Download - Visual Exploration of Temporal Data in Electronic Medical ... · Visual Exploration of Temporal Data in Electronic Medical Records Josua Krause, Narges Razavian, PhD, Enrico Bertini,

Transcript
Page 1: Visual Exploration of Temporal Data in Electronic Medical ... · Visual Exploration of Temporal Data in Electronic Medical Records Josua Krause, Narges Razavian, PhD, Enrico Bertini,

VisualExplorationofTemporalDatainElectronicMedicalRecords

JosuaKrause,NargesRazavian,PhD,EnricoBertini,PhD,DavidSontag,PhD

NewYorkUniversity,NewYork,NY

Electronic medical records (EMRs) andadministrativedatacontaina largeamountofdistinct events, like diagnoses, laboratorytests,etc.,makingitdifficultto“tell”thestoryofapatient.Weproposepatient-viztoaddressthis issue by using a visual representation ofthe data. This allows us to provide the largenumberofdistincteventtypesandadditionalinformation like costs and hospital stays in amanageable form.Usingbothananonymizedpublic and an unaltered private dataset weexploretheusefulnessofourtool.

Abstract

Longitudinalstudiesandinsuranceclaimsdatageneratealargeamountofelectronicmedicalrecords (EMRs) and administrative data. Thisdatacontainsalargenumberofdistincteventsthroughout theobservedtimewindowof thelife of patients. Understanding, interpreting,and finding relations in those records is achallengingtaskthatishardtoachieveusingatabularorsimilarrepresentation.There-fore, visualization is needed to e.g. betterunderstand predictivemodels built on top ofthedata,impactofcomorbidities,progressionof chronic diseases, or contributors of healthcarecosts.Our proposed tool patient-viz has a visuallyrichdesignaimedtomakethehugeamountofadministrative data manageable for datascientistsandmedicaldoctors.Thegoalofthetool is to provide a quick overview of onepatient which can be further explored toinspectdetailedinformation.Theinputcanbeanytemporaleventdatawitha largenumberof differently typed events. We test our toolwith online accessible semi-synthetic dataprovided by CMS [1] and privately collecteddatafromamajorUSinsurancecompany.

Introduction

JosuaKrause([email protected]) NargesRazavian([email protected]) EnricoBertini([email protected])DavidSontag([email protected]) Website:https://github.com/nyuvis/patient-viz

ContactInformation 1. CMSmedicareclaimssyntheticpublicusefiles.www.cms.gov/Research-Statistics-Data-and-Systems/Downloadable-Public-Use-Files/SynPUFs/,2008–2010.[Online;accessed16-July-2015].

2. HCUP.ClinicalClassificationsSoftware(CCS)forICD-9-CM.www.hcup-us.ahrq.gov/toolssoftware/ccs/ccsfactsheet.jsp,2012.[Online;accessed16-July-2015].

3. C.Plaisant,B.Milash,A.Rose,S.Widoff,andB.Shneiderman.Lifelines:Visualizingpersonalhistories.InProceedingsoftheSIGCHIConferenceonHumanFactorsinComputingSystems,CHI’96,pages221–227,NewYork,NY,USA,1996.ACM.

4. D.Klimov,Y.Shahar,andM.Taieb-Maimon.Intelligentvisualizationandexplorationoftime-orienteddataofmultiplepatients.ArtificialIntelligenceinMedicine,49(1):11–31,2010.

References

Verticalselectioncanbeusedtoseealleventsofaparticular day. Here a set of lab-tests wasperformed.Out-of-rangevaluesaredepictedinred.

Steepascendsinthetimelinecausedbymanyneweventtypesareanindicatorforanincidenthappeningintheliveofapatient(left).TheCMSprovideddata(right)mergesthreepatientsforanonymizationpurposeswhichmitigatestheimpactofincidentslettingsteepascendsmostlydisappear.

A user can use a lens to get information aboutinteresting events. The horizontal granularity ofsingle events is by day and the vertical position isdeterminedbythefirstoccurrenceoftheparticulareventinthehistoryofthepatient.

Thetoolprovedtobehelpfulforanalyzingtheoutputofmachine learning algorithms in predictive healthcareanalysis.Itprovidesaquickwaytoidentifytheproblemsof an automatically generated model by looking atpatients that were classified wrong or turn out to beoutliers in the cohort. Additionally, the tool enablesclinicians to quickly “tell the story” of the patient usingjust the administrative data often found in insuranceclaims.Asfutureworkweplantoworkmorecloselywithphysiciansand investigatehowwecan improveour tooltointegratemorewiththeirtasks.

Conclusions

CCS[2]hierarchyofeventtypes Generalpatientinformation Selectedevents

Costsperday

Time

Firstoccurrence

Oneevent(eg.Diagnosis593onJune5th)

Eventsofsametype(eg.Diagnosis250)

Hospitalstays

Our design is similar to LifeLines [3] withadditional information shown in overlays andusingasmallerverticalfootprintwhichallowsforalargernumberofdistincttypestobeshownatthe same time. Another notable similar work isthe tool VISITORS [4]which requires evenmoreverticalrealestatepertype.Every event is represented as rectangle whosecolor indicates its type, ie.adiagnosis (green),aperformedprocedure (orange), a laboratory testresult (blue),aprescribedmedication(purple),aphysician(pink),orahospital(brown).

VisualDesign