ACDC: Alpha-Carving Decision Chain for Risk Stratification
-
Upload
yubin-park -
Category
Engineering
-
view
252 -
download
0
Transcript of ACDC: Alpha-Carving Decision Chain for Risk Stratification
ACDC:Alpha-CarvingDecisionChainforRiskStratification
YubinPark,AccordionHealth,Inc.JoyceHo,EmoryUniversity
JoydeepGhosh,TheUniversityofTexasatAustin
1ICMLWHI2016
WhatisDecisionChain(DC)?
• AlsoknownasRuleLists(Wang&Rudin,2015)• Asequenceofrules,appliedtooneafteranother,wheretheratioofpositiveclassincreasesoverthesequenceofrules• ToyExample:AdecisionchainforpredictingthelikelihoodofbeingaLonghornfan,• IfTomlivesinAustin,TXà 25% chanceofbeingaLonghornfan• AndTomlikestowatchfootballgamesà 50% chance• AndTomgoesoutforatailgateoneverySaturdayà 75% chance• AndTomwearsburntoranget-shirtsallthetimeà 95% chance
2ICMLWHI2016
Conceptually,somethinglikethis
3
+ +� �
+
�
�
�
· · ·
✓A ✓B
R1
S1S1
S1
S0
S0
S0
S2
S2
S2
SD�1
RD�1
RD
R2
+ +� �
+
�
�
�
· · ·
✓A ✓B
R1
S1S1
S1
S0
S0
S0
S2
S2
S2
SD�1
RD�1
RD
R2
DecisionTree DecisionChain
ICMLWHI2016
IsDCMoreInterpretablethanDT?
• InDecisionChain(DC),• Riskisproportionaltothenumberofrules• Lesstomemorizeforfilteringoutlow-riskpopulation(orsamples)• Moretomemorizeforcapturinghigh-riskpopulation• UsingDC,onecanimplementaneconomicallyefficientbusinessprocessbasedonjobmaturity-level
• WhileinDecisionTree(DT),• Thenumberofrulesisagnostictorisk• Low-riskcanbecapturedwithoneruleaswellashundredsofrules
• Thus,DCmaybehelpfulforsomeapplications
4ICMLWHI2016
InHealthcareApplications,
• Classimbalanceproblemsareprevalent• Majorityclassexamplescanbeoftencarvedout(orfilteredout)withasimpleconjunctionofif-elsestatements• Animplementationstrategy• Filteroutmajorityclasswithrulesà ObtainalessimbalanceddatasetàApplyafancymachinelearningalgorithm• Onequestion: HowmanymajorityexamplesshouldIfilterout?• Apossiblesolution: Ifwebuildadecisionchain,thenwecanstreamlinegrid-searchmucheasily
• DCcanbemoreinterpretableaswellasmoreefficient(sometimes)
5ICMLWHI2016
QuestionisHow
• Wewilluseagreedyapproach• Notethatdecisiontreeisalsoagreedyalgorithm• Pickasplittingfeaturethatmaximizes{informationgain,purityscore,etc.}• Splitthedatasetintopartsbasedonthevalueofthesplittingfeature• Repeatfromthebeginningforeachdataset
• Wewillgrowadecisionchainasfollows• Pickasplittingfeaturethatcarvesoutthemostamountofmajoritysamples• Splitthedatasetintopartsbasedonthevalueofthesplittingfeature• Repeatfromthebeginningononlyonepartitionthathasmorepositiveclassexamples
6ICMLWHI2016
MoreDetailsonHow
• Selectingthebestsplittingfeature• WewilluseAlpha-Divergence• Alpha-DivergenceisthesameasKL-DivergencewhenAlpha=1• Alpha-DivergenceisthesameasHellingerdistancewhenAlpha=0.5• Alpha-DivergencecanbealotofdifferentthingsbasedonthevalueofAlpha
• WewillchangethevalueofAlphaadaptively(withasimplestrategy)toachieveourgoal• Moredetailsareinthepaper
7ICMLWHI2016
EffectofDifferentAlphas
• HighAlpha• Purepartitions
• LowAlpha• Balancedpartitions
8
α = 1α = 16
α = 48α = 64
0
25
50
75
50 100 150nbp.systolic
count
Shock F T
ICMLWHI2016
Experiments:SepticShock
• Alpha-CarvingDecisionChain(ACDC)showscomparableperformancewithotherdecisiontreealgorithms• ATree(a=1):C4.5• ATree(a=2):CART• ATree(a=x):otheralpha-trees
9
●
●
●
●
●
0.5
0.6
0.7
0.8
ATree(a=1)
ATree(a=2)
ATree(a=4)
ATree(a=16)
ATree(a=64)
ATree(a=128)
ACDC
ModelAU
C
ICMLWHI2016
Experiments:SepticShock
• SinceACDCisadecisionchain,wecanmakethiscoolvisualization• Putdecisionrulesandperformancemetricsinasinglechart• Riskisproportionaltothenumberofrulesapplied
10
●
●
●
●
●
L1: nbp.systolic<=132
L2: nbp.systolic<=98.4
L3: min.nbp.systolic<=90.4
L4: nbp.diastolic<=46
0
1
2
3
0.00 0.25 0.50 0.75 1.00Coverage
Lift
ICMLWHI2016
Experiments:CardiacArrest
• Anotherexample:Asystole• Again,comparabletootherdecisiontrees
11
●
●
●●
●
●
●
0.5
0.6
0.7
0.8
ATree(a=1)
ATree(a=2)
ATree(a=4)
ATree(a=16)
ATree(a=64)
ATree(a=128)
ACDC
ModelAU
C
ICMLWHI2016
Experiments:CardiacArrest
12
●●
●
●
●
●
●
●
L1: min.nbp.diastolic<=48.767L2: min.nbp.systolic<=104
L3: min.spo2<=90
L4: spo2<=93.6
L5: min.spo2<=78L6: avg.pp>57.495
L7: hr<=90.9
0.0
2.5
5.0
7.5
10.0
0.00 0.25 0.50 0.75 1.00Coverage
Lift
●
●
●
●
●
●
●
●
●
L1: min.nbp.diastolic<=48.767L2: min.nbp.systolic<=104
L3: min.spo2<=90
L4: spo2<=93.6
L5: min.spo2<=78L6: avg.pp>57.495
L7: hr<=90.9
0.00
0.25
0.50
0.75
1.00
0.00 0.25 0.50 0.75 1.00FPR
TPR
ICMLWHI2016
Experiments:CardiacArrest
• Youalsocanmakeariskpyramid
13
min.diastolic < 48 mmHg
min.systolic < 104 mmHg
min.spo2 < 90 %
spo2 < 93 %
Baseline Risk
1.3 times higher
1.5 times higher
3.7 x
5.8 xAsystole !Risk Stratification !
Decision Chain
Higher risk
ICMLWHI2016
Contacts
• YubinPark• yubin[at]accordionhealth [dot]com
• JoyceHo• joyce [dot]c[dot]ho[at]emory [dot]edu
• JoydeepGhosh• jghosh [at]utexas [dot]edu
14ICMLWHI2016