Despiterecentadvancesincancertreatmentsthatprolongpatients’lives,treatment-inducedcardiotoxicityremainsoneseveresideeffect.Theclinicaldecision-makingofcardiotoxicityischallenging,asnon-clinicalsymptomscanbemisseduntillife-threateningeventsoccuratalaterstage,andcliniciansalreadyhavehighworkloadcenteredonthetreatment,notthesideeffect.Ourprojectstartswithaparticipatorydesignstudywith11clinicianstounderstandtheirpracticesandneeds;thenwebuildamultimodalAIsystem,CardioAI,thatintegrateswearablesandLLM-poweredvoiceassistantstomonitormultimodalnon-clinicalsymptoms.Also,thesystemincludesanexplainableriskpredictionmodulethatcangeneratecardiotoxicityriskscoresandsummariesasexplanationstosupportclinicians’decision-making.WeconductedaheuristicevaluationwithfourclinicalexpertsandfoundthattheyallbelieveCardioAIintegrateswellintotheirworkflow,reducestheirinformationoverload,andenablesthemtomakemoreinformeddecisions.
Insummary,ourcontributionsareasfollow:
Weconductacomprehensivereviewofexistingchallengesassociatedwithcancertreatment-inducedcardiotoxicity,varioustechnologiesusedforremotemonitoringofcancerpatients,aswellasAI-basedclinicaldecisionsupportsystems(AI-CDSS)inthiscontext.
OurPDsessionsconsistofthefollowingparts:(i)wefirstaskparticipantstodescribetheirchallengesinmonitoringanddetectingtreatment-inducedcardiotoxicityduringtheircurrentclinicalpractice,andtheirperspectivesonthepotentialofAItechnologiestoaddressthesechallenges;(ii)wethenintroduced,ifneeded,abriefoverviewofcurrenttechnologiesthathavebeenusedinRPMandAI-CDSS,suchaswearables,LLM-basedCA,andAI-drivenpredictionmodels;(iii)wepresentourUIdesigndraftandcollectfeedbackonvisualdesign,interactivefunctionalities,andsystemspecifications.
Ourparticipantsreiteratedthechallengeofsubtlesymptomsintheearlystagesindetectingtreatment-inducedcardiotoxicity,wheresuchsymptomsaredifficultforpatientstodetectduringroutineclinicalassessmentsorwhenpatientsareoutsideofclinicalsettings.BothP6andP4furtheremphasizedthecomplexitybypointingoutthatsymptomsmaynotonlybesubtle,but,insomecases,entirelyabsentevenwhenpatientsareexperiencingsevereconditions.P4sharedacasewhereapatient,completelyasymptomaticathome,wasfoundtohaveaseveresevereconditiononlybecausehewaswearingaheartmonitor,whichisnotprovidedroutinelytoallpatients:
“Hehadbeenontheheartmonitorforabout3or4days…Igotthismessagehehadasevereformofarrhythmia…Hewascompletelyasymptomatic.Hewassittingathomerelaxing,andthemonitorpickedupthisthing.”(P4)
Participantshighlightseveralreasonsbeyondthelackofhealthliteracythatmaypreventpatientsfromaccuratelyreportingtheirsymptoms.Theynotedlogisticalbarriersthatpossiblyhindertheeffectivenessofself-reporting,evenwhenpatientsareawareoftheirsymptoms.Thesebarriers,suchasdelayedcommunicationbetweenpatientsandhealthcareproviders,areparticularlypronouncedinresource-constrainedsettings,leadingtodiscouragementandincreasedcognitiveloadonpatients,asdescribedbyP1:
“Insomeareas,patientsmaynothearbackrightawayfromtheirphysiciansornursesifthey’rehavingsymptoms…Theymayfeellikethey’recomplainingtoomuchorbeingaburden.”(P1)
Anotherlimitationidentifiedbytheparticipantsisthedifficultyinrecognizingcardiotoxicity-specificsymptomswithincurrentself-reportingsystems.Manyoftheexistingtoolsforsymptommonitoringaredesignedforgeneralpurposesanddonotfocusoncardiotoxicity,whichcanleadtomissedorunderreportedcardiacproblems.Forexample,P5emphasizedthatthesurveystheyuseare“notspecifictocardiotoxicity…unlessthere’sanythingontheexamthatsuggestsacardiacproblem.”
Severalparticipants(P1,P2,P7)expressedtheirfrustrationwiththeabsenceofreliablepatternsorcorrelationsbetweenpatientcharacteristicsandcardiotoxicityrisk,makingitchallengingtodeterminewhichpatientsrequiremoreintensivemonitoring.AsP1noted,“Wedon’thaveagoodgrasponwhatfeaturescorrelatewiththelikelihoodofsomeonedevelopingcardiotoxicity.”Suchuncertaintyoftenresultsinreactivedecision-making,whereclinicianshavetowaitforsymptomstomanifestbeforeintervening,ratherthanbeingabletoanticipateandaddressriskssooner.Suchdelays,compoundedbylatesymptomreportinganddiagnosticprocesses,canleadtobelatedintervention.P2highlightedthisissuebysharingacaseinwhichdelaysinreportingandtestingprolongedtreatmentdecisions:“Thesymptomswerereportedlate…everythingwasdelayed,fromtheechotothestresstesttotheEKG.”
AfterweintroducedtheinitialUIdesigndraftoftheinformationdashboard,participantsexpressedstronginterestinsuchamultimodalAI-poweredsystem.Theypointedoutseveralopportunitiesforthesetechnologiestoaddresstheirchallengesandsharedtheirexpectations.
Participantsreportedthatsituationalfactorscouldarisewhenpatientsareinclinicalsettings,suchas“whitecoathypertension”,wherepatientsmayshowelevatedbloodpressureduetoanxietyinaclinicalsetting.Inaddition,participantspointedtotheadvantageofusingwearablesforpatientswhoaregeographicallydistantfromtheclinicorunabletovisitregularlyastheyprovideremoteaccesstovitalhealthdata.P1highlightedhowwearables,particularlyinnon-clinicalsetting,couldhelpcliniciansgainamoreaccuratepictureofapatient’scondition:
“Oftentimes,ourpatients,especiallyontheoncologyoffice,theyhavewhitecodehypertension.Sotheycomeandseeus,andtheirbloodpressuresareskyhigh,butit’sbecausethey’renervous.Idon’tknowwhattheyarelikeoutsideoftheclinic…Thosethingscansometimesbehelpful.It’salmostlikepeoplewhohavecardiacmonitors.”(P1)
Inadditiontothepotentialofwearablesincapturingphysiologicaldata,participantsalsosawgreatpotentialinusingLLM-basedCAstotrackpatient-reportedsymptomsovertime,withoutrequiringmanualinputfrompatients.Severalparticipantsmentionedthatthiscouldeasetheburdenonpatients,manyofwhomstruggletorememberordocumenttheirsymptomsaccuratelybetweenappointments.P2furtheremphasizedtheconvenienceofvoiceinteraction,especiallyforpatientswhomayhavedifficultywithtraditionalreportingmethods,suchasfillingoutforms,“Thisisareallygoodidea.Andthatwaytheytalk,andtheydon’tneedtobetypingorspendingtime.”
Furthermore,clinicianspointedoutthatcombiningwearablesandLLM-basedCAscouldbeparticularlyuseful,aseachtechnologyaddressesdifferentaspectsofpatientmonitoring.Thesetwomodalitiescomplementeachotherandoffertheabilitytolinkphysiologicalvitalsignswithpatient-reportedsymptoms,asP1commented,“beingabletocorrelatethetimewiththeirbiometricscouldbereallyhelpful.”
Severalparticipantsnotedthatnavigatingextensivemedicalrecordscouldbetime-consumingandchallenging,andAI-poweredsummarizationtoolscanhelpalleviatethisburdenbyprovidingconciseoverviewsofpatientrecords.P4echoedthissentimentbyhighlightingthegrowingissueofclinicianburnout,muchofwhichisdrivenbytheneedfordocumentationinEHRs:
“There’salotofburnoutinmedicine,andtheburnoutisfromallofthisdocumentationthatwedo.IseeahugeroleofAIinthatdocumentation…fulfillouttheEHRdataautomatically.”(P4)
“IthinkwithAItherewillbetoolsinplacetoevensuggestwhatthenextordershouldbe…AIwouldhelpyouunderstandwhatthenextstepcouldbeevenbeforeyoustarttalkingtothepatient.”(P7)
Inadditiontoitspotential,cliniciansstressedtheimportanceoftheexplainabilityofAImodels.TheyexpressedtheneedforAImodelstoclearlycommunicatewhatriskscoresmeanandwhatactionsshouldfollow.Withoutexplanations,theymaybeunsurehowtoactontheAI’spredictions.Forinstance,P1discussedthekindofexplanationneeded:“likethetriggereventofhavingthishigherscoreismaybeconsiderationforcardiologyreferral.”
TheneedforexplainabilityisalsocloselytiedtohowcliniciansperceiveAIintheirworkflows.Formanyparticipants,theyexpressedthesamesentimentthatAItoolsshouldserveasassistants,ratherthanreplace,theirclinicaljudgements.ThisviewreinforcestheneedforexplainabilitythatclinicianswanttoretaincontroloverfinaldecisionsandhaveAItoprovideactionableinsightsinawaythatisunderstandable,allowingthemtocomprehendhowAIarrivedatspecificrecommendationsorriskscores.
Participantsprovidedcriticalfeedbackthatreflectedtheirexpectationsandhighlightedseveralareasforimprovement,allowingustorefinethesystemandUIdesigntobetteralignwithclinicians’decision-makingworkflow.
Cliniciansemphasizedtheimportanceofcontinuouslymonitoringacomprehensivesetofphysiologicalparametersandsymptomsthatareclinicallyimportanttocardiotoxicity.Manycliniciansmentionedheartrate,bloodpressure,oxygensaturation,and,wherepossible,additionalindicatorssuchasbloodsugarlevels.Theywerecuriousaboutthepotentialofwearabledevicestotrackthesediversemetrics,particularlyinpatientswithcomplexhealthprofiles.Cliniciansalsonotedtheimportanceofmonitoringcriticalsymptomssuchaschestdiscomfort,palpitations,shortnessofbreath,whicharetypicalearlyindicatorsofcardiotoxicity.
Additionally,clinicianssuggestedtheneedforpatient-specificbaselinesthatcanalertthecareteamonlywhensignificantdeviationoccur,ratherthanbombardingthemwithexcessivedata.Forexample,asuddenincreaseintheheartratethatdeviatesfromapatient’sestablishedpatterncouldtriggeranalertforfurtherinvestigation,asP6suggested,“Thecareteamreallyshouldonlybealertedifthere’sasignificantchange”
Cliniciansexpressedaclearneedforadynamicandflexiblevisualizationofpatientdata,withhigh-levelsummariesanddetaileddataviews.Thisdualapproachwouldenablequickassessmentswhilealsoallowingdeeperdivesintospecifictrendsandvariationswhennecessary.Forinstance,asummarygraphofdailyaverageheartratescanofferasnapshotofapatient’scardiovascularstatus,whiletheoptiontoviewhourlydatawouldprovideinsightintospecificfluctuationsandtheirpotentialcauses.P5emphasizedtheimportanceoftrendsonisolateddatapoints,stating“Thetrendisextremelymoreimportantthanjusttheonetime”.
Thisflexibilityinvisualizingdataenablesclinicianstoidentifypatternsandcorrelationsthatmightnotbeimmediatelyobviousfromasinglemetric.Forexample,beingabletodrilldownintohourlyheartratedatacanhelpcliniciansunderstandthecontextofadailyaveragethatdeviatesfromthebaseline,facilitatingmoreinformeddecision-making.
WhilewepresentedAImodulesandLLM-basedvoiceassistantstoclinicians,cliniciansexpressedtheirexpectationsthattheyshouldbemoreexplainabletopreventmisunderstandingsandensureproperusage.Forexample,understandingthebasisofAIriskscoresiscrucialforclinicianstotrustandeffectivelyusethesystem.Theywouldliketoknowhowtheyarecalculatedandwhattheypresent.
ForLLM-basedvoiceassistant,theyexpressedtheneedtoprovideclearexplanationsabouthowAIrespondstodifferentpatientscenarios.AsP8pointedout,understandingwhentoescalatecarebasedonAIalertsiscrucial:“Shortnessofbreath,discomfort,andpassingoutshouldalwayspromptcontactwiththehealthcareprovider…somethingthatisacute,thatischanging,shouldbeevaluatedbyahealthcareprofessional.”
Basedonclinicians’insightsandfeedbackondesignsuggestions,werefinedourinitialuserinterfacedesigns.WeproposedamultimodalAI-basedsystemforcontinuousmonitoringcancerpatientswithanti-cancertreatmentinanon-clinicalsetting(i.e.,athome)byintegratingadvancedtechnologiesacrossthreemaincomponents:wearableandsmartspeaker,thebackend,anddashboardwithdatavisualization.
Thebackendarchitecturestores,processes,andanalyzesthedatacollectedbythewearableandsmartspeakerwiththefollowingmodules.
Weemployacloud-baseddatabasetostorecollecteddatafromthewearableandsmartspeaker,conversationallogs,andhistoricalEHRs.Thesedatawillfacilitatethefunctionalityoftheothermodulesinthesystem.Thedatabaseisencryptedtoprotecteachindividual’sprivacy.
Sec.A:PatientInformationOverviewintegratesseamlesslywithexistingEHRsystems,providingimmediateaccesstocriticalpatientinformationwithouttheneedtoswitchbetweendifferentplatforms,suchasdemographics,cancertype,stage,andtreatmenthistory.
Sec.B:DailySummaryaggregateandsummarizekeyinformationfromcollectedself-reportedsymptomsandwearablesensordatainaconciseformattosupportaquickoverviewofpatient’scurrentstatus,particularlyfocusingonanysignificantchangesorissuesthatmightrequireimmediateintervention.
Sec.C:WearableSensorDatadisplaysthecollectedphysiologicaldata,suchasheartrate,respiration,oxygensaturation,andskintemperature.Thevisualizationispresentedinaninteractiveformat,allowingproviderstoexploredifferentdatamodalitiesoverspecifictimeperiods,quicklyidentifyingtrendsoranomaliesinthepatient’sphysiologicalsignals.
Sec.D:AIRiskPredictionpresentsthepredictedcardiotoxicityriskscoreanditstrends.WealsopresentitsShapleyValueasabreakdownofthemostinfluentialfactorsthatpredictedtheriskscore,andanLLM-generatedplainlanguageexplanationoftheriskscoreandShapleyValuetoenhanceexplainability.
Sec.E:ConversationalLogshowstherawconversationlogbetweenpatientsandthesmartspeaker.Thevisualizationleveragesaclearcolorschemetodifferentiate”normal”(green)and”abnormal”(yellow)symptomsreportingandaselectionfeaturetoenableclinicianstoselectspecificdatestoreviewthedetailedconversationlogs.
Atthebeginningofthestudy,participantsreceivedanonlinelinktoaFigmaprototype.Theprototypedisplayedrealisticandsyntheticpatientdatathatwehadgeneratedspecificallyforthestudy.Participantswereaskedtosharetheirscreensandnavigatethroughvarioustabsanddashboardinterfacefeatures.Duringthisprocess,theywereencouragedtothinkaloud,explainingeachsteptheytooktointerpretpatientdataandhowtheinformationwasrelatedtotheircurrentdiagnosticworkflow.Theyalsosharedtheirthoughtsontheinterfacedesign.
Inthefirststageofthediagnosisofcardiotoxicity,cliniciansoftenbeginbyidentifyingsymptomsbasedonpatientself-reports,eitheroutsidetheclinicorduringin-personvisits.Clinicianspraisedthesystem’sstreamlinedaccesstosymptomreportsviathevoiceassistantmodule.Theytypicallyexaminedtheconversationlogfordetailedinformationaboutwhenandhowsymptomswerereported.Clinicianshighlightedtheimportanceofcorrelatingthesereportedsymptomswiththepatient’sphysiologicaldata,andtheyappreciatedthesystem’sabilitytoseamlesslyalignsymptomswithphysiologicaldatafordeeperinsight.Forexample,whenpatient-reportedsymptomssuchaschestpainorshortnessofbreathwerepresent,clinicianswouldimmediatelycheckthewearablemoduleforspecificphysiologicalmetricslikeheartrate,bloodpressure,andbloodsugarlevels.AsP7describedhisthinkingprocess:“WhatwasthesymptomSochestpain,shortnessofbreath…thequestioniswhatwastheheartrate,whatwasthebloodpressureWhatwasthebloodsugar”
Also,cliniciansappreciatedthesystem’sabilitytoprovidequickandeasyaccesstocrucialmetrics,suchasbloodpressureandheartrate,whichtheyconsistentlycheckduringpatientassessments.P3emphasizedtheimportanceofthesevitalsignsintheirroutineevaluations,stating:“Bloodpressureandheartratearethekeythingswalkingintotheroom,likeanyandeverytime.”
Afteridentifyingpotentialsymptomsofcardiotoxicity,clinicianstypicallymovetothediagnosticstage,wheretheyreviewrecenttestresultsorordernewteststoconfirmorruleoutcardiotoxicity.Clinicianspraisedoursystemforstreamliningtheretrievalofdiagnosticdata,allowingthemtoeasilyaccesspastandcurrenttestresultsinoneplaceandinterpretdiagnosticresultsinconjunctionwiththepatient’smedicalhistorywithoutswitchingbetweendifferentplatformsorsystems.Cliniciansfrequentlyutilizedthe”Results”and”Therapy”tabstoreviewkeydiagnostictestsforheartfunction,suchasechocardiograms,cardiacMRIs,andEKGs,whicharecrucialforassessingcardiotoxicity.AsP3emphasized:“Fortheheartlaboratorytests,thefirstthingsIlookatarethecardiaclaboratorytests…likedidtheyhaveanechocardiogram,acardiacMRI,orevenEKGs.”Beyonddiagnosticresults,cliniciansvaluedhowthesystemintegratesapatient’streatmenthistorywiththeirdiagnosticdata.
Oncethediagnosticevaluationiscomplete,cliniciansneedtodecidewhethertoadjustthepatient’streatmentplan,initiatenewinterventions,orincreasemonitoringtomitigatecardiotoxicityrisk.Cliniciansemphasizedthevalueofoursystem’sreal-timecardiotoxicityriskscorestoguidethesecriticaldecisions.Thedynamicnatureoftheriskscoresprovidedaclearpictureoftheriskofcardiotoxicityofapatientovertime.
Cliniciansappreciatethatthesystemallowsthemtoassesscardiotoxicityriskbeforeproceedingwithtreatment.P10highlightedhowthisfeaturesupportsinformeddecision-making,particularlywhenchoosingbetweentreatmentswithvaryinglevelsofrisk:
“Ifyoucouldatbaseline,incorporatealltheinformationthatyouhaveonthepatientandsay,‘Hey,thispatient’sactuallysuperhighriskforcardiotoxicityfromtreatmentoptionA,buttreatmentoptionBislessrisky,’thenthatcouldactuallyinformyourtreatment.”(P10)
Byofferingthiscomparison,clinicianscouldbalancethepotentialrisksofcardiotoxicityagainsttheefficacyofcancertreatments,enablingmoretailoredandsafertreatmentplans.
Inadditiontotheinitialriskassessment,cliniciansvaluedtheabilitytotrackchangesinthecardiotoxicityriskscoreovertime.Thisfeatureenabledthemtodetectincreasingrisksandintervenebeforeseriouscardiaceventsoccurred.Thiscapacityforongoingriskmonitoringallowedclinicianstodynamicallyadjusttreatmentplans,ensuringthatpatientswereprotectedfromseverecardiotoxicitywhilestillreceivingeffectivecancercare.
Moreover,cliniciansappreciatedhoweasilyaccessibleandinterpretabletheriskscoreswere,whichcontributedtomakingfast,informeddecisionswithoutbeingoverwhelmedbyexcessivedata.P11expressedtheirsatisfactionwiththesystem’sstraightforwarddesign,“Thisriskscoreisnice.Iwouldlookatthatquickly…Itdoesn’tlookcumbersome.”Thissimplicityallowedclinicianstointegratetheriskscoreintotheirworkflowseamlessly,supportingtheirdecision-makingprocessefficientlyandeffectively.
Duringthestudy,clinicianswereabletointerpretthepresentedinformationandincorporatethemintotheirexistingworkflow.Astheyinteractedwiththesystem,clinicianshighlightedseveralkeyaspectsofthesystem’susabilityanditspotentialtosupporttheirclinicaldecision-making.TheaverageSUSscoreis72.33±1.89plus-or-minus72.331.8972.33\pm1.8972.33±1.89(outof100),suggestingitoffersasolidlevelofusability.
Amajorthemehighlightedbyparticipantswasthesystem’ssimplicityandeaseofuse,whichtheyconsideredessentialforsmoothintegrationintotheirclincialworkflow.AllparticipantsratedthesystemveryhighlyforeaseofuseinSUS,indicatingthattheyfoundthesystemveryeasytonavigate,suchasP3’scomment:“likethisinterfacesofar,isrelativelysimple…Itdoesn’tlookcumbersome.”Similarly,P2echoedthesesentiments,notingthattheintuitivedesignwasanasset:“Thedesignisgoodandsimple.Ithinkthat’sagreatwaytostart.”ThissimplicityindesignwasfurthersupportedbytheNASA-TLXeffortscore,whichaveraged1.13±0.70plus-or-minus1.130.701.13\pm0.701.13±0.70(outof7.0),indicatingthesystemrequiredminimalefforttooperateandcouldbeadoptedeasily.
Participantsalsoexpressedhighconfidenceintheirabilitytousethesystemeffectively,supportedbytheSUSconfidencemetric,andmanystatedthattheyfeltcomfortablecompletingtaskswithoutneedforadditionalhelp.Theyagreedthatmostpeoplewouldbeabletolearntousethesystemveryquickly,asentimentcapturedbythehighSUSscoreforthisquestion,highlightingitslowlearningcurveandeasyfitwithintheirexistingworkflows.
Participantsemphasizedthatthesystem’seaseofaccesstorelevantinformationwasasignificantadvantage.Bypresentingthedatainastreamlinedmanner,thesystemminimizedtheneedtomanuallysearchthroughmultiplerecords,improvingworkflowefficiency.AsP4explained,“Ifyouhavemoreinformation,thiswillmakeiteasierwithoutushavingtolookbackinthechart.”TheNASA-TLXresultsfurtherhighlightedthatthesystemdidnotimposeahighworkloadonusersasmentaldemand(1.13±0.23plus-or-minus1.130.231.13\pm0.231.13±0.23)andphysicaldemand(0.73±0.42plus-or-minus0.730.420.73\pm0.420.73±0.42)arelow.Thissuggeststhatthesystemeffectivelysupportedtheretrievalandorganizationofinformationwithoutimposinganadditionalburdenonusers’mentalorphysicalresources.
Participantsalsoappreciatedthesystem’sabilitytodeliverhighlyspecificandorganizedinformation,particularlyinthecontextofcardiotoxicity-relateddecision-making.Thistargetedpresentationofdatawasseenasessentialforsupportingthemtomakeinformeddecisions.P4pointedoutthatthesystemeffectivelyprioritizedcardiotoxicity-relatedinformationwithamorefocusedview:
“Inthechart,there’ssomuchmoredata.Thisismorespecific,organized,mostlyforcardiotoxicity.Sotheappprovidesthatmorespecificconcernsinformation.”(P4)
Thesystem’scapacitytofacilitateproactivedecisionmakingwasasignificantbenefithighlightedbytheparticipants.Byprovidingreal-timeupdatesandpredictiveinsights,thesystemallowedclinicianstoanticipatepotentialcomplicationsandinterveneearly,shiftingfromareactivetoamoreproactiveapproach.P5remarked,
“Ithinkworkflowwise,thiswouldactuallyjustpreemptus…soatleastthisway,wehavealittlebitmoredatatosaylike,hey,youshouldgo..orhey,justfollowupwithyourcardiologist…Ithinkthatpartwouldbereallyhelpfulintermsofhowweactuallymanagepatientsandwhatwereferthem.”(P5)
Inadditiontoenhancingproactivecare,thesystemwaspraisedforitsroleincomplementingexistingclinicaltoolsratherthanreplacingthem.Cliniciansfeltthatthesystemcouldaugmenttheircapabilitiestomonitorpatients.CliniciansalsorecognizedtheAI-basedsystem’spotentialforsupportinglong-termmonitoringandfollow-upcare,particularlyforpatientsinthesurvivorshipphaseofcancertreatment.Thesystem’sabilitytocontinuouslymonitorpatienthealthforextendedperiodswasseenasavaluabletooltodetectlater-emergingcardiotoxiceffectsandensureongoingpatientsafety.P10suggested,
“Maybeinyoursurvivorship,youknowyourpatientswhoare5,10yearsoutfromtheirdoxorubicinortheirchestradiation…havinga30-daymonitoringperiodbeforetheirsurveillancevisitwouldhelpmakesurethere’snoworryingsignstherethatwouldpromptmorecardiac.”(P10)
Duringthink-aloudsessions,cliniciansalsoprovidedtheirfutureexpectationsforthesystem,highlightingwheretheyenvisiontheprototypecouldbefurtherimproved.
Whileclinicianshighlightedthesystem’simprovedaccesstorelevantinformationasasignificantadvantage,theyalsoexpressedtheneedformoreclinicallyrelevantdatatofurthersupportdecision-making.Onekeyrequestwasforcomprehensivetrackingofallpatientmedications,includingnon-cancertreatmentsthatcouldinteractwithcancertherapiesorexacerbatecardiotoxicity.P7pointedout,“Thisisonlythecancertreatment.Normally,it’snotjust3or4medications…thereareothermedications,”highlightingthecomplexityofmanagingpatientsundergoingcancertreatmentandthenecessityforthesystemtoprovideaholisticviewofallmedicationsandpotentialinteractions.Byincorporatingamorethoroughmedicationhistory,thesystemcouldbettersupportcliniciansinrecognizingpotentialdruginteractionsandsideeffectsthatcouldimpactpatienthealthoutcomes.
Cliniciansalsoexpressedtheneedformoredetailedinformationaboutpatients’anti-cancertreatments,suchasthetypeofchemotherapyadministered,itslastadministrationdate,cumulativedose,andassociatedcardiotoxicrisks.Theyemphasizedthatnotallanti-cancertreatmentshavethesamecardiotoxicpotential,andunderstandingthesespecificsisessentialforeffectivemanagement.P7remarked,“Iwouldwanttoknowwhatexactlyisthechemotherapythatwe’retalkingabout,becausenotallchemotherapieshavecardiotoxicity,”pointingtothenecessityofprovidinggranulardetailstoassessrisksaccuratelyandmakeinformedtreatmentdecisions.
Furthermore,whileclinicianspraisedthesystem’sabilitytostreamlineaccesstorelevantcardiotoxicity-relatedinformation,reducingtheneedtosearchthroughmultiplerecords,theyalsoexpressedtheexpectationforittointegratemoreseamlesslywiththetoolstheyalreadyuse.Reflectingonpastexperiences,theydescribedthechallengeofnavigatingbetweenvarioussystemsandmanagingthemultipleriskscoresandalertsalreadyinEHRs.AsP1noted,
“Wedohave,inourelectronicmedicalrecord,alotoftheseriskscoresthatarealreadykindofpoppingupandshowingupforus.There’sariskscoreforopiateabuse,ariskscoreforhospitalreadmission,acoupleofdifferentthings.Sowewantthesetools,butthere’salsosomealarmfatigue.Sometimesweseesomuchwhenweloginthatwejustkindofignoreitandgettowhatweneedtodointhechart.”(P1)
Anotherprominentthemefromthefeedbackwasthecollaborativenatureofcardiotoxicitydecision-making,emphasizingtheneedforthesystemtooffercustomizationbasedontheuser’sspecialty.Decision-makingincancertreatment-inducedcardiotoxicitytypicallyinvolvesmultiplespecialists,suchascardiologistsandoncologists,whoeachprioritizedifferentaspectsofpatientdata.Forinstance,cardiologistsareprimarilyconcernedwithcardiacmetrics,whileoncologistsneeddetailedinformationaboutanti-cancertreatments.P7,acardiologist,highlightedthisdivergenceinpriorities,explained,“Foracancerdoctor,themammogramisuseful,important.Butforme,itprobablydoesn’tgiveanyadditionalinformation,”pointingoutthevaryingrelevanceofcertaindatatypesdependingonthespecialist’sfocus.Incontrast,P5,amedicaloncologist,stressedtheneedforaquick,high-leveloverviewofcumulativeanthracyclinedosage,reflectingtheirfocusoncancertreatment-relatedmetrics.
Moreover,thecollaborativenatureofdecision-makingoftennecessitatesinputfrombothcardiologistsandoncologists,makingitcrucialforthesystemtofacilitateinformationsharingbetweenspecialties.P11highlightedthiscollaborativedynamic,emphasizingtheimportanceofreviewingnotesfromotherspecialiststogainacomprehensiveunderstandingofthepatient’streatmentandcondition:
“Idolookattherecentprogressnotes.Imean,sometimesthepatientforgotwhathappened.Soit’salsogoodformetogoandlookatthosenotesbeforewalkingin.SoIknowwhatotherprovidershaveseen.AndthenIcanalsoestimatewhatissuesIshouldaddressevenifthepatientdoesn’tbringitup,becausethecancerteammayhaveissuesthattheywantsomelevelofinputon.Ispentalotoftimelookingthesenotestotrytounderstandwhat’sgoingon.Sowecanmakedecisionsaboutwhetherit’sthecancerdrugsornot.”(P11)
Bycateringtothespecificneedsofdifferentspecialistswhileenablingcollaboration,thesystemcanmakeiteasierforclinicianstoaccesstheinformationmostrelevanttotheirrespectiverolesandsharecriticalinsightsacrossspecialties.
Whileremotemonitoringsystemshavethepotentialtoimprovepatientcarethroughreal-timedatacollection,cliniciansexpressedethicalconcernsaroundtheirpracticalimplementation,especiallyaroundtimelyinterventions,ambiguityinresponsibility,andthelackofpatienteducation.
Onesignificantconcernisthegapbetweenreal-timedataavailabilityandtimelyinterventions.Althoughthesystemcouldalertclinicianstoapatient’sacutemedicalcondition,thereremainsuncertaintyabouthowquicklythesealertscanbeactedupon,especiallyoutsideregularofficehours.Cliniciansexpressedworryabouthowthesystemwouldfunctionwhentheyarenotreadilyavailabletorespond.AsP5explained,
“I’llbehonestthatit’sgonnabehardtohave,ortoevenremembertolookatthisdashboardsometimes,rightOrevenifyoucheckit8ameverymorning,you’regonnamiss,you’reinevitablygonnamisssomethinglateroryou’llcatchitlaterafterthefact.”(P5)
Afurtherconcernrelatedtotimelyinterventionistheambiguityofresponsibilitywhencriticalalertsaregenerated.Cliniciansexpressedthat,evenwhenalertsareissuedpromptly,theuncertaintyaroundwhoisresponsibleforactingonthemaddsanotherlayerofcomplexity.AsP11pointedout,“identifyingwho’sgonnatakechargeifthere’saredflagthatcomesoutisasignificantchallenge.”P10expandedonthisconcernbyposingcriticalquestionsabouthowthesealertswouldbemanagedinpractice,giventheconstraintsofstaffingandavailability:
“Thinkingaboutstaffing,thinkingaboutlike,who’sgoingtobecheckinginwiththepatientObviously,ifthepatientjuststartstalking,there’snoguaranteethatsomeone’savailabletolistenandtakethatmessage.Becauseifthehealthcaresystemisbeingnotifiedthatthepatienthasaacutemedicalissuelike,howdoesthatgettotherightpersontofigureout,hey,dotheyneedtocall911Ordoweneedtogetthemtotheemergencyroomrightnow”(P10)
Thishighlightsthelogisticchallengeofensuringthattherightpersonisavailabletoreceivethealertandempoweredtomakecriticaldecisionsbasedonit.
Anotherrecurringthemeinthediscussionswastheriskofpatientsplacingtoomuchtrustinthesystem,assumingthatenteringdataintoanapporchatbotwouldsufficeforurgentmedicalissues.Clinicianswereconcernedthatmisplacedtrustcanleadpatientstodelayseekingimmediatemedicalattention,relyingsolelyonthechatbotforguidance.ThispointstoacriticalneedforpatienteducationasP11emphasizedtheimportanceofclearcommunicationwithpatientswhentheyneedtotakeaction:
“Wehavetomakesurethatpatientsknow…Somepatientsmightthink,’Oh,well,ItoldmychatbotthatIhadchestpain,soit’sfine’…buttheyneedtoknowwhatredflagsymptomsareandthattheyneedtoseekimmediatemedicalattention.”(P11)
P10furtherstressedthatpatientsneedtobeawareofthesystem’slimitationsduringoff-hours,“Patientsneedtoknowthatiftheytoldtheirchatbotthattheyhadchestpainat9pmonaFriday…noone’sgonnacheckthechatbotuntilTuesdaywhentheygetback.”
Cliniciansalsoraisedconcernsabouttheaccuracyandvalidationoftheinformation.Theywerecautiousaboutrelyingondatageneratedbytechnologieswithoutrigorousvalidation.P7noted:
“Ifitisnot(validated),howshouldIbesurethatthisisallaccurateinformationthereBecauseifI’mnotconvinced,likeImightbelookingattheinformation,butifI’mnotsurewhethertheinformationisaccurate,thenwe’llgobacktothepatient’schartinourownsystemandlookforthatinformation.”(P7)
Cliniciansexpressedthatwithoutconfidenceintheaccuracyofthedata,theyareinclinedtoverifyitagainstexistingsystems,whichpotentiallyunderminestheirefficiency.P10echoedthisconcern,emphasizingthatfornewtechnologiestogainadoptioninclinicalsettings,theyneedtoundergothesamerigorousvalidationsasexistingmedicaltools.
Ourworkbuildsonthesestudiesbyspecificallyoperationalizingthesevisionsinthecontextofcancertreatment-inducedcardiotoxicity.WhilepreviousresearchhasdemonstratedthepromiseofRPMandCDSS,theunpredictablenatureofcardiotoxicitynecessitatesamoreproactiveandcontinuousapproach.OurfindingssuggestthatCardioAI,amultimodalAI-basedsystem,hasthepotentialtoextendclinicians’capabilitiesbeyondclinicalsettingsbyprovidingcontinuousmonitoringofsymptomsandreal-timeriskprediction.Byintegratingwearabledevices,LLM-basedvoiceassistants,andAI-drivenanalysis,oursystemfacilitatesseamlessinformationcollectionfromnon-clinicalsettingsintoclinicalworkflows.Ourstudyoffersapracticalapplicationthat’blurstheboundaries’betweenclinicalandnon-clinicalsettings.
WhileourstudydemonstratesthepromiseofmultimodalAI-basedsystemsinextendingclinicalcapabilities,wealsounderscoretheimportanceofconsideringthebroaderimplicationsofthesesystems.Continuousdatastreamscanbeinvaluableforconditionslikecardiotoxicity,whereearlyinterventioniscritical,buttheymayposechallengesinothercontexts,suchascausingpatientanxietyorinformationoverloadinlessdynamicorchronicconditions.Thesetechnologiesshouldnotbeapplieduniformly;instead,wearguethattheiruseneedstobetailoredtospecifichealthcarecontexts.Ratherthanattemptingtoerasetheboundarybetweenclinicalandnon-clinicalsettings,weadvocateforthoughtfullyextendingclinicalcapabilitiestobetteraddresstheneedsofcliniciansandpatients,particularlyforconditionsthatbenefitfromearlydetectionandintervention.Thisperspectiveacknowledgesthestrengthsoftraditionalclinicalsettingswhilerecognizingtheneedforflexibilityindeliveringcareacrossdiverseenvironments.
WealsoidentifycrucialdesignconsiderationsthatcomewiththeintegrationofAI-assistedproactivedecision-makingthroughcontinuousmonitoring.Acentraldesignconsiderationisitsseamlessintegrationintoexistingclinicalworkflows.Giventhatcliniciansarealreadyworkinginhigh-pressureenvironmentswithwell-establishedroutines,theintroductionofthesesystemsneedstobecarefullyalignedwithcurrentpractices.Oursystemminimizesdisruptionbyautomatingroutinetasks,streamliningdatacollection,anddeliveringreal-timeinsightsdirectlywithintheclinicalcontext.Anothercriticalfactoristheriskofinformationoverload,acommonchallengewithcontinuousdatastreams.Oursystemmitigatesthisbyfilteringandprioritizingthemostrelevantdata,ensuringthatcliniciansonlyreceiveactionableinformation.
Thisshifttowardcontinuousmonitoringandreal-timedecisionsupportalsonecessitatesrethinkingtraditionalclinicalworkflowsandresponsibilities.Cliniciansraisedconcernsaboutaccountabilityinacontinuouslymonitoredenvironment,particularlyregardingwhowouldmonitorthedataandintervenewhenabnormalitiesaredetected,particularlyduringnon-workinghours.Theseconcernsintroducequestionsabouttheliabilityandworkloadof24/7monitoring,whichrequiresafundamentalshiftfromthetraditionalepisodiccaremodel.Theneedforconstantvigilanceandpotentialinterventionsmaychallengecurrenthealthcaredeliverystructures,andaddressingtheseissueswillbecrucialtothesuccessfulimplementationofAI-basedsystemsinclinicalpractice.
OurstudyrevealsseveralethicalandpracticalconcernsthatarisewhenintegratingAI-basedcontinuousmonitoringsystemsintoclinicalpractice.Acriticalconcernistheneedforrobustinfrastructuresupport,includingreliableinternetconnections,functioningdevices,andinteroperabilitybetweensystems.Cliniciansnotedthattechnicalfailures,suchasmalfunctioningdevicesorpoorWi-Ficonnectivity,coulddisruptreal-timesymptommonitoring,delayinterventionsandunderminetheeffectivenessofthesystem.EnsuringreliableITinfrastructureandtechnicalsupportwillbecriticalforthesuccessofthesesystems.Inadditiontoinfrastructure,humanresourcesareessentialforongoingmonitoringandmaintenance.WhileAIcanautomatedatacollectionandanalysis,humanoversightremainscrucialforinterpretingalertsandrespondingtoabnormaldata.Cliniciansexpressedconcernsabouttheincreasedworkloadthatthiscanplaceonhealthcareteams,particularlyin24/7monitoringenvironments.Managingthesedemandswillrequirecarefulplanningtoavoidoverburdeningstaffandensurecontinuous,high-qualitycare.
Beyondtechnicalchallenges,thereareethicalconcerns.OnekeyissueistheoverrelianceofpatientsonAI-basedsystemsthatcandelaycriticalmedicalattention.Cliniciansexpressedconcernthatpatientsmightassumeloggingsymptomsintoasystemguaranteesimmediateintervention,evenwhenhealthcareteamsareunavailable,withoutfullyunderstandingthelimitationsoftechnology.Thisriskisfurtheramplifiedinunderservedcommunitieswhereaccesstocliniciansisalreadylimited,potentiallyleadingtoafalsesenseofsecurity.
Inthiscontext,theroleofpatienteducationbecomesparamount.Ourfindingssuggestthatempoweringpatientsthroughtechnologymustbeaccompaniedbyclearcommunicationaboutthesystem’scapabilitiesandlimitations.WhileAI-basedsystemscanprovidevaluableinsightandsuggestions,patientsneedtobeawareofwhichsymptomswarrantimmediatemedicalattentionandwhichcanbemonitoredthroughthesystem.Withoutthiseducation,thereisariskthatpatientsmisunderstandtheurgencyoftheirsymptomsandrelyonthesysteminappropriately,leadingtodelayedinterventions.ThisunderscorestheimportanceofdesigningAI-basedsystemsthatnotonlyprovideaccurateinformationbutalsoguidepatientsinmakinginformeddecisionsaboutwhentoescalatetheirconcernstohumanclinicians.
PrivacyandsecurityconcernsarealsocentraltotheimplementationofAIandLLM-basedsystems.Asthesetechnologieshandlesensitivepatientdata,ensuringtheconfidentialityandintegrityofthisinformationbecomescrucial.Ourstudyrevealedthatcliniciansareparticularlyconcernedabouthowpatientdataisstored,transmitted,andaccessedwithinAI-basedsystems.TheincreasingrelianceonAIraisesquestionsaboutdatabreachesandunauthorizedaccess,whichcouldunderminepatienttrustandposesignificantlegalrisks.AddressingtheseconcernswillrequirerobustdatagovernanceframeworksthatprioritizepatientprivacywhileallowingfortheeffectiveuseofAIinhealthcare.
Ourworkhasseverallimitations.First,whilewegatheredvaluableinsightsduringtheprototypedesignphasewithinputfromelevenexperts,ourevaluationstudyincludedonlyfourparticipants.Futurestudiesshouldinvolveamorediversegroupofparticipantsandconductbroaderevaluations.Second,oursystemwasdevelopedasaprototypeandusedasadesignprobetocollectfeedbackfromcliniciansduringtheevaluation.Thisapproachmayaffectthevalidityandgeneralizabilityofourfindings,astheresultsmightdifferifthesystemweredeployedinreal-worldclinicalsettings.
Inthisstudy,weexplorethechallengesthatcliniciansfaceinmanagingcancertreatment-inducedcardiotoxicityanddevelopedamultimodalAI-basedsystem,CardioAI,tosupportsymptommonitoringandriskprediction.Throughparticipatorydesignsessionswith11clinicians,weuncoveredthecomplexityofmanagingcardiotoxicityandidentifiedgapsinexistingmonitoringapproaches.OursystemprovidescontinuousmonitoringofsymptomsviawearabledevicesandLLM-basedVA,andexplainableAI-basedpredictiveriskscores,whichofferactionableinsightstosupportclinicaldecision-making.Evaluationbyclinicalexpertshighlightedthesystem’sabilitytoreduceinformationoverload,streamlineworkflows,andsupportproactivedecision-making.ThesefindingscontributetothegrowingfieldofHCIinhealthcare,providingafoundationforfuturedevelopmentoftechnologiestoaddresssimilarchallengesinclinicaldecision-makingtasks.WeenvisionthatourworkcaninspirefuturedesignsofmultimodalAI-basedsystems.