Levi:
Higuys,thanksforjoiningus.Thisisthehands-onmachinelearningbroadcast.
ThisisTaylorLarson.
Taylor:
Yup.
Levi:
I'mLeviThatcher.Taylor'sjoiningusforthefirsttime.He'sonthedatascience
team.
AndTaylor,howareyoudoing?
Taylor:
Yeah,I'mdoinggood.
Levi:
Good,good.
Taylor:
Thanksforhavingmeontoday.I'mjustgoingtositinforMiketoday.I'mnota
Mike2.0but—
Levi:
Yeah,yeah.No,—
Taylor:
--gladtobesittingin.
Levi:
We'regladtohaveyouhere.Sowhat'sgoingon?
Taylor:
Notmuch.
Today,we'vegotsomeexcitingthingsontab.But,beforewediveintoallofthat,
Ijustwanttomentionthatifyou'dliketocontributetothechatwindowbesure
tologintoYouTube.We'regoingtomaybebesharingacoupleofthingsonthe
screen,soadjustyourvideoresolutionifyouneedtoanddon'tforgettosubscribe
tothechannelaswellasourblogso.Thatway,canfollowonandgetsomegood
notificationsaswegoso.
Levi:
Yeah.OntheYouTubethough,youcanturnupthehighresolution.Ithinkthat
kindofhelpswithwhatweshowhere.
Taylor:
Yup.
Levi:
Soyeah,checkthatout.Andthen,intermsofwhat'sonthedocket,mailbagfirst?
Taylor:
Yup,mailbag.Then,we'regoingtodosomemachinelearninginthenewslikeyou
guysnormallydo.Andthenwearegoingtochataboutsomehealthcareusecases
forhealthcare.ai.Andthenwe'regoingtoopenupthechatandjusttryandget
anyquestionsansweredorkindofhavesomefeedbackgoingon,soitshouldbe
good.
Levi:
Yeah,yeah.That'sawesome.Andwewanttosaythatthisisnothingwithoutyou
all.Wewanttoreallymakethisacommunityandlearnfromyouandhearlike
whatproblemsyou'refacinginyourdatawork.So,please,logintothechat.Let
usknoworyoucanemailus.There'sacontactpageonHealthcare.ai,socheck
thatout.
Andlet'sgetstartwiththemailbag.DoyouhavethosequestionsuporshouldI
pullthemupoverhere?
Taylor:
Yeah,I'vegotthem.
Levi:
Cool.Let'sseewhatwe'vegotthisweek?
Taylor:
Yeah. So, the first one we had, that one came from Samir - emailed to
healthcare.ai.Hewasworkingonaprojectinahealthcaresystemandhassome
real-timedata-thingsliketemperature,heartrate,oxygenlevel,andwantedto
knowwhattypeofmodelwouldbeagooduseforthatandkindofhowtosetit
upso.
Levi:
Oh,that'sareallyinterestingusecase.
Taylor:
Yup.
Levi:
So, of course, in a health system when you're dealing with an EMR or health
metrics in general your measurements - the things that you're using to train a
model,thesecouldbehourly,daily,orevenweekly.Andso,howdoyoukindof
like--howdoyoustartwiththat?Howdoyoukindofgetthemalltogether?Like
what—
Taylor:
Yeah,itgetskindofcomplicatedbutit'sexcitingdatatohave-potentially,some
reallyusefuldata.Butyeah,ifyou'vegot--differentpeopleinthehospital--one
person might be getting their temperature checked every couple of hours, so
they've got many, many records of their temperature. Where then, someone
mightjusthavetheirtemperaturecheckedacoupleoftimesduringtheirhospital
stay.Andso,you'vegottoaggregatethatdatauptosomesortofgrain,kindof,
forthemachinelearningalgorithmtobeabletohandleithandleitinastandard
way. And so, some ways that we like to do that include things like maybe the
earliesttemperaturereading,thelatest,thehighest,thelowest.
Levi:
Ohyeah,allsortsofthingsyoucandowiththat–themean.
Taylor:
Yes.
Yup,exactly.Justgetittosomesortof,youknow—sothatyourcolumnsdon't
havetobeinfinitelywidebecauseyoudon'tknowiftheperson'sgoingtohave4
readingsor24readings.
Levi:
That'sagoodpoint.
Taylor:
Yup.
Levi:
So,Ihadneverreallyheardofthatconcept“grain”beforeIcametohealthcare.
So,TaylorcomesfromColorado.AndwhenyouwereColoradoMedicAid,wasthat
conceptused?
Taylor:
Yeah. Yeah. You know, you kind of have some examples of grains are like the
patientlevel.Maybeyou'realsotalkingaboutthevisitlevel.So,onepatientcan
havemultiplevisits.Thenonevisitcouldhavemultipledays,soyoumighthavea
daygrain.Andso,that'skindoftheconceptofgrain.Andthinkingaboutthatyou
wantallofyourrowsformachinelearningtobeatthesamegrainsothatyoucan
feeditinaccordingly.
Levi:
Okay,that'sawesome.
Taylor:
Yeah.
Levi:
Yeah. So basically what does a row represent? Is that a an hour, a day, or an
encounter–basically?That'sgreat.
Taylor:
Yeah.AndIthinkthatthoseotherdatafieldsthatSamirmentioned-heartrate
andoxygenlevel,thosewouldkindoffollowthesameformat.Andyoujustsort
ofreallywanttothinkaboutandprobably-atleast,inmycase,I'dneedtochat
withsomeonethatknowsmuchmoreabouttheclinicalsettingthanIdoandask
forsomeadviceonwhatwouldbethemostclinicallyrelevanthandfuloffields
thatyoucankindofstartwith?Andthentestthoseinyourmodeltoseehow
thosehelp.
Levi:
Yeah, it's definitely helpful when you can kind of pair the analyst and the data
analysiswiththeclinicalexpert.That'sthewaywetrytodothingshereatHealth
Catalyst.
Anyotherquestionscomeacross?
Taylor:
Yeah,yeah.Wehadacoupleothers.Anothergoodonethatwe'dactuallyseenon
acoupleofoccasionswas“Howdoesforecastingdifferfrommachine-learning?
Andthatcanthesamealgorithmsbeused?”
Levi:
Yeah,that'sagreatquestion.So,we'veheardthatalotandforecastingsounds
pretty similar to prediction and machine-learning. You're looking out into the
future. And how we think about it is, with machine learning, we often are
predictingkindofa“yes–no”.“Willthispatientget[inaudible00:05:47]kindof
on a personal level. Whereas, forecasting seems to roll up to, “Okay, well how
manybedswillbeutilizedinthisdepartment,thisday?”Almostmoreoflikea
summarytypecalculation.Andperhapsit’llbringinseasonaleffectsalotmore
thanyouwouldwithtypicalmachinelearning.So,youknow,substancedoesn't
carewhatdayoftheweekitis.Butifyou'relookingtopredict,“Okay,wellhow
busyisthisdepartment?”You'lldefinitelywanttobringinthingslikemonthof
year,dayoftheweek-thingslikethat.
Taylor:
Yeah.Yeah.Anditseemstobeespeciallyapopularusecaseforoperationsfor
financetoknow“Howmanybedsaregoingtobeused?Howmanypeopleare
goingtocomeinwiththisspecificcondition?”Sothattheycan,kindof,doabit
moreresourceplanning.Or“whatisourcostforthisdepartmentgoingtobefor
acertainamountoftime?”Andyeah,it'sdefinitelykindofalittlebithigherlevel,
kindoftheaggregationofthosekindsofgrainsthatImentionedearlier.
Levi:
Yeah,soyoucanusethesamealgorithmsbecause,essentially,justrollingupto
differentgrain.
Taylor:
Yup,yup.Ithink,aslongasyousetyouroutcomevariableupcorrectly,tobeout
inthetimeframethatyouactuallyneed–howlikelyisthispersontocomebackto
the hospital within 30 days, you know?” Now, you know when they've been
discharged.Youknowwhattheirlikelihoodis.Andyoukindofcreatethatkindof
amoving30-daywindoworsomethinglikethat.Andatleasttomethat's—
Levi:
Yeah,yeah.
Taylor:
Theuseofforecasting—
Levi:
No,exactly.
Taylor:
Withthealgorithmsthatalreadyexist.
Levi:
Yeah,soitseemsreallyvaluableinmoreoftheoperationalandfinancialsense—
Taylor:
Especially—
Levi:
Alotmorethanintheclinicalsense.
Taylor:
Yeah,yeah.Alotoftheusecases,whichwe’llkindofgettoalittlebitlaterinthe
show,endupreallywantingtofocusonthepatientwhichIthinkisreallygreat
butthenthere'sotherpartsofthehealthsystemthathavetobeaccountingfor
thewholesystemandplanningaccordingly.
Levi:
Yeah,yeah.Fantasticquestions.Keepthemcoming,guys.Sowereallyappreciate
youreachingoutandlovediggingintowhatyou'reuptooutthere.
Taylor:
Yeah.
Levi:
Anythingelsegoingon?
Taylor:
Yeah,wehaveonemoreaboutdefiningandthetrainingversusthetestingdata
forthe—it'ssomethinglikea30-dayre-admission.
Levi:
Yeah,that'sagreatquestion.Sowheneveryou'redoingpredictionyou'llhavea
trainingsetandatestset.We'vegonethroughthisalittlebit.Butonahighlevel,
thealgorithmneedstolearnfrompastfolks-whatweretheirattributes?Didthey
haveagoodoutcomeornot?Andthenyou'repredictingonadifferentsetoffolks.
And the idea with this training set is those are the people that you're learning
from.Andthetestsetarethosepeoplethatyou'regivingapredictionfor.Soeach
day,ifyou'repredictingcentrallineinfection,yourtestsetarethosepeoplethat
haveacentrallineinthem.Andthetrainingsetareallpastfolksthathavehada
centrallineinthepast.
Taylor:
Yup.
Levi:
And so, when you've done this for a 30-day re-admission. It’s fairly
straightforward,justkindofwithatimewindowthere.
Taylor:
Yeah.There'sacoupleoftimewindowsthatcomeintoeffectwiththe30-dayreadmission,kindofdependingontheusecasewhereyou'vegotpatientsthatare
stillinthehospital.You'dliketoknowhowlikelytheyaretobere-admittedafter
they've been discharged. And then you also have patients that have been
dischargedbutthey'restillwithina30-daywindow.Theyhavenotyetbeenreadmittedtothehospitalbutthey'reoutsideofthehospitalsomaybethereare
someotherinterventionsthatcomeintoplay.
Basically,thetestwindowforthatwouldbeanyonewhohasnotbeendischarged
morethan30daysandwhohasnotalreadybeenre-admitted.Andthosewould
bethepatientsthatwouldbeyourtestwindow.Youwouldpredictonthem.And
everyoneelsethateitherhasalreadybeenre-admittedorisoutsideofthat30
days-thosewouldbeyourtrainingset.
Levi:
Yeah.That'sawesome.That'sareallypracticalquestion.Andyoucankindofapply
thoseprinciplestomanydifferentusecases.
That'sthenicethingwithmachinelearning.Youreallydon'tneedaPhDaboutit.
Youcankindoflearnhowtheseconceptsworkandthensay,”Okay,well,with
this particular business question, let's map what we’re doing in the past, what
we'redoingtoday.”Andalotofthesethingsareverysymmetricacrossusecases.
Taylor:
Yup.Andifyoudefinitelyjustkindoflookcloselyatyourusecase,relyonyour
personalexperiencewiththedataandjustmakesurethatitmakeslogicaland
practicalsense.Andthat'skindofthebestwaytostart.Andthenyoucanworkon
refiningthatdefinitionalittlebitifyouneedto.Butrelyingonyourownexpertise
isagreatplacetostart.
Levi:
Yeah,yeah.Thankssomuchguysforreachingout.Welovehearingfromyou.
Andagain,whetherit'sinthechatwindoworinthecontactpageonourwebsite,
criticism,dress-fashionsenserecommendations.Icouldusethat—
Taylor:
Yeah.Idon'tknowifthisshirtwastoobusy.
Levi:
--sothat’snice,that’snice.
Taylor:
ButIthrewitonanyway.
Levi:
[inaudible00:10:33].Ilikeit.
Taylor:
Yup.
Levi:
We'rehappytotakeanythingyou’vegotoutthere.
Andso,that'sthemailbagfortoday.Sowhereshouldwegonext?
Taylor:
Yeah. And now we're onto machine learning and the news. And I'd heard
somethinginthenewsthatGooglehadacquiredKaggle,isthatright?
Levi:
Kaggle.Yeah,yeah.
Taylor:
WhatisKaggle?
Levi:
Kaggle.So,notKegel.Youknow,there’ssomemispronunciationoutthere.
Taylor:
Yeah.I'dhearditonapreviousshowthatIwaswatchingsoIdidn'twanttogo
downthatpathagain,so.
Levi:
Yeah.Yeah.Ibelieveit'sKaggle.
Taylor:
Yeah.
Levi:
Okay. And so Kaggle – we’ll show here on the screen. Kaggle is this amazing
machine-learningwebsitewhereyoucangetyourhandsdirtywithdataandyou
canfinddatasetsthatyoucandownloadandplaywith.Andwhatyouessentially
do is there’s a little competition where you compete against other folks in the
field.Thesecouldberesearchers,orpeopleintheirbasement,orwhoever.And
you'realltryingtopredictsomethingthat'sinthisdataset.
Andso,ifwegotothiscompetition’stabreallyquick,wecankindofshowyou
howthatworks.There'salwaysabunchofcompetitionsopenatonetime.Andas
wescrolldown,you'llnotice,okay,well,somethingtodowithfisheriesthatyou're
predictingorYouTubevideopredictions.Andyou'llnoticethere'squiteabitof
prizemoneyinvolvedinthesecompetitionsbutthere'salsoonesthatarejusthere
forpractice.
That'sactuallyhowIgotmystartinmachinelearning.AndIfinishedmyPhD,went
intoananalystroleandwaslearningSQL.AndIwaseagertolearnaboutmachine
learningandIwasgettingkindofthisbuzzaroundit.AndIwaslike,“Wait.Well,
itsoundslikestatistics.”
AndhereatKaggle,youcanpracticethethingsyou'rehangingontheshow.You
cantrydifferentalgorithms.Andso,checkitout.It'sfantastic.
AndGoogleboughtit.Youknow,they'rebuyingquiteafewtypecompanies.So
they bought DeepMind a couple of years back which is a British AI firm that
actuallytaughtacomputerhowtoplayGo.HaveyoueverheardofGoorplayed
Go?
Taylor:
Idon'tthinkso.
Levi:
IhadneverheardaboutiteitheruntilIreadabouttheworktheyweredoingin
Britain but Go is this ancient Chinese game. It's fairly simple but like has an
incrediblenumberofperturbationsaboutit.Andso,it'sreallyquiteafeatthatthe
computerlearnedhowtoplayGowell.Butgettingback—
Taylor:
Isthattheonewhereitgotmoreaggressiveovertimeorsomethinglikethat?
[inaudible00:12:54]
Levi:
Oh,oh.Yeah,yeah.
Taylor:
[inaudible00:12:56]differentgamebutyeah.
Levi:
That'snotsomethingrelatedtothatracistbotoutthere,right?Therewasaracist
botIheardthat--
Taylor:
Yeah.No,Idon'tknowaboutthatone.
Levi:
Yeah,thatwastheaggressivethingthatlikeconnected.
Taylor:
Okay.
Yeah,withthatKaggleandGoogle,doyouhaveanyotherthoughtsonmaybewhy
theydidtheacquisition?I'dreadalittlebitthatthey'reinterestedingettingaccess
tomoredatascientistsand—
Levi:
Yeah.
Taylor:
Kindofgettingafootinthedooronthatgrowingfieldso[inaudible00:13:23]
Levi:
Yeah,yeah.AndKaggle-that'sagoodquestion.SoKaggle’skindofthecenterof
public data science, you can think of it as. They have data sets. They have
competitions.Theyhavekernelswhichmightsoundalittlebitabstractbutifyou
gotothekernelstab,itbasicallygivesyouabunchofdifferentsetsofcode.They'll
letyouseehowtodoanalysisonthisorthatkindofdata.
Taylor:
Oh,cool.
Levi:
Yeah,inPython,andRandotherlanguages,sodefinitelycheckitout.
Taylor:
Yup.
Levi:
Sothat’swhat'sinthenewsthisweek.Sotheheartoftoday'sshowisgoingtobe
aroundsomeoftheworkthatTaylorandI'mdoingandothersonthedatascience
team where we've been actually going across a lot of different use cases in
healthcareandsaying,“Well,howdoesthemachinelearningfitintothisorthat?
Andwhattypeofmachinelearninggoesinhere?Andhowdoyouactuallymake
animpactontheseoutcomesimprovementprojects?
So,Taylor,doyouwanttorundownacoupleofthesedifferentusecasesthatyou
findthatarelikedifferentcategoriesofthem?
Taylor:
Yeah,yeah.Ifeltlikeatleastinthetypeofworkthatwedowheremostofour
workisaroundkindofanoutcomesimprovementmethodology.Andso,allofthe
projectskindoftieouttoafewdifferentcategoriesthatI'llgooverbuttheyall
haveto,ofcourse,drivesomesortofoutcomeeithersomesortofoperational
efficiency,financialefficiencyandthendefinitelytheonethatwefocusonalotis
clinicaloutcomesimprovement.Sothosearekindofthethreecategories–the
operationsandperformancemanagement,finance,andclinical.AndthenIthink
thatwecanbreakclinicaldownevenalittlebitfurtherintoacuteandchronic.So
ifyouwant,Icanjustchataboutacoupleofusecases—
Levi:
Yeah,yeah.Let’s—
Taylor:
Andwecouldevendiveintoonedeeper,ifyoulike?
Levi:
Yeah.No,that’sawesome.
Taylor:
Cool.
Levi:
It’skindofnicetogetthehigh-levelcategoriesandmaybekindofexploresome
ofthesub-categoriestherein?
Taylor:
Yeah.Forsure,yeah.
Soanexampleofanoperationsandperformancemanagementonemightbekind
of at the practice level - the outpatient practice level around appointment no
showswherethere'ssomelostefficiency.Andsomefolksthataren'tgettingthe
proper follow-up care, if they're not showing up to their primary care
appointments.Andso,that'sagreatspotforpredictingwhetherornotpatients
will show up to their appointment. And then either use that to increase the
interventionofmaybecallingthem,orfollowingup,orhelpingworkaroundtheir
scheduleortransportationneeds,orimprovingyourschedulingandefficiencyas
farasdoublebookingcertainslots.Andthere'salotofdifferentwaystouse-what
seems like one prediction can have many different use cases within that
application.
So, another area is finance. Like I mentioned, propensity to pay - how likely
someoneistopayfortheirmedicalcare.Andthisisanareathatcanalsohelpwith
findoutwhatresourcesmaybeapatientneedsthehospitalsystemtohelpwith,
whetherit'smaybetimingofthepayment,orsomeothersortsofarrangements,
orjustmaybeareminderthatthebillhadgoneouttoabadaddressandyoucan
followupso.
Levi:
Yes,alotofdifferentbenefitsrightthere.It'snotjustaboutthehospitaldeciding
whotodenycareto.It’smore,Iwouldsay,mutuallybeneficialaswithalotof
theseimbalancesthataremaybe—sayit's$6000andthepersoncan'tpaythat.
But maybe if they do some sort of balance modification the person could pay
$3000andthenbothpartiesbenefit.
Taylor:
Exactly.Noteverythingisgoingtobaddebt.Youcouldhavejusthavealotbetter
outcomesforthehospitalsystemandforthepatientwhichiskindoftheultimate
goal.
Andthenontheclinicalside,there'ssomereallyinterestingusecasesspecifically
aroundacutecare.Hospital-acquiredconditionsisaninterestingone.Thingslike
CLABSI which I know we've touched on before. Hospital-acquired pneumonia.
Sepsis-that'saninterestingusecase.
Andthatcouldbeanythingfrompredictingwhetherornotthepatientwillget
thatconditionorhelpingidentifythattheyhavethatconditionandmaybewere
misdiagnosedwithsomethingelse.Andso,helpingkindofhoneinonabitofabit
ofaregistrywhichisreallynice.Andthen,ofcourse,youcantiesomeofourmost
commonoutcomevariablestothosecohortsaswell,sopredictinglengthofstay
foranacutecondition,predictingre-admissionsafterthey'vehadthatcondition.
Levi:
So,it’skindoflikethismatrix,youcankindoflikeapplyalotofdifferentscenarios
forfocusingonheartfailure,forexample.
Taylor:
Yup.
Yeah,soyoucangofrompredictingtheconditionitself.Asimilarthingwouldwork
forchronicconditionsorpredictingsomesortofoutcomefollowingaprocedure,
aconditionorwhateverhaveyou.
And like I said, same with on the on the chronic side, you could use predictive
analytics to help with registry prediction or any of those other outcomes. And
we'vedefinitelygotintoalotofthosebefore–COPD,heartfailure,diabetes-
thingslikethat.
Levi:
Yeah,that'sfantastic.So,itreallyboilsdowntowhatismostbeneficialtoyour
businessgoals.Ofcourse,you'relookingtoincreasetheefficiencyofpatientcare
andthehospitalslovesavingmoney,helpingthebottomline.Andso,wekindof
letthatdrivethediscussionhere.We'retheonessaying,“Okay,well,it'sgreatto
createmodelsbutifthey'renotdoingthingsthatarehelpfultothebusinessthen
they're really not really helpful at all. I mean, it’s just kind of an exercise— it’s
moreofanacademicexercise.
Taylor:
Yup.
Levi:
So,we'repracticalfocusedandlovesaying,“Okay.Well,howcanwehelpthisunit
be better?” And then kind of say, “Okay, well, here are the scenarios where
machinelearningcouldcomeinandhelp-whetherit'slengthofstay,likehewas
saying,orre-admissionsprediction,oreventhediseaseitself.
So, CLABSI’s a model that we found to be quite accurate. And it's interesting
becauseitdoesn'tbringinalotofthoselongertermvariables.CLABSIisacentral
lineinfection.Itdoesn'treallymattersomuchwhetheryouareofacertainweight
orhavehighbloodpressure.Sowhat’skindofthedistinctionthereisthat'smore
oflikein-hospital.
Taylor:
Yeah,soIthinkifyou'rekindoflookingatausecasewhereyou'retryingtopredict
aspecificconditionthenasyourconsultingwiththeclinicianorsomething,maybe
foryourCLABSIexample,-howlongthey'vehadthecentralline,howlongthey've
beeninthehospital.Somespecificimpactsaroundthatwhereachroniccondition
thatthey'veeitheralreadydevelopedandyou'retryingtofindoutsomesortof
outcomeafterwards.Yeah,you'regoingtoneedtolookintokindofthehistoryof
othervisits.Yeah,soIthinkthattheusecaseiskindoffollowingthatspectrumof
shorttermversusreallylongterm.
Levi:
Yeah,yeah.
Taylor:
It’simportanttoconsiderthat.
Levi:
Oh,totally.
Andyouseealltheconnectionpointsbetweenthemachine-learningfolkswho
aretheanalysts,becauseonlythehealthcareanalystcandothismodelcreation
themselves,andtheconnectionwithcliniciansbecauseIdon'tknoweverything
aboutCLABSIandCOPDandCAUTIandallthesedifferentthings.Butwhenyou're
talkingtoaclinicianorsomesortofsubjectmatterexpert,they'llbeabletosay
like,“Okay,wellherearethethingsthatmightdrivethisprediction.Andthenwith
healthcare.ai,itboilsitdownandsays,“Okay,fromthese40variables,hereare
the20thatactuallyshouldbeinthefinalmodel.”Soit'skindofabalancebetween
clinicalSMEEandalgorithm.
Taylor:
Yeah.AndI'dsaythatsomethingelsethatwe'refindingtobereallyimportant,as
youconsidersomeoftheseusecasesatyourhealthsystemorinyourroleorjust
things that you find interesting, is having some other sort of process or
interventionaroundit,sousingtheusingthepredictionstodrivesomesortof
improvement. So we're on a call the other day and someone kind of used the
phrase“notdoingmachinelearningjustforthesakeofmachinelearning”.Well,
there's definitely some interesting problems and that's kind of maybe where
somethinglikeKagglecomesinoryou'relearningit.That'sreallygreat.Butthen
in actual practice, using machine learning to drive some sort of outcome or
improvesomesortofcareiswhereyouendupwiththebestadoption.
Levi:
Yeah,exactly.
SogotoKaggletopractice.Don'tnecessarilypracticeyourmachinelearningon
yourhealthsystem.
Taylor:
Yup,youcandefinitelyfindsomethingusefultodo.Therearesomanyawesome
use cases out there. And a lot of people that are interested in having some
additionalsupportintheirdecisionmakingso.
Levi:
Exactly,exactly.
So we actually have a couple of comments that speak to some of those
discussions.
Taylor:
Oh,cool.
Levi:
Yeah,soDanWellish,andthisisatinybitofanonsequiturbutitsays,“Canyou
speaktoyourexperienceswithcleaningupdataintermsofremovingcolumns
whereimputingfornullsdoesn'tmakesense?”Thatkindofgoesbacktotheinitial
chatabout,perhaps,grains.Whenyou'rerollinguptoaparticulargrain,whatif
youdon'thaveenoughdatainthatcolumn?
Taylor:
Yeah.Yeah.Iwouldthinkthatyouwouldwanttoprofilethedataprettycarefully
upfront.Yeah,iftherearesomanynullsthatimputingisjustnottherightanswer
then,yeah,Ithinkyou'dwanttodropthatcolumn.
AndIthink,inalotoftheclinicaldata,ifmaybeacertainfieldisprettyraretobe
populatedthenmaybethat'snotagoodonetoconsider.Butyouwouldalsowant
tobecarefulnottotossoutafieldjustbecauseitlookslikeithasmostlynulls
becausethatcouldbeareallyimportantevent.Anditistrulyjustasrareasyou
thinkitisorasitlooksinthedata.Soyoudefinitelywouldwanttomaybeconsult
someone else that has more experience with the process that's actually
populatingthedatabeforeyouthrowitoutentirely,butitcouldbenegatively
impactingyourmodelforsure.
Levi:
Yeah,that'sagreatpoint.Andhealthcare.aicomeswithtoolsthathelpyousee,
“Okay,whatpercentageofmycolumnsornull?DoIneedtoactuallyfillinthese
null values with something that, you know, maybe the column mean or the
column’smostfrequentvalue”-thatsortofthing?Sowe'vetriedtokindofguide
youalongthatprocess,whatthetoolsareinhealthcare.ai.
Soacoupleofothercommentsherecomingthrough.Sowehadoneotherthat
askedusisJohn.So,actuallyasimilarquestion.So,alotofquestionsaboutdata
preparationaskingaboutthegranularityofdataanddealingwithmissingdata.So
keepthosecomingin.That'ssuperhelpful.
Andlet'sseehere.Oh,okay.Thomasasked,“Doyouknowagoodresourcefora
catalogoflongitudinaldataorpatient-leveldatainhealthcare?”Sothat'soneof
thehardpartsofmachine-learninginhealthcareisthatdataavailabilityisalittle
bit–well,it'sdifficult[inaudible00:24:13]anything.
Andwithhealthcare.ai,weactuallyputinathousand-rowlongitudinaldataset
thatisit'smadeupbutithasprettyrealisticvaluesinit.Andso,youcanplaywith
it,runtheexamplesonit,andkindofgetanideaasto,“Okay,well,ifwehave
multiplerowsperperson,howdothealgorithmsinteractwiththat?”Socheckout
thatbuilt-indatasetwithhealthcare.ai.
Taylor:
Yeah. And that would probably be a good use, like the more longitudinal data
[inaudible00:24:41]haveagooduseforthelinearmixmodelwith—
Levi:
Yeah,exactly.
Taylor:
We'vegotsomedataforthosepatientswherethey'vehadmultiplevisitsinyour
datasetbutit'sthesamepatient.Youmightwanttokindofaggregatethose.But
yeah,someofthatreallyspecific,realdataishardtocomebythough.
Levi:
Yeah,maybeweshoulddoanepisodeonthat.
Taylor:
Yeah.
Levi:
It’dbekindofnicetobrowsearoundandseewhat'soutthere.
Taylor:
Yup.
Levi:
Yeah. It's definitely something where Kaggle can help a little bit. So check out
Kaggleandseespecificallytheirdatasetpage.Butwe’llbecomingtoyoumore
withthatinthefuturebecausethat’sdefinitelyahottopic.
Taylor:
Yeah.That’sagoodthought,yup
Levi:
Awesome.
So,Thomasherehasacoupleofotherquestions.Andweappreciatey'allreaching
out.Feelfreetoaskmore.We'reeagertointeracthere.
Okay.Thomasaskedme,“Iplantoaddexampledatasetstotestthefunctionsin
yourRpackage?”Andthat'sagreatquestion.Andso,yeah,wedohavetheone
diabetesdataset.But,youknow,weshouldputinacouple,Iwouldsay,because
thesmallcsvfilesdon'ttakeupmuchspace.Andmaybewecoulddosomething
likeaCLABSIdatasetoraheartfailure.
Taylor:
Yup.
Levi:
Hmmm.Huh?
Taylor:
Yeah,maybetryandworkinsomeexampledatasetseventowardssomeofthe
usecasesthatImentionedearlier.Butyeah,soobviously,itsoundslikethere'sa
littlebitofaneedforsomelongitudinaldataoutthere.Wecouldseeifthere's
kindofawaytogetatthatalittlebittoanswerthatquestion.
Levi:
Yeah, yeah. That's an awesome point. So we'll follow up that and do maybe a
wholeepisodeondataouttherethatyoucanplaywith.Opendata,we’llcallit.
Taylor:
Yup.Andwe'remissingMiketoday.Iknowthathehaslookedatsomeinteresting
datasourcesaswellsowe'llcirclebackwithhim.
Levi:
Yeah.That'sdefinitelyagreatpoint.We'llpassitontoMike.
So, that's it. We went through the mailbag, talked about data and the news –
GoogleboughtKaggle,andtalkedalittleaboutmodelsthatyoucouldbuildinyour
healthcareenvironment.
Justrealquick,Taylor,whataresomeofthethingsthathealthcare,youknow,
healthsystemsaremostconcernedabout?Soifyouthinkabout,“Okay,wellwe
candomodelsonallthesedifferentthings.Whereshouldthehealthsystemstart?
Youknow,maybeCMS?Justkindofpushingtheminacertaindirection?”
Taylor:
Yup.Ifeellikere-admissionsisdefinitelyanareathathealthsystemsarefocusing
on.Andit'salsoatoughproblem.It'satoughproblemtopredict.There'salotof
variablesthatgointoitthatmaybearenotavailableinjusttheEHR.Sometimes,
the socio-economic data that's a little bit harder to get at is quite predictive.
Anyways,Ithinkthat'sabigareathatwe'reseeingdemandinthatisgoodtofocus
on.
And then things like sepsis or some of the hospital-acquired infections. That's
anotheroneofthoseareasthatimpactsreimbursementandnegativelyimpacts
patientqualityoflife.Andhospitalsarestartingtofocusonalloftheabovenow
which is exciting for patients. But also, challenging for data scientists, data
architects,dataanalysts.Sothere'salotofinterestingchallengesoutthere.
Levi:
Yeah.Andthere'salotofthingsaffectingreimbursements.AndisthisinanACO
senseyou'resayingorjusthospitalsingeneral?
Taylor:
Ah, even in the hospital. Yeah, it can affect whether— you know, if there was
somethingthatyouacquiredwhileyou'retherethatcanaffectaninsuranceor
especiallysomethinglikeaMedicarereimbursement.
Levi:
Interesting.
Taylor:
Andso,that'sanareawhereifitimpactsrevenueandpatientqualityoflife,the
hospitalsystemisgoingtostarttolookintoitorshould.
Levi:
Yeah,that'sagreatplacetostart.That'sawesome.
Now,thankssomuchforjoiningus.Dowehaveanythingelseweneedtoaddress?
Taylor:
No.
Levi:
[inaudible00:28:26]wekindofranthroughourpointsthere?
Taylor:
Yeah,yeah.Ithinkwekeptthingsontracktoday.So,ifthere’sanyotherthingsin
thechatwindow,wecouldmonitorthatforashortbit.Butotherwise,Ithinkwe
gotthroughmostofthequestionshere.
Levi:
Yeah.We'lltakesomefornextweek.Andbesuretokeepsendingthoseinthrough
email.Welovetohearwhatyou'reupto.
Thanksforjoiningus,Taylor.
Taylor:
Iappreciateyouincludingmetoday,sothanksfortheopportunity.
Levi:
Oh,forsure.Thankyou.
Taylor:
Allright.
Levi:
We’llseeyouallnextweek.Thanks,everyone.
Taylor:
Allright,thankyou.