How Machine Learning is Helping us Get Smarter at Healthcare€¦ · How Machine Learning is...

Post on 22-May-2020

4 views 0 download

Transcript of How Machine Learning is Helping us Get Smarter at Healthcare€¦ · How Machine Learning is...

How Machine Learning is Helping us Get Smarter at Healthcare

ClearDATA’s Chief Technology Officer Matt Ferrari shares insights on:

•Machinelearninginthepubliccloud

•Industrysectorusecases

•Examplesofcloud-managedservicesformachine learningapplications

This whitepaper is designed for healthcare execs looking for a foundational understanding of what machine learning means to healthcare in the public cloud. These are the observations and opinions of ClearDATA’s CTO Matt Ferrari and does not imply endorsement of his comments from any of the three public clouds mentioned: Amazon Web Services, Google Cloud Platform and Microsoft Azure.

2H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

Fascinatingusecasesformachinelearningarespringingupacrossallhealthcaresectors.Machinelearningpromisestoimprovethequalityandspeedofcare,buildefficiencies,andofferbetterpatient-providerrelationships.Inthispaper,weexplorehowmachinelearningisbeingappliedtodifferenthealthcaresectorsacrosspubliccloudenvironmentssuchasAmazonWebServices(AWS),MicrosoftAzureandGoogleCloudPlatform(GCP).Weconcludewithadescriptionofmanagedservicesinthecloudthatcanhelpyourownorganizationreplacemanualorcumbersomeprocesseswithsmarter,fastermachinelearning.

Defining Machine LearningIn1950,AlanTuringproposedamethodtodetermineifamachinecould“think”inamannerindistinguishablefromahumanbeing.TheoriginalTuringTestrequiredonehumanandonecomputertocommunicateonlyviatextonacomputerscreen,withanotherhumanactingasjudgetodeterminewhichofthosecommunicatingwashuman,andwhichwasmachine.Sincethen,we’vecomealongwaywithartificialintelligence.InMay2018,attendeesatGoogle’sdeveloperconfabI/O2018watchedasGoogleDuplex1,Google’snewdigitalassistantmimickedthehumanvoiceinconversation,andmadephonecallstobookappointmentsonbehalfofthehumanrequestingthem.So,whatexactlydowemeanwhenwerefertomachinelearningtoday?Therehasbeensomedebateregardingauniversally-approveddefinitionformachinelearningordeeplearning,butinsimplestterms,itinvolves“training”computersystemstofind,rememberandactonpatternsindata.Andsimilartostatistics,machinelearningrequireslargedatasetstoreachmeaningfuloutcomes.Healthcareisanareathathasgottenveryproficientatproducingmassiveamountsofdata,muchmorethananyclinicianorphysiciancouldbegintoprocessalone.Assuch,machinelearningoutcomesstandtobenefitpatientsmorethanthesoloeffortsofoneclinicianorhealthcarepractitioner.

Machinelearningtodaycanbesupervisedorunsupervised.Withsupervision,thevariablesandoutcomesareknown.Weteachthemachine(thatis,createalgorithms)bymonitoringandensuringthetrainingmodelisreachingthedesiredresults.Wecanexplainhowtheresultswerereached,asthemachinessortthroughmassivedatasets.Withunsupervisedlearning,themachine’soutputsareunknownattheonset.Algorithmsareessentiallysetloosewiththedatatogleaninsightsthatcanbeverymeaningfulandimportant,buthardertoexplain.Supervisedismorecommonthanunsupervised,butbothmethodsaregrowing,andbothhavemerit.

Inthepast,activitiescontrolledbyartificialintelligenceandmachinelearningwereverylinearandtheendgoalswerequitebasic:programamachinetoaccomplishasimpletaskandthensimplypresstheEnterbutton.InputAandexpectOutputA.Thenewapproachistoprovidethemachineastart“state”withaspecificendgoal,thenallowthesystemtodevelopitsownintermediatestatesandletitprogressfromonetothenextuntilthatgoalisaccomplished.2

3H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

Whichspecificmachinelearningapplicationshelpuscollectmountainsofoncedisparatedataandproducemeaningfuloutcomesthatcanimprovehealthcare?Examplesfollowfromeachofthethreemajorpubliccloudstoday.Welookatthetechnologyrunningthesolution,whatthesolutiondoes,andwhyitmatterstohealthcare.

Amazon Web ServicesLiketheotherprimarycloudplatforms,Amazonisinvestingheavilyinmachinelearning.Inthelastyear,AmazonlaunchedLex,whichisdesignedtoenabledeveloperstobuildconversational

interfacesforapplications.ItispoweredbythesamedeeplearningtechnologiesasAlexa.Itusesadvanceddeeplearningormachinelearningforautomaticspeechrecognitiontoconvertspeechtotextandnaturallanguageunderstandingtorecognizetheintentofthetext--allofwhichhelpcreateengaginguserexperiences.

Amazonmachinelearning,orAML,providescompanieswithamethodforinterpretingdata.Itoffersvisualaidsandeasily-digestibleanalyticsreportsandisbasedinpartonAmazon’sexistinginternalsystemalgorithms.It’sveryaccessibleeventothosepractitionerswhomaynothaveextensiveexperiencewithdatascienceanditattemptstohelpbusinessesbuildmachinelearningmodelsoftheirownwithouthavetoinvestincountlesshoursspentoncreatingthenecessarycodetodosootherwise.3Bycomparison,inthepast,apractitionermayhaveusedDragontocreatespeech–to-text,thenaddnaturallanguageprocessingontopofthat.Today,publiccloudsareembeddingthisnaturallanguageprocessingstraightintotheirplatforms,makingthesamedeeplearningthatpowersAlexaavailabletoanydeveloperwhowantstoquicklybuildsophisticatedconversationalbotsorchatbots.Lexscalesautomatically,takingawaytheneedtobeconcernedaboutinfrastructurecapacity,andusersonlypayforwhattheyuse.It’sanimpressiveachievement.Expecthealthcareorganizationstoincreasinglyrequestthiskindoffunctionalitybehindbusinessassociated-coveredservicesandtomakeitHIPAAcompliant.Therearethird-partyservicesavailablethatprovideasecure,HIPAAcompliantenvironmentformachinelearninginhealthcarewhichwe’lllookatfurtherinthispaper’sconclusion.

AmazonalsolaunchedSageMaker,whichenablesdevelopersordatascientiststodeploymachinelearningmodelsatanyscaleandmovequicklythroughproduction,sotheycangaininsightsfromdatamuchsooner.SageMakerbypassestheneedfordatascientiststogothroughacquisitionprocesses,buyCap-exassets,andgetinvolvedwithinfrastructuredemandsthatareoutsideoftheirareaofexpertise.Instead,thesedatascientistsgaineasyaccesstotheirdatasourcesforanalysis,deployingamachinelearningmodelwithasingleclickfromtheSageMakerconsole.

4H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

WhileSageMakerisnotcurrentlyHIPAAeligible,Amazonismovingthatway,andinthemeantime,healthcareisbenefittingbyusingdepersonalized,anonymizeddatawithmachinelearningtolearnvolumesaboutpredictingandpreventingdisease.

Machinelearninghasmovedfromjustspeechandtextapplicationstoimaging.Thepossibilitiesformachinelearningoffimagesinhealthcareareprofound,withmanyusecaseshappeningrightnow.AmazonRekognitionisamachinelearningimagerecognitionservicethatrecognizesobjects,peopleandactivitiesinimagesandmedicalvideos.Currentlyusedpredominantlyforfacialrecognitionandsecuritytodetectunsafeorinappropriateimages,thepossibilitiesforhealthcarearetremendous,suchasdecisionsupportinoncologyandradiology.

Moreover,AmazonMachineImage4technologyprovidespractitionersandresearcherspre-installedtoolsthatcanscaleandintegratewithdeeplearningframeworks,whetherdevelopedbyAmazonorothers.Forexample,CognitiveToolkitbyMicrosoftTensorFlowisdevelopedbyMicrosoft,butmanyTensorFlowprojectsarealsorunningonAWSandGCP.AWSGlueisanothermachinelearningfriendlysolution.It’sanextract,transformandloadservicethatenablesuserstoeasilyprepareandsubmitdataforanalytics.Itisagoodexampleofoneofthenew“serverless”technologiesbeingadoptedbyhealthcare.Serverlessmeansnoinfrastructuretobuy,manageandmaintain–theenvironmentisautomaticallyprovisioned,andcustomersonlypayforwhattheyuse;inthiscase,todotheirdatapreparation.

Amazon’sautoscalingclusterisyetanothertoolthatspeedsuseofmachinelearning,byenablingdatascientiststorunamachinelearningalgorithmagainstmassivedatasetsinminutes.Anemerginguseistorunqueriesagainst“datalakes,”4withoutthepriororconcurrentneedtomoveorsynchronizethedata.

Google Cloud PlatformGooglehassignificanthistorywithartificialintelligenceandmachinelearning,includingitsGoogleCloudMachineLearningEnginethatingestsstructuredorunstructureddataintotheGoogleCloudPlatform(GCP).Theengine’strainingandpredictiveservicescanbeusedindependentlyortogether.Theinterfaceallowsdatascientiststoautomaticallydesignandevaluatetheirmodelarchitecturestoreachoptimalresults.Googlealsosupportstheimportofmachinelearningmodelsthathavebeen“trained”elsewhere.

GCPisoneofthepremiercloudcomputingplatformsonthemarkettoday.Whenitbeganin2008astheGoogleAppEngine,itwasusedmostlytoallowdeveloperstocreateapplicationsandhosttheminGoogle’secosystem.Soonthereafter,GooglelaunchedtheirComputerEngineapplicationtohelpsupporttheuseofvirtualmachines.GCPnowprovidesaverybroadspectrumofcomplementaryservicesincludingcomputing,storage,APIsystems,productivitytools,IoTservicesandofcoursetheCloud.5

5H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

GoogleCloudalsooffersbigdatasolutions,includinganintegrated,serverlessBigDataplatformthatcaptures,processes,storesandanalyzesdataallwithinoneveryscalableplatformfreefromtraditionalconstraints.LikeAWS,Googlehasitsownspeechdetectionandnaturallanguageprocessingtools.Italsooffersspecialtoolsforanonymizingdata.CloudDataprepcleansstructureddata(suchasfromaSQLdatabase)butcanalsouseunstructureddatainaserverlessenvironment.Thisisahugestepforward,asthehealthcareorganizationsidestepsinvestinginandmaintainingserversandsecurity.WithCloudDataprep,datascientistscanruntheirdataprepwithjustclicksinsteadofcode,simplydragginganddropping.There’satrendinthisdirection,whichspeedsthetimeittakesforthedatascientiststostartmakingdecisionsbasedontheanonymizeddata.Wewilltalkmoreaboutusecaseslaterinthispaper,butfornowitisworthnotingthatthishaspowerfulimplicationsforacceleratingthetimeittakestocompletedrugstudiesandclinicaltrials.

Awell-knownmachinelearningtechnologyisDeepMind,theartificialintelligenceventurebyGooglewhichhascreatedneuralnetworksthatlearnedtoplayvideogamesinafashionlikehumans.OfgreaterimportancetothoseofusinhealthcareisDeepMind’simpressiveabilitytoprocessimages.ThisincludesitsworkexaminingmammogramswiththeCancerResearchUKImperialCentre6,aswellasworkwithmaculardegenerationinagingeyes,andmuchmore.Byfindingearlywarningsignswecanofferproactivehealthcareoptionsratherthanattendingtomoredisastrousconsequencesforpatientslater.And,asisalwaysthecasewithmachinelearning,themoreimagesDeepMindprocesses,thesmarteritisgoingtobecome.

Micosoft AzureMicrosoftAzurewaslaunchedin2010andleveragedexistingMicrosofttechnologiessuchasRemoteApp,ActiveDirectoryandSQLServer.Essentiallyitisacollectionofvariouscloudcomputingservices.Azureisatemptingoptionforcompaniesasresourcecapacityisavailableondemandandasaproduct,itdoesnotrequireanyinitialcoststoimplement.Microsoftonlychargesfortotalresourceconsumptionasopposedtobillingforserverinstallationorleasingfees.TheAzureumbrellahasnowgrowntoincludemultipleservicessuchasanIoTsuite,amanagedsearchserviceandamediaserviceforcontentprotectionandvideoplayback.7

Microsoftwasoneoftheearliestpioneersinanalyticstools,baseduponwhathasbecometheCortanavirtualassistantthatusesvoicecommandtechnologyacrossmultipledevices.Microsoft

6H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

wasoneofthefirsttoofferserverlesscomputingwithnoinfrastructuretomanage.OnesuchtoolistheAzureMachineLearningStudio,whichMicrosoftdescribesas“acollaborative,drag-and-droptoolyoucanusetobuild,test,anddeploypredictiveanalyticssolutionsonyourdata.MachineLearningStudiopublishesmodelsaswebservicesthatcaneasilybeconsumedbycustomappsorBItoolssuchasExcel.”Designedforappliedmachinelearningwithbuilt-intemplates,datascientistsneednotwriteasinglelineofcodetoutilizethismachinelearningplatform.

Infact,oneoftheinterestingthingsaroundMicrosoft’smachinelearningcapabilitiesistheplethoraofpre-configuredenvironmentsitoffers.Mostdatascientistsaren’tinterestedinvetting,trialingandthendeployingthevarioustoolsneededtodrivedecisionsfrommachinelearning.Tothatend,Microsofthaspre-configuredenvironmentstogetstartedquickly,workingwithfamiliarprocessingsystemslikeAnacondaPythontoApacheSpark.

Microsoftalsohasadeeplearningvirtualmachinestraightoutofthebox,runningonAzure.TheDeepLearningVirtualMachinemakesitmorestraightforwardtouseGPU-basedVMinstancesfortrainingdeeplearningmodels.Lastly,Azure’smachinelearningcapabilitiesincludeenablingcustomerstobuildAIappsthatsenseprocessandactoninformation.Thiscanincreasespeedandefficiency,andcanleadtobetterhealthcareoutcomes.

Asmachinelearningandhealthcaremoveforward,expectAmazon,GoogleandMicrosofttomakemoreoftheirserviceseligibleforprotectedhealthinformation(PHI).Amazon,forexample,addsmoreservicesallthetimebehindHIPAAcompliance.Inthemeantime,healthcareorganizationsareanonymizingdatatousepubliccloudmachinelearningcapabilities.Themovementhasbegunandisgainingmomentumdailyofferingbenefitstothoseweserveindustrywide.

Use Cases by SectorWhileweareonlyatthebeginningofunderstandingalloftheusecasesformachinelearning,wearefarenoughalonginadoptiontohavemanynotableexamples.Let’sexplorethesebysector.

ProvidersSecuring Data

Fordoctorsandhospitals,agood‘bestuse’caseformachinelearningisincybersecurity.Aprimaryadvantagemachinelearningisbringingtoproviderstodayistohelpdefendagainstthegrowingthreatofransomware.Withhealthcare-basedransomwareattackshappeningeveryday8,cybersecurityfocusedonmachinelearningcanidentifypotentialvulnerabilitiesandpreventattacksbeforemaliciousactorsfindthem.Intelligentalgorithmsflagout-of-the-ordinaryactivityandsend

7H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

alertsfortimelyinterventions.Bycalculatingscoresforwherethedatashouldbemoving,andbywhom,andinwhatways,machinelearningcanhelptargetproblemareaslongbeforeanITorsecurityteammembermaynoticetheanomaly.

Providersseekingsuchmachinelearningexpertiseincybersecurityshouldlooktomanagedservicesvendorsthatdealexclusivelyinhealthcare—andcanprovetheirexperienceandwillingnesstoassumeriskofabreachviaacomprehensiveBAA.

In-depth Research

Imagingoffersanotherbestusecase.Machinelearningisprovidingconsistentlycorrectdeterminationsfromimagerecognition,andinmanycases,toahigherdegreethanahuman.Assuch,we’restartingtoseemajorinvestmentsinthisarea.Microsoft,forexample,issucceedinginthisareawithbiomarkersandphenotyping,andtheresultsaredrivingcancerresearchforwardasneverbefore.9Providerorganizationsarealsogatheringandcleaningtheirdatainanticipationofbeingabletouseitforimageanalysisbehindabusinessassociateagreement.Anidealmanagedservicespartnerforsuchausecasewouldhavehealthcareexpertise,HITRUSTCertification,andofferscalable,securedatamanagementinanyofthedominantpublicclouds.

Improving Patient Care

AthirdbroadadoptionexampleofmachinelearninginhospitalsandclinicssurroundstheuseofNaturalLanguageProcessing(NLP);particularlytoprocessinformationintheEHR.ClinicianscanresearchwhatprescriptionsapatientisonanduseNLPtosearchtheunstructureddataatscaletodeterminetheanswer.Theycanaskhowmanypatientsafacilityhaswithsimilarconditions.Inadditiontodrivingdowncosts,NLPcanhelpdriveupconsistencyofcareinhealthcarepractices.

Andforproviders,theabilitytousemachinelearningtoextractsignificantmeaningfrommassiveamountsofunstructuredtextishelpingwithcriticaldecisionmaking.Thedoctorornursewhoentersapatient’sroomtypicallybeginswithaseriesofimportantquestions.Whetherit’sbeingrecordedonaudio,notepadoriPad™,thereisgreatpotentialinthatdata,ifitgetsused.Accordingly,providersareinvestingalotoftimeandmoneyintomachinelearning,sotheycanaugmentandbetterpositiondiagnosis.Injustonesuchexample,machinelearningisbeingusedtodistinguishbetweenheartconditionsthathaveverysimilarresults.Thisreducestheriskofprovidersoptingforthediagnosistheyaremostfamiliarwithwhenadifferentdiagnosismaybethebestdecision.

8H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

Tosummarize,machinelearningismakingitpossibletoreadallthestructuredandunstructureddatasittinginEHRapplicationsandanalyzewhatisreallyhappening.Practitionerscansearchforcommonalitiesordeterminewhatkindoftraumaapatientmayhavebeenthroughtobetterdeterminethecriticalcareorlong-termcarethepatientneedstomanagetheirdiseaseorillness.Inadditiontoexistingillness,thepractitionercanmoreaccuratelypredictwhoisatriskintheneartermofcertaindiseases.Bybeingmoreaccurateinthediagnosis,thereislessrepeattreatmentandthefacilitycanoptimizeitscode,reducecosts,andmostimportantly,improvepatientoutcomes.

Payers Whenwepivottoinsuranceproviders,thebestusecasesformachinelearningchangedramatically.Theyarehighlyfocusedonpaymentandrevenuecyclemanagement,participationwithHIEs,andpatientengagement.

Revenue Cycle Management

Inonescenario,payersareusingmachinelearningtoquicklydeterminehowsoon,basedonaparticulardiagnosis,ahospitalgetspaidforservicesrendered.Howquicklydoesthepatientgettheletterinthemailthatsayswhattheirfinancialresponsibilityis,andhowquicklydotheypayit?Inanotherpatient-orientedexample,machinelearningcanbeusedbypayerstohelptheproviderbetterunderstandsomethingsaboutthepatients,suchaswhichonesarelikelytobe“noshows,”whicharelikelytoreceiveservices,butthennotpay,andmore.Payersarealsousingmachinelearninginparticipatoryhealthcarescenarios.Patientsarenowdoingtheirownresearchbeforetheyselectadoctor,andpayerscanhelpthemunderstandwhichoneswillprovidethemthebestcoverageontheirplan.

Improving Clinical Efficiency

Mergingtheclinicalandfinancial,healthcarepayersarealsofocusingonimprovingclinicalefficiencywhiletheyimproveclinicaloutcomes.Heretheyusemachinelearningtolookatwhichtreatmentsdeliverthemostclinicallyeffectiveoutcomeatthelowestcost.

Frauddetectionisanotherbestusecaseforpayersthatneedtoknowifthepersonsubmittingaclaimisinfactthatperson,anddidinfacthavethattreatmentorsurgery,especiallywithever-escalatingmedicalidentitytheft.Withreal-timeauthorizationsthatworkinawaysimilartothebankingindustrymodel,payerscanensuretheauthenticityofaclaim.

9H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

Pharma and Life SciencesWeseedeepadoptionofmachinelearninginthepharmaceuticalspace,primarilytoenhanceclinicaltrialsbyenablingpharmaandlifesciencescompaniestousebigdataandanalyticsenginestonormalizeandde-identifypatientdata.Theycanusethisdatatorunmachinelearning-driventrialstounderstandoutcomesofadrug,findingwaystoimproveitbeforebringingittomarket,allwithouthavingtotestthousandsofpatientswhomayhavenegativesideeffectsfromdrugtrials.Withtheabilitytomorequicklyidentifypotentialoutcomesanditerateandretest,researcherscanhelpspeedtomarketlife-changingtreatment.

Independent Software VendorsHealthcaretechnologycompaniesaremakinguseofmachinelearningtocreatemoreintelligentapps.ManyClearDATA®customers,forexample,usemachinelearningtounderstand,withspecificity,howthepatientordoctorusestheapp;inturn,thecompany’sdeveloperscanfurtherimprovefeaturesbasedonusability.

Ortheymayfeedtheirdataintoadatalaketoprojectandpredictscenariosandoutcomes;forinstance,toexaminethepathofapatientwhohasaspecificdiseaseandhasbeentoaneurologistandanophthalmologist.Machinelearningcandeterminethenextlikelysteptobestservethispatient.Andyes,allofthisdatacanbefedbackintothepatient’selectronicrecord.

Applications of Machine Learning in the Public CloudNowthatwe’velookedatwhatmachinelearningis,howthecloudsareexpandingmachinelearningopportunities,andthewaysinwhichvarioushealthcaresectorsarefocusingtheirmachinelearningefforts,let’sdivealittledeeperintoafewexamplesbasedonspecificobjectives.

10H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

Data preparation

Overthepastseveralyears,biomedicalresearchershavefiguredouthowtousemachinelearningtoperformstudiesbasedonde-identifiedpatientdataderivedfromelectronichealthrecordsandothersources.Themajorpubliccloudproviders–Amazon,GoogleandMicrosoft–nowenabledatascientistsforpharmaceuticalfirmstodosuchresearchinaHIPAA-compliantenvironmentwithoutbeingmachinelearningexperts.

Oneofthemajorusecasesforthisapproachisresearchthatseekstounderstandhowgeneticdifferencesdeterminetheresponsesofindividualstoparticulardrugs.Toaccomplishthisgoal,dataonthephysicalcharacteristicsandmedicalhistoriesofindividualpatients,knownasphenotypes,arecomparedtotheirgeneticfeatures,knownasgenotypes,whicharederivedfromgenomicsequencing.

Beforemachinelearningwasusedforthistask,phenotypeshadtobederivedmanually--averylaboriousandtimeconsuming10job.Machinelearning,incontrast,allowedphenotypestobegeneratedusinganautomatic,unsupervisedandhigh-throughputprocess.Butresearcherswhowantedtousethisapproachinonsitedatacentershadtobuyadditionalequipment,hiredeveloperswiththerequisiteskillset,andbuildtheirownalgorithms.

Today,apharmaceuticalfirmcanbuythesameinfrastructureasaservice(laaS)inapubliccloud,whichcanbescaledasneeded.Machinelearningalgorithmsandrelatedservicesinthecloudareprebuiltforspecificpurposes.TheGoogleCloud,forexample,hasproductscalledCloudDataLab,whichexploresbothstructuredandunstructureddatasets,andCloudDataPrep,whichcleansandpreparesdataforanalysisandcanalsode-identifyclinicaldata.

Large scale data analysis

Thelifesciencescommunityisveryfocusedongenomicresearch,whichpromisestorevolutionizeourunderstandingofhealthandillness.Thechallengeisthemassivescaleofgenomicinformation.Ittakesapetabyteofdata,forinstance,torepresentasinglehumangenome.Tostoreandanalyzethisamountofdatainadatacenterwouldbeverydifficultforanyorganization.Sopubliccloudsareincreasinglybeingutilizedforgenomicresearchandareprovidingmachinelearningtoolstoanalyzethedata.

TheNationalInstitutesofHealth’sAllofUsproject11,inconjunctionwithacademicmedicalcenters,aimstoanalyzegeneticinformationfromamillionpeopleisthebest-knownoftheinitiativeslaunchedtodiscovercuresthroughgenomicresearch.Manyhealthcareorganizationsaredoingsimilarworkonprecisionmedicine,usingmachinelearninginthecloud.

Forexample,theColoradoCenterforPersonalizedMedicine(CCPM),abranchoftheUniversityofColoradoDenver,ishelpingdoctorsevaluatepatients12atthemolecularleveltopredicttheirriskfordiseaseandtodeveloppersonalizedtherapiesbasedontheirDNA.Thisrequireslarge-scaleanalysisoftheirgeneticsandthehealthhistoriesofthousandsofpatientstofindpatterns.

11H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

Becauseon-premisestechnologyistooexpensiveforthiskindofresearch,CCPMmigrateditsdatainfrastructuretoaHIPAA-compliantpubliccloud.

CCPMisalsousinganalgorithmicdatabaseapplicationtoaccelerateitsresearch.Thisapplicationhasreducedthetimerequiredforresponsestodataqueriesby97%.

Algorithm-driven diagnosing

Cardiology,radiology,andpathology,whichuselargeimagedatabases,arenaturalcandidatesformachinelearning.Algorithmshavebeendeveloped,forexample,toidentifyabnormalitiesinimagesthatmayindicatethepresenceofdisease.Thesealgorithmsareasaccurateasandsometimesmoreaccuratethanphysicianswhoaretrainedto“read”theseimages.

Google,forinstance,attractedmediaattentionin2017whenitusedmachinelearning,predictiveanalyticsandpatternrecognitiontodetectbreastcancerbylookingforcellpatternsintissueslides.Astudyshowed13thatmachinelearningcouldidentifybreastcancerwith89percentaccuracy,comparedtothe73percentscoreachievedbyahumanpathologist.

Similarly,StanfordUniversityresearchersusedadeeplearningalgorithmtoidentifyskincancer14aswellasdermatologistsdid.Thecomputerscientistsmadeadatabaseofnearly130,000skindiseaseimagesandtrainedtheiralgorithmtovisuallydiagnosepotentialcancer.Insteadofwritingcodetotellthecomputerwhattolookfor,theytrainedthealgorithmtodecidewhichlesionswerecancerous.

Sepsis prevention modeling

Afewyearsago,alargemedicalresearchuniversitysetouttopredictwhichpatientswereatriskofdevelopingsepsis,ahighlydangerousconditionthatisoneoftheleadingcausesofmortalityinhospitals.Usingmachinelearningtoolsinthecloud,researchersfiguredouthowtoidentifythesepatients24hoursearlierthancouldhavebeendonewithtraditionalmethods.

Theteamthatdevelopedthisclinicalprotocolstartedwithanexpertmodelthatusedspecificthresholdsoftemperature,heartrate,respiratoryrate,andwhitebloodcountaskeyindicatorsofsepsisrisk.Afterloadingintheavailabledataona

patient,includingthedataintheirEHR,thecomputerusedanalgorithmtodeterminehowcloselyapatient’scharacteristicsmatchedthoseofpatientswhohadpreviouslydevelopedsepsis.Whenapatientmatchedtheprofile,theirphysicianreceivedanalert,actedonitordidn’t,andfedtheirreactionbacktothealgorithmtoimproveit.Overtime,thisearlywarningsystembecamehighlysensitiveandisnowusedroutinelyatthatuniversity.

12H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

Managed Cloud Services for Machine Learning ProjectsAssophisticatedastheaboveusecasesmaybe,almostanyhealthcareorganizationcanembarkonamachinelearninginitiativewithouttheneedtoaddITinfrastructure.Asafull-servicecloudprovider,ClearDATAisattheforefrontofhelpingorganizationsofeverysizefromeverysectorshedtheconstraintsofoutdated,expensiveinfrastructureandparticipateinthecloudeconomy.

ClearDATA’smulti-cloudabilitiesaresomeofitsmostrelevantandusefulfeatures.However,thatflexibilityrequiresextrascrutinyaroundsecurity.Purpose-builtsafeguards,automationandhealthcareexpertisearejustafewofthetoolsleveragedbytheplatformtoensurefullprotectionofsensitivehealthcaredata.Astheplatformwasdesignedexclusivelyforusewithinhealthcareorganizations,itcomesstandardwithcomprehensive,real-timesupportfromtheClearDATAteamandisbackedbyanindustry-leadingBAA.

Additionally,theplatformisalsoHITRUSTCertifiedensuringthatClearDATA’scloudcomputingandbackupservicesmeetthehighestindustrystandardsformanagingprotectedhealthinformation.

Moreover,inadditiontopossessingmultiplecertifications,ClearDATAisalsofullycompliantwithHIPAA,GxPandGDPRregulationsandframeworks.Mitigatingriskiseasilymanagedviathecompliancedashboardwhichallowsuserstomonitorsthousandsofcomponentsacrossmultiplecloudenvironments.Thedashboardisaneasywaytogetaglanceintothestateoforganizationalcompliancewheneverit’sneeded.Lastly,thesegreattoolsandfunctionsarefurtherstrengthenedwiththesupportofClearDATA’sarrayofprofessionalservices.Whetherthechallengeisdevelopingaprocessdesign,customdevelopmentorariskassessment,ClearDATAdeliversextensivehealthcareandITexpertisetoprovidethenecessarysolutionstoresolveanyissue.Evenwithpotentiallyoverlookedneedssuchasbreachsimulation,penetrationtesting,securityauditsupportandapplicationcodereviews,ClearDATAappliesahands-onconsultativeapproachtoeverychallenge.ClearDATAalsooffersserviceconsultationsforvarioustypesofprojectspecifications.

13H O W M AC H I N E L E A R N I N G I S H E L P I N G U S G E T S M A RT E R AT H E A LT H C A R E |

ConclusionMuchlikehowAmazonPrimehascompletelyredefinedconsumeraccesstoon-demandentertainment,machinelearningisalsoredefiningthestateofhealthcarein2018.Machinelearningallowshealthcaresystemstobreaktheirnormalwayofthinkingintermsofexistingmodelsandallowsclinicalteamstotestforpotentialoutcomesusingmorecomplexalgorithmsthantheyevercouldbefore.Machinelearningisultimatelyagame-changingtechnologythathealthcaresystemsneedinordertohelpovercomethechallengestheyencountereveryday.15Whiletherewasatimewhentalkofmachinelearningwasfuturisticandvisionary,thehealthcareindustryhasmovedpasttheideatoactualbroadadoption.Allsignspointtoincreasinglydeepuse,withtheopportunitiesonlylimitedbyourcollectiveimaginationsandexpertise.Theoutcomestandstobenefitpatientsaswellasproviders,payersandthoseworkingtoservethem,asweallworktomakehealthcaresmarter.

Resources1Rawes,E.(2018,July2).WhatisGoogleDuplex?DigitalTrends.Retrievedfromhttps://www.digitaltrends.com/home/what-is-google-duplex/

2Drepper,U.(2017,Sept20).Anintroductiontomachinelearningtoday.Retrievedfromhttps://opensource.com/article/17/9/introduction-machine-learning

3Reese,H.(2016,Dec.1).ShouldAmazonbeyourAIandmachinelearningplatform?Retrievedonfromhttps://www.zdnet.com/article/should-amazon-be-your-ai-and-machine-learning-platform/

4AWS.(n.d.).Whatisadatalake?[Blogpost].Retreivedfromhttps://aws.amazon.com/big-data/datalakes-and-analytics/what-is-a-data-lake/

5Forrest,C.(2018,May14).GoogleCloudPlatform:Acheatsheet.Retrievedfromhttps://www.techrepublic.com/article/google-cloud-platform-the-smart-persons-guide/

6DeepMindHealth.(2017November24).DeepMind,theCancerResearchUKImperialCentreandOPTIMAM[BlogPost].Retrievedfromhttps://deepmind.com/applied/deepmind-health/working-partners/health-research-tomorrow/cancer-research-imperial-optimam/

7Sanders,J.(2018,June18).MicrosoftAzure:Acheatsheet.Retrievedfromhttps://www.techrepublic.com/article/microsoft-azure-the-smart-persons-guide/

8Donovan,F.(2018,May3).HealthcareIndustryTakesBruntofRansomwareAttacks.HealthITSecurity.Retrievedfromhttps://healthitsecurity.com/news/healthcare-industry-takes-brunt-of-ransomware-attacks

9Thatcher,L.(n.d.).HowMachineLearninginHealthcareSavesLives.Retrievedfromhttps://www.healthcatalyst.com/how-machine-learning-in-healthcare-saves-lives

10JetteHenderson,MS,1RyanBridges,BS,2JoyceC.Ho,MA,PhD,3ByronC.Wallace,PhD,4andJoydeepGhosh,PhD1.(2017,Jul26).PheKnow–Cloud:AToolforEvaluatingHigh-ThroughputPhenotypeCandidatesusingOnlineMedicalLiterature.AMIAJtSummitsTranslSciProc.Retrievedfromhttps://www.ncbi.nlm.nih.gov/pmc/articles/PMC5543339/

11NIH.(n.d.).AllofUsResearch.[Blogpost].Retrievedfromhttps://allofus.nih.gov/

12Google.(n.d.)ColoradoCenterforPersonalizedMedicine:Integratinggeneticdataandpatientrecordsinthecloudtopersonalizecare.[Blogpost].Retreivedfromhttps://cloud.google.com/customers/colorado-center-for-personalized-medicine/

13Stumpe,M.(2017,March3).AssistingPathologistsinDetectingCancerwithDeepLearningPostedbyMartin,TechnicalLead,andLilyPeng,ProductManager.Retreivedfromhttps://ai.googleblog.com/2017/03/assisting-pathologists-in-detecting.html

14Kubota,T.(2017,January25).Deeplearningalgorithmdoesaswellasdermatologistsinidentifyingskincancer.Retreivedfromhttps://news.stanford.edu/2017/01/25/artificial-intelligence-used-identify-skin-cancer/.

15Lee,P.(2018,Feb28)Microsoft’sfocusontransforminghealthcare:IntelligenthealththroughAIandthecloudMicrosoft[BlogPost].Retrievedfromhttps://blogs.microsoft.com/blog/2018/02/28/microsofts-focus-transforming-healthcare-intelligent-health-ai-cloud/