issatisfiedinthisprogramthenamountstocheckingifQispresentinthecomputedanswerset.
Incontrast,agoal-directedsystemisquerydriven.GivenananswersetprogramPandaquery
goalQ,agoal-directedsystemforexecutinganswersetprogramswillsystematicallyenumerate
all answer sets that contain the propositions/predicatesin Q. This enumeration employs Coin-
ductiveSLDresolution(orCo-SLDresolution),whichsystematicallycomputeselementsofthe
greatestfixedpoint(GFP)ofaprogramviabacktracking(Marpleetal.2012).Withs(ASP),our
goal-directedmethod lifts the restriction that answer set programsmust be finitely groundable
(Marpleetal.2012).
Oneramificationofthetop-downexecutionstrategyofagoal-directedASPsystemisthatthe
answersetproducedmaybepartial(Marpleetal.2012).1Considerthefollowingprogram:
p :- not q. r :- not s.
q :- not p. s :- not r.
Usinga goal-directedASP system, the query:- q will produce{q, not p} asthe answer be-
causetherulescontainingrandsareindependentoftherulesforqandp.Hadthequerybeen
:-q,s,thesystemwouldgive{q, not p, s, not r}astheresult.Hencethepartoftheanswerset
wegetisdeterminedbythequery.Thepartialanswersetessentiallycontainstheelementsthat
arenecessarytoestablishthequery.
TheGalliwaspsystem(MarpleandGupta2012)wasthefirstefficientimplementationofthe
goal-directed method. It is essentially an implementation of A-Prolog (Baral2003) that uses
groundednormallogicprogramsasinput.TheunderlyingalgorithmofGalliwaspmakesuseof
1Becauseanswersetsarepartial,weenumerateelementsthataretrueaswellasfalseinthepartialanswerset.
4 Z.Chen,E.Salazar,K.Marple,L.Tamil,G.Gupta,S.Das,A.Amin
coinduction (Guptaetal.2007) to find partial answer sets containing a given query. In query-
driven,top-downexecution,coinductionishelpfulinprocessingevenloops2 wherecyclicalna-
tureofexecutionneedstobedetectedandusedasabasisforgoal-success(MarpleandGupta2012).
Specifically,Galliwaspusescallgraphstoclassifyrulesbasedontwoattributes:(i)aruleissaid
tocontainanoddloopovernegation(OLON)ifitcanbecalledrecursivelywithanoddnumber
ofnegationsbetweentheinitialcallanditsrecursiveinvocation,and(ii)itiscalledanordinary
ruleifithasatleastonepathinthecallgraphthatwillnotresultinsuchacall.Galliwaspuses
ordinaryrulestogeneratecandidateanswersetsandOLONrulestorejectinvalidcandidatesets.
AnimplementationofGalliwaspisavailable(MarpleandGupta2012).
The s(ASP) predicate ASP system (Marpleetal.2016a; Marpleetal.2016b) operates in a
manner similar to its predecessor, Galliwasp, in that rules are classified according to nega-
tion (OLON rules and non-OLON rules) and executed in a goal-directed manner. However,
s(ASP) completely removes the need to ground input programs. Therefore, unlike Galliwasp
and other ASP solvers, s(ASP) programs do not need to have a finite grounding. That is, the
s(ASP) system executes predicate answer set programs directly, without grounding them at
all. Unlike traditional ASP systems, s(ASP) programscan contain arbitrary terms; lists, terms
and complexdata structurescan appearas argumentsof predicatesin s(ASP) programs.Thus,
s(ASP)isamuchmorepowerfulandexpressivesystem.Theintroductionoffullpredicateswith
termsrequiresanumberofinnovativetechniquestosupportthem.Thesetechniquesincludeco-
SLD resolution (Guptaetal.2007), constructive negation, and dual rules (Marpleetal.2016a;
Marpleetal.2016b).Animplementationofthes(ASP)systemisavailable(Marpleetal.2016b)
atsourceforge.
Abduction:Thetermabductionreferstoaformofreasoningthatisconcernedwiththegener-
ationandevaluationofexplanatoryhypotheses.Abductivereasoningleadsbackfromfactstoa
proposedexplanationofthosefacts.Accordingto(Harman1965),abductivereasoningtakesthe
followingform:
The fact B is observed.
But if A were true, B would be a matter of course.
Hence, there is reason to suspect that A is true.
In this form, B can be either a particular event or an empirical generalization. A serves an
explanatory hypothesis and B follows from A combined with relevant background knowledge.
Note that A is not necessarily true, but plausible and worthy of further validation. A simple
exampleofabductivereasoningisthatonemightattributethesymptomsofacommoncoldtoa
viralinfection.
2.1 AbductiveReasoninginASP
Moreformally,abductionisaformofreasoningwhere,giventhepremiseP⇒Q,andtheobser-
vationQ,onesurmises(abduces)thatPholds(Kakasetal.1992).Moregenerally,givenatheory
T,anobservationO,andasetofabduciblesA,thenE isanexplanationofO(whereE⊂A)if:
1. T∪E|=O
2Anevenloopoccurswhenarecursivecallisencounteredwithaneven,non-zeronumberofnegationsbetweenthe
callanditsancestorinthecallstack.(Marpleetal.2016a)
TheoryandPracticeofLogicProgramming 5
2. T∪E isconsistent
Generally,Aconsistsofasetofpropositionssuchthatifp∈A,thenthereisnoruleinT withpas
itshead.WeassumethetheoryT tobeananswersetprogram.Underagoal-directedexecution
regime,anASPsystemcanbeextendedwithabductionbysimplyaddingthefollowingrule:
p :- not _not_p.
_not_p :- not p.
foreveryp∈A. Posingthe observationO asa queryto thisextendedanswer set programwill
yieldalltheexplanationsaspartofa(partial)answerset.Thereasonwhythisworksissimple:
rulesoftheformabovehavetwoanswersets,oneinwhichpisfalseandanotherinwhichpis
true.Thesesettings(alongwithassignmentstootherpropositions)maymaketheobservationO
succeed.Asanexample,considertheprogrambelow:
p :- a, not q.
q :- a, b.
q :- c.
andtheabducibles{a, b, c}.Ifweaddtheclausesbelowtoprogramabove:
a :- not _not_a. b :- not _not_b.
_not_a :- not a. _not_b :- not b.
c :- not _not_c.
_not_c :- not c.
thenthequery(observation)?- pwillsucceed,producingtheanswerset{p, not q, a, not
b, not c} and abducibles {a, not b, not c}. First, not a, not b and not c are not
included in the answer set as they are auxiliary predicates introduced for convenience. Note
thats(ASP)omitsfromoutputthepredicatesstartingwithanunderscore(’ ’).Second,ingoal-
directed execution,since the models generatedmay be partial, propositions/predicatesthat are
knownto be true, aswell aspropositions/predicatesknownto be false, are shownexplicitlyin
the(partial)answerset.
It should also be noted that abductive reasoning is more precise under goal-directed execu-
tion.Otherexecutionstrategies,mostofwhicharebasedaroundSAT-solvers(e.g.,theCLASP
system(Gebseretal.2014)),computetheentirestablemodel,whichcanproduceconfusingan-
swerswhereitisnotclearastowhichoftheabduciblesareinvolvedintheexplanation.Consider
the exampleprogramabove.Supposewe extendthe set of abduciblesto includea new propo-
sition {d}andaddthe two customaryrulesinvolvingd and not d. Forobservation{p},if we
use a traditionalASP solver,nowwe will gettwo answer sets: {p, not q, a, not b, not
c, d} and {p, not q, a, not b, not c, not d}. These two answer sets essentially are
extensionsofthesingleanswersetproducedbythegoal-directedmethod,however,thisanswer
sethastobereplicatedtwice—onceforincluding{d}andonceforincluding{not d}.Thus,if
thereareirrelevantabducibles,thenforagivenobservation,thenumberofanswersetswillpro-
liferateexponentially.Considerableanalysiswillberequiredtoextractthetrueabducibles.The
goal-directedexecutionmethoddoesnotsufferfromthisproblem,asitonlyexploresthespace
ofknowledgethatisrelevanttotheobservation.Thus,onlyrelevantpropositions/predicatesare
abduced.
6 Z.Chen,E.Salazar,K.Marple,L.Tamil,G.Gupta,S.Das,A.Amin
3 PhysicianAdvisorySystemforHFmanagement
InthissectionwediscusstheguidelinesforHFmanagement,aswellasgiveanoverviewofour
physicianadvisorysystem(Chenetal.2016).
The2013ACCF/AHA GuidelinefortheManagementofHeartFailureisbasedonfourpro-
gressivestagesofheartfailure.StageAincludespatientsatriskofheartfailurewhoareasymp-
tomaticanddonothavestructuralheartdisease.StageBdescribesasymptomaticpatientswith
structuralheartdiseases;itincludesNewYorkHeartAssociation(NYHA)classI,inwhichordi-
naryphysicalactivitydoesnotcausesymptomsofheartfailure.StageCdescribespatientswith
structuralheartdiseasewhohavepriororcurrentsymptomsofheartfailure;itincludesNYHA
classI,II(slightlimitationofphysicalactivity),III(markedlimitationofphysicalactivity)and
IV(unabletocarryonanyphysicalactivitywithoutsymptomsofheartfailure,orsymptomsof
heartfailureatrest).StageDdescribespatientswithrefractoryheartfailurewhorequirespecial-
izedinterventions;itincludesNYHAclassIV.Interventionsateachstageareaimedatreducing
riskfactors(stageA),treatingstructuralheartdisease(stageB)andreducingmorbidityandmor-
tality(stagesCandD).
ThetreatmentrecommendationsinthesefourstagesarearticulatedintheACCF/AHAClass
ofRecommendation(COR)andLevelofEvidence(LOE).CORisanestimateofthesizeofthe
treatmenteffectconsideringrisksversusbenefitsinadditiontoevidenceand/oragreementthata
giventreatmentorprocedureisorisnotuseful/effectiveorinsomesituationsmaycauseharm.
CORissummarizedasfollows:
• ClassIrecommendation:thetreatmentiseffectiveandshouldbeperformed.
• ClassIIarecommendation:thetreatmentcanbeusefulandisprobablyrecommended.
• ClassIIbrecommendation:thetreatment’susefulnessisuncertainanditmightbeconsid-
ered.
• ClassIIIrecommendation:thetreatmentisharmfulorunhelpfulandshouldbenotadmin-
istrated.
Onerulefortreatmentrecommendationsintheguidelinelooksasfollows:
“Aldosteronereceptorantagonistsarerecommendedtoreducemorbidityandmortalityfollowinganacute
MI in patients who have LVEF of 40% or less who develop symptoms of HF or who have a history of
diabetesmellitus,unlesscontraindicated.”
To overcome the difficulties that physicians face in implementing the guidelines, we have
developeda physicianadvisorysystem thatautomatesthe 2013ACCF/AHA Guidelineforthe
ManagementofHeartFailure.Ourphysicianadvisorysystemisabletogiverecommendations
withjustificationslikearealhumanphysicianwhostrictlyfollowstheguidelines,evenunderthe
conditionofincompleteinformationaboutthepatient.
The physician advisory system for HF managementconsists of the rule database and a fact
table. The rule database covers all the knowledge in the 2013 ACCF/AHA Guideline for the
Management of Heart Failure. The fact table contains the relevant information of the patient
withheartfailuresuchasdemographics,historyofpresentillness,dailysymptoms,medication
intolerance, known risk factors, measurements as well as ACCF/AHA stage and NYHA class
(Yancyetal.2013).Theserulesandfactscanbeloadedintothes(ASP)system,andthequery?-
recommendation(Treatment, Recommendation class) posed to find possible treatments
for this patient. There may be more than one potential treatment that can be discovered via
backtracking.
TheoryandPracticeofLogicProgramming 7
Therulesabovecanbegrounded,andthenthegroundedprogramcanberunontheCLASP
system(Gebseretal.2014)aswell.NotethatSAT-basedASPsystemssuchasCLASPwillre-
portall the recommendationsin a single answer set, as they cannotdistinguish between treat-
mentsthatareinthesamemodelbutareindependentofeachother.However,knowingallappli-
cablerecommendationsforapatientisuseful,aswillbediscussedlater.Theseapplicablerecom-
mendationsarepharmaceuticaltreatments,managementobjectivesanddevice/surgicaltherapies.
s(ASP)isabletogenerateallvalidrecommendationsforagivenpatientaswell,however,CLASP
ismuchfasterintermsofexecutionefficiency.Themainadvantageofthes(ASP)systemforour
HFapplicationisthatitisabletogenerateindividualrecommendationsseparatelyalongwithall
thepredicates/propositionsneededtosupportthatrecommendation.
Theguidelinerulesarefairlycomplexandrequiretheuseofnegationasfailure,non-monotonic
reasoning and reasoning with incomplete information. A fairly common situation in medicine
is that a drug can only be recommended if its use is not contraindicated (i.e., the use of the
drugwillnothaveanadverseimpactonthatpatient).Contraindicationisnaturallymodeledvia
negationasfailure.Theabilityofanswersetprogrammingtomodeldefaults,exceptions,weak
exceptions, preferences, etc., makes it ideally suited for coding these guidelines. Not surpris-
ingly, the program for the physician advisory system is unstratified. Both odd and even loops
overnegationoccurintheprogram.Asanexample,theindispensablechoiceknowledgepattern
(Chenetal.2016) used in our system contains an even loop. As a concrete example of non-
stratification,considerthefollowingrulefromACCF/AHAstageC(Yancyetal.2013):
“Inpatientswithacurrentorrecenthistoryoffluidretention,betablockersshouldnotbeprescribedwithout
diuretics.”
The pattern describes a case in which if a choice is made, some other choices must also be
made;if those choices can’tbe made, then the first choice is revoked (Chenetal.2016). Note
thatPrologconventionsarefollowed(variablesbeginwithanuppercaseletter).
recommendation(beta_blockers, class_1) :-
not absent_indispensable_choice(beta_blockers),
not rejection(beta_blockers), evidence(accf_stage_c),
diagnosis(hf_with_reduced_ef).
absent_indispensable_choice(beta_blockers) :-
not recommendation(diuretics, class_1),
diagnosis(hf_with_reduced_ef), evidence(accf_stage_c),
current_or_recent_history_of_fluid_retention.
recommendation(diuretics, class_1) :-
recommendation(beta_blockers, class_1),
diagnosis(hf_with_reduced_ef),
not rejection(diuretics), evidence(accf_stage_c),
current_or_recent_history_of_fluid_retention.
This paper focuses on the physician advisory system’s ability to perform abductive reason-
ing.Usingthisabductivecapability,aphysiciancan,forexample,figureoutthesymptomsthat
aparticularpatientmustexhibitfora giventreatmentrecommendedbythe doctor(sothatthis
8 Z.Chen,E.Salazar,K.Marple,L.Tamil,G.Gupta,S.Das,A.Amin
treatmentisconsistentwiththe2013ACCF/AHAGuidelinefortheManagementofHeartFail-
ure).Theautomaticcheckingforcomplianceusingthephysicianadvisorysystemisdiscussedin
detailinSect.4.
4 AutomaticCheckingofAdherencetoGuidelines
AbductiveReasoningins(ASP):Abductivereasoningissupportedbythes(ASP)system,where
theabduciblespredicatesarespecifiedasfollows,forexample:
#abducible goal(X).
Thisdeclarationmeansthatthespecifiedgoalmaybeabducedwhenfailureofthequerywould
otherwiseoccurduringexecution.Specifically,s(ASP)createsthefollowingrulesforthedecla-
rationstatementabove:
goal(X) :- not _neggoal(X), abducible(goal(X)).
_neggoal(X) :- not goal(X).
abducible(goal(X)) :- not _negabducible(goal(X)).
_negabducible(goal(X)) :- not abducible(goal(X)).
Case Study: As mentioned earlier, many times available patient information is incomplete or
guidelines are not complied with. A physician may give prescriptions using his/her intuitions
basedon incompleteinformationabouta patient. Insuch a case, the physician’srecommenda-
tioncanberuninthes(ASP)systemintheabductivemodewherethephysician’srecommenda-
tionisposedasaquery.Thesystemwillthentellthephysician,viaabduction,anysymptoms,
conditions,orrequiredevidencethatmustbepresentfortherecommendationtobecorrect.The
physician can then work on ascertaining the presence of those symptoms, conditions, etc., or
revisetheirrecommendationifthosesymptoms/conditionsarenotpresent.Sointhiscase,while
thephysicianmayhavemadeanincorrectrecommendation,it canstill beapplicableif the ab-
duced symptoms/conditions/evidencewere to be present. In contrast, if the query representing
thephysician’srecommendationfails,thenitmeansthatthephysician’srecommendationisin-
correctwithrespecttotheguidelines,andshouldberejectedorcarefullyre-evaluated.
Ouraim,thus,istohelpimprovecompliancewiththeclinicalpracticeguidelinesviaabduc-
tivereasoning.Thisisachievedasfollows:allguideline-complianttreatmentrecommendations
arefirstgeneratedinasingleanswersetwiththehelpoftheCLASPsystem.Anytreatmentrec-
ommendationmadebyadoctorthatisnotamongtherecommendationsmadebyCLASPsystem
isconsideredanon-guideline-complianttreatmentrecommendation.Anon-guideline-compliant
treatmentrecommendationmadebyaphysicianisposedasanobservationandfedasaqueryto
s(ASP)system,whichisthenrunintheabductivemodetodiscoverthenecessaryevidencethat
mustbeavailableforthegiventreatmenttobeapplicable.
Figure1displaysthefacttabledistilledfromtheprofileofoneofthe10representativepatients.
Therepresentativepatients’datawereprovidedbytheUniversityofTexasSouthwesternMedical
Center.
According the patient’s profile, her doctor prescribed the combination of hydralazine and
isosorbidedinitrate(isordil/hydralazine)tothispatient,However,weknowthattherecommen-
dationofisordil/hydralazineisnotreasonableforthispatientundertheguidelinesoncewerun
theCLASPsystemonthedata.
Sinceisordil/hydralazineishighlyeffectiveintreatingHFforAfricanAmericans,wewantto
TheoryandPracticeofLogicProgramming 9
%doctor’s assessments %demographics
evidence(accf_stage_c). evidence(female).
evidence(nyha_class_3). evidence(age, 60).
%history of the patient
diagnosis(hf_with_reduced_ef). diagnosis(diabetes).
diagnosis(dilated_cardiomyopathy). evidence(angina).
history(standard_neurohumoral_antagonist_therapy).
%measurements from the lab
measurement(potassium, 3.0). measurement(lvef, 0.16).
measurement(creatinine, 1.49). measurement(non_lbbb, 110).
measurement(nt_pro_bnp, 5051).
measurement(glomerular_filtration_rate, 44).
%contraindications
contraindication(crt). contraindication(ace_inhibitors).
Fig.1. Representationofthepatient’sinformation
makesurewedidnotmissanyvitalinformationregardingtheirrecommendation.Theapplicable
rulesforthisclassIrecommendationofisordil/hydralazineisreproducedbelow(Yancyetal.2013):
“ClassI : The combination of hydralazine and isosorbide dinitrateis recommended to reduce morbidity
andmortalityforpatientsself-describedasAfricanAmericanswithNYHAclassIII-IVHFrEFreceiving
optimaltherapywithACEinhibitorsandbetablockers,unlesscontraindicated.”
“ThecombinationofhydralazineandisosorbidedinitrateshouldnotbeusedforthetreatmentofHFrEFin
patientswhohavenoprioruseofstandardneurohumoralantagonisttherapyandshouldnotbesubstituted
forACEinhibitororARBtherapyinpatientswhoaretoleratingtherapywithoutdifficulty.”
Nowwe declareall possibleabducibleswhichare notshownin the patient’sprofiles(thisstep
canbeautomatedusingsimplestringmatching):
#abducible contraindication(beta_blockers).
#abducible contraindication(icd).
...
#abducible diagnosis(atrial_fibrillation).
#abducible diagnosis(hypertension).
...
#abducible evidence(fluid_retention).
#abducible evidence(sleep_apnea).
...
#abducible history(angioedema).
#abducible history(thromboembolism).
...
Abduciblesinvolvingnumericvaluesaredeclaredastextualpropositions,e.g.,“lvef less than 30”,
“mi post 40 days”,etc.Andcorrespondingauxiliaryrulesareintroduced.Forexample:
lvef_less_than_30:- measurement(lvef, Data), Data =< 30.
mi_post_40_days:- measurement(mi, Data), Data >= 40.
10 Z.Chen,E.Salazar,K.Marple,L.Tamil,G.Gupta,S.Das,A.Amin
{ contraindication(ace_inhibitors), contraindication(arbs),
diagnosis(hf_with_reduced_ef), evidence(accf_stage_c), history(ace_inhibitors),
recommendation(hydralazine_and_isosorbide_dinitrate,class_2a),
treatment(hydralazine_and_isosorbide_dinitrate),
not absent_indispensable_treatment(hydralazine_and_isosorbide_dinitrate),
not contraindication(hydralazine_and_isosorbide_dinitrate),
not rejection(hydralazine_and_isosorbide_dinitrate),
not skip_concomitant_treatment(hydralazine_and_isosorbide_dinitrate),
not superseded(hydralazine_and_isosorbide_dinitrate),
not taboo_choice(hydralazine_and_isosorbide_dinitrate) }
Abducibles: { history(angioedema), contraindication(arbs) }
Fig.2. Resultsofabductivereasoning
The declaration of abducibles should be placed after facts and rules in s(ASP) since we do
notwantto abduceknownfactsor deduciblefacts. Thisarrangementguaranteesthatwhatever
s(ASP)abducesleadsustooverlookedormissedevidencewhichisindispensabletojustifythe
treatmentrecommendationinquestion.
Thenweposethefollowingquery(theobservation)tothes(ASP)systemwiththeabductive
reasoningflagset:
recommendation(hydralazine_and_isosorbide_dinitrate, class_1)
Withtheabduciblesdeclared,s(ASP)mayabducethemwhenfailureofthequerywouldoth-
erwiseoccur.Inthiscase,s(ASP)returnsfalse,whichmeansthereistrulynothingwecandoto
makeclassIrecommendationofisordil/hydralazine,giventhepatient’sdata.
NextwewanttoknowifitispossibletomakeclassIIarecommendationofisordil/hydralazine.
Therelevantrulesintheguideline(Yancyetal.2013)regardingtheclassIIarecommendationof
isordil/hydralazineareshownbelow:
“Class IIa : A combination of hydralazine and isosorbide dinitrate can be useful to reduce morbidity or
mortalityinpatientswithcurrentorpriorsymptomaticHFrEFwhocannot begivenanACEinhibitoror
ARBbecauseofdrugintolerance,hypotension,orrenalinsufficiency,unlesscontraindicated.”
Next,weposethefollowingquery(theobservation)tos(ASP)runninginabductivemode:
recommendation(hydralazine_and_isosorbide_dinitrate, class_2a)
Oneoftheanswersetsaswellasthelistofabduciblesgivenbys(ASP)aredisplayedinFig.2.
We immediately know from Fig. 2 that two things are abduced by the system: 1) the con-
traindicationofARBs and2)thehistoryofangioedema.ItisclearfromtheFig.1thatbothof
theabduciblesarenotinthepatient’sprofile.Withsomemedicalknowledge,weknowthatthe
historyofangioedemaisanindicationforthecontraindicationofACEinhibitors,whichislisted
inthefactsheet.ItisreasonableforustomakeaclassIIarecommendationonceweconfirmthe
intoleranceofARBsforthispatient.Inthiscase,thepatient’sprofilesaysnothingaboutherin-
toleranceofARBs.Thus,thedoctorcouldlookforthispieceofinformationelsewheretoensure
thatshedoesnotmissusefulinformationregardingtherecommendationofisordil/hydralazine.