medical decision sciencesinformatics.snubi.org/2004fall/medicaldecisionsciences.pdf ·...
TRANSCRIPT
1
Medical Decision Sciences Medical Decision Sciences
Ju Han Kim, M.D., Ph.D., M.S.Ju Han Kim, M.D., Ph.D., M.S.SNUBI: SNUBI: SNUBiomedicalSNUBiomedical InformaticsInformatics
Medical Decision SciencesMedical Decision Sciences
•• Medical decision makingMedical decision making•• Data, Information, & KnowledgeData, Information, & Knowledge•• Knowledge representationKnowledge representation•• BayesBayes rulerule•• ROC curveROC curve•• Clinical decision analysisClinical decision analysis•• Machine learningMachine learning•• Clinical decision support systemClinical decision support system•• Grade CGrade C
Clinical practice as decision makingClinical practice as decision making
•• The The ability to make good decisionsability to make good decisions is the is the hallmark of the good performing hallmark of the good performing professionalprofessional..
•• Information systemInformation system can be used for good can be used for good decision makingdecision making and and decision supportdecision support..
•• Therefore, it is important to understand the Therefore, it is important to understand the theory of decision makingtheory of decision making..
A working definition of a decisionA working definition of a decision
An irrevocable allocation of resourcesAn irrevocable allocation of resources
•• Resources can be time, money, attention Resources can be time, money, attention cycles, cycles, emotional energy, or anything else emotional energy, or anything else of of valuevalue to the decision maker.to the decision maker.
•• IrrevocableIrrevocable allocation is required, because allocation is required, because otherwise there is no crisp way do define otherwise there is no crisp way do define when a decision should be made.when a decision should be made.
Decision is different than worryingDecision is different than worrying
Worrying typically does not involve Worrying typically does not involve allocation of resources.allocation of resources.
Medical Decision SciencesMedical Decision Sciences
•• Medical decision makingMedical decision making•• Data, Information, & KnowledgeData, Information, & Knowledge•• Knowledge representationKnowledge representation•• BayesBayes rulerule•• ROC curveROC curve•• Clinical decision analysisClinical decision analysis•• Machine learningMachine learning•• Clinical decision support systemClinical decision support system•• Grade CGrade C
2
Data, Information, & Knowledge Data, Information, & Knowledge
•• DataData
•• InformationInformation
•• KnowledgeKnowledge
Informatics: Informatics: knowledge about knowledge
1.1. Combining several simple ideasCombining several simple ideas into one compound one, into one compound one, and the all and the all complex ideascomplex ideas are made. are made.
2.2. The second is The second is bringing two ideasbringing two ideas, whether simple or , whether simple or complex, together, and setting them by one another so as complex, together, and setting them by one another so as to take a view of them at once, without uniting them into to take a view of them at once, without uniting them into one, by which it gets all its ideas of one, by which it gets all its ideas of relationsrelations. .
3.3. The third is The third is separating them from all other ideasseparating them from all other ideas that that accompany them in their real existence: this is called accompany them in their real existence: this is called abstractionabstraction, and thus all its , and thus all its general ideasgeneral ideas are made.are made.
John Locke, John Locke, An Essay Concerning Human UnderstandingAn Essay Concerning Human Understanding (1690)(1690)
The acts of the mind wherein it exerts its power over The acts of the mind wherein it exerts its power over simple ideassimple ideas, are chiefly these three: , are chiefly these three:
Medical Decision SciencesMedical Decision Sciences
•• Medical decision makingMedical decision making•• Data, Information, & KnowledgeData, Information, & Knowledge•• Knowledge representationKnowledge representation•• BayesBayes rulerule•• ROC curveROC curve•• Clinical decision analysisClinical decision analysis•• Machine learningMachine learning•• Clinical decision support systemClinical decision support system•• Grade CGrade C
•• Binding Binding
•• CombiningCombining
•• RelatingRelating
•• AbstractingAbstracting
Data, Data Structure, Abstraction, RepresentationData, Data Structure, Abstraction, Representation
FROG
2
BIRD
HUNTER BIRDMAN
MAN BIRDHUNTS
Grumpy Happylikes
Manager
Is_a
Grumpy Happylikes
Manager
Is_a
Managers
Grumpy
Is_a
likes
Happy
Happy
Data, Data Structure, Abstraction, Data, Data Structure, Abstraction, Knowledge Representation, & FormalismKnowledge Representation, & Formalism
Frames: an exampleFrames: an exampleNAME : Acute glomerulonephritisTriggered by
Confirmed by
Caused byCauses
Complications
Differential diagnosis
facial edema, not painful, not erythematical, symmetrical, etc.malaise, asthenia, anorexia, etc.
recent streptococci infectionsodium retention, acute hypertension, nephroticsyndrome, etc.acute kidney failure
(If chronic high blood pressure then chronicglomerulonephritis)(If recurrent edema then nephrotic syndrome)
3
Semantic Network & Medical OntologySemantic Network & Medical Ontology
생물학적 기능
병리학적 기능
질병 또는 증상
is_a
is_a
is_ais_a
결핵 저칼륨혈증
Semantic Network : UMLSSemantic Network : UMLSOrganism
Finding
Laboratoryor test result
Biologicfunction
Bodysystem
Injury orpoisoning
Anatomical structure
Organismattribute
Sign orsymptoms
Embryonicstructure
Fully formedanatomicalstructure
Organismfunction
Cell or molecular
dysfunction
Molecularfunction
Cellfunction
Organ or tissue
function
Mentaldysfunction
Experimentalmodel ofdisease
Disease orsyndrome
Physiologicfunction
Pathologicfunction
Plant Virus Animal
Vertebrate
Mammal
Human
Mentaldysfunction
Bodypart
part of
disrupts
evaluation of
prop of
disrupts
conceptual part of
process of
evaluation of
conceptual part of
Medical Language & Classification SystemsMedical Language & Classification Systems•• ICDICD•• MeSHMeSH•• CPT 4 CPT 4 •• SNOMEDSNOMED•• HLHL--77•• LOINCLOINC•• Arden SyntaxArden Syntax•• UMLSUMLS
MetathesaurusMetathesaurusSemantic NetworkSemantic NetworkInformation Source MapInformation Source Map
•• Controlled Vocabularies and Vocabulary ServerControlled Vocabularies and Vocabulary ServerPTXT / COSTAR / TMR / RMRSPTXT / COSTAR / TMR / RMRSMED (Medical Entities Dictionary)MED (Medical Entities Dictionary)PEN & PADPEN & PADGALENGALEN
Medical Decision SciencesMedical Decision Sciences
•• Medical decision makingMedical decision making•• Data, Information, & KnowledgeData, Information, & Knowledge•• Knowledge representationKnowledge representation•• BayesBayes rulerule•• ROC curveROC curve•• Clinical decision analysisClinical decision analysis•• Machine learningMachine learning•• Clinical decision support systemClinical decision support system•• Grade CGrade C
ProbabilityProbability
)()|()()|()()|(
)()()|()|(
)()|()()|()(/)()|(
)()()()(
1)(1)(0
DpDSpDpDSpDpDSp
SpDpDSpSDp
ApABpBpBApBpBApBAp
BApBpApBApApAp
i
⋅+⋅⋅
=⋅
=
⋅=⋅∴∧=
∧−+=∨
=
≤≤
∑AA BB
Bayes Bayes Rule Rule
)|()|(
)()(
)|()|(
)()(
)(1)(
DSpDSp
DpDp
SDpSDp
ApAp
ApApOdds
•=
=−
=
•• oddsodds--likelihood ration form of likelihood ration form of Bayes Bayes formulaformula
ProbProb = 1 / 3= 1 / 3Odds = 1 / 2 =0.5 Odds = 1 / 2 =0.5
AA
Post. Odds = prior odds x likelihood ratioPost. Odds = prior odds x likelihood ratio
4
Diagnostic test characteristics Diagnostic test characteristics
D +D + D D ––T + aT + a bb a+ba+bT T -- cc dd c+dc+d
a+ca+c b+db+d a+b+c+da+b+c+d
•• Sensitivity = a / (a+c) = p(T+|D+)Sensitivity = a / (a+c) = p(T+|D+)•• Specificity = d / (b+d) = p(TSpecificity = d / (b+d) = p(T--|D|D--))•• Positive Predictive Value = a / (a+b) = p(D+|T+)Positive Predictive Value = a / (a+b) = p(D+|T+)•• Negative Predictive Value = d / (c+d) = p(DNegative Predictive Value = d / (c+d) = p(D--|T|T--))
Diagnostic test characteristics Diagnostic test characteristics
예제예제)) Sensitivity=99.99%, specificity=99.9% Sensitivity=99.99%, specificity=99.9% 인인 최신의최신의
에이즈에이즈 검사가검사가 개발되었다개발되었다. . 철수는철수는 이이 검사에검사에 양성반응을양성반응을
보였다보였다. . 철수가철수가 에이즈에에이즈에 감염됬을감염됬을 확률은확률은 얼마인가얼마인가? ? ((현재현재 한국인의한국인의 에이즈에이즈 유병율은유병율은 0.00010.0001이라고이라고 한다한다.).)
1.1. 99%99%2.2. 95%95%3.3. 80%80%4.4. 50%50%5.5. 10%10%
Diagnostic test characteristics Diagnostic test characteristics
AIDS AIDS no AIDSno AIDST +T +T T --
9,9999,99911
10,00010,000
•• Sensitivity=99.99%Sensitivity=99.99%•• Specificity=99.9%Specificity=99.9%
11999 999
10001000
Prevalence=0.0001Prevalence=0.0001
00,00000,00000,000 00,000 00,00000,000
109,999109,99999,900,00199,900,001
100,010,000100,010,000
•• Positive predictive value = 9,999 / 109,999 < 10%Positive predictive value = 9,999 / 109,999 < 10%•• Negative predictive value = 1.0Negative predictive value = 1.0
Diagnostic test characteristics Diagnostic test characteristics
T+ (TP)T+ (TP) TT-- (FN)(FN)
T+ (FP)T+ (FP) TT-- (TN)(TN)
T+ (TP)T+ (TP) TT-- (FN)(FN)
T+ (FP)T+ (FP) TT-- (TN)(TN)
Medical Decision SciencesMedical Decision Sciences
•• Medical decision makingMedical decision making•• Data, Information, & KnowledgeData, Information, & Knowledge•• Knowledge representationKnowledge representation•• BayesBayes rulerule•• ROC curveROC curve•• Clinical decision analysisClinical decision analysis•• Machine learningMachine learning•• Clinical decision support systemClinical decision support system•• Grade CGrade C
ROC CurveROC Curve
5
Medical Decision SciencesMedical Decision Sciences
•• Medical decision makingMedical decision making•• Data, Information, & KnowledgeData, Information, & Knowledge•• Knowledge representationKnowledge representation•• BayesBayes rulerule•• ROC curveROC curve•• Clinical decision analysisClinical decision analysis•• Machine learningMachine learning•• Clinical decision support systemClinical decision support system•• Grade CGrade C
Why decision making is so hard?Why decision making is so hard?
UNCERTAINTY!!UNCERTAINTY!!
•• If you have all information about the If you have all information about the probabilitiesprobabilities of different events, then it of different events, then it typically is NOT a hard decisiontypically is NOT a hard decision
•• If there is no uncertainty, and the decision If there is no uncertainty, and the decision is still hard, then the problem is that the is still hard, then the problem is that the utility functionutility function is not clearis not clear
Decision TheoryDecision TheoryA theory for decision making that is A theory for decision making that is rationalrational..
Expected value decision makingExpected value decision makingMaximize expected valueMaximize expected value among the among the decision alternativesdecision alternatives
Clinical Decision AnalysisClinical Decision Analysis
Chance node
Decision node
Transition probability
Active transition
Alternating chance & decision nodesAlternating chance & decision nodes
사망
Simple decision treeSimple decision tree
즉각적족부절단
관찰
생존
사망
반흔형성
병변확장
회복
사망
무릎위
절단
Result
사망
무릎아래절단
무릎위절단
생존
A trauma caseA trauma case
Simple decision treeSimple decision tree
즉각적족부절단
관찰
생존 (0.99)
사망 (0.01)
반흔형성 (0.7)
병변확장 (0.3)
회복
사망
무릎위
절단
Result
사망
무릎아래절단
무릎위절단사망 (0.1)
생존 (0.9)
PROBABILITY: PROBABILITY: objectiveobjective
6
Simple decision treeSimple decision tree
즉각적족부절단
관찰
생존 (0.99)
사망 (0.01)
반흔형성 (0.7)
병변확장 (0.3)
0.00
1.00 회복
0.95
사망
무릎위
절단
0.98
0.00
Result
사망
무릎아래절단
Utility
무릎위절단사망 (0.1)
생존 (0.9)
UTILITY: UTILITY: subjectivesubjective
Simple decision treeSimple decision tree
즉각적족부절단
관찰
생존 (0.99)
사망 (0.01)
반흔형성 (0.7)
병변확장 (0.3)
0.00
1.00 회복
0.95
사망
무릎위
절단
0.98
0.00
Result
사망
무릎아래절단
Utility
무릎위절단사망 (0.1)
생존 (0.9)
EXPECTED VALUEEXPECTED VALUE
0.970
0.956
0.855
0.98 X 0.99 = 0.970
0.95 X 0.9 = 0.855
0.7 X 1.0 + 0.855 X 0.3 = 0.956
Sensitivity analysisSensitivity analysis
즉각적족부절단
관찰
생존 (0.99)
사망 (0.01)
반흔형성 (0.7)
병변확장 (0.3)
1.00
무릎위절단사망 (0.1)
생존 (0.9)
0.970
0.956
0.855
.5
.6
.7
.8
.91.0
0.8
반흔형성
What if you donWhat if you don’’t know the numbers?t know the numbers? Standard gambleStandard gamble
Utility is not PROBABILITYUtility is not PROBABILITY
Immediate deathImmediate deathon surgery on surgery
Perfect visionPerfect vision
blindnessblindness pp
11--pp
Getting utilityGetting utility
Risk seeking vs. risk averseRisk seeking vs. risk averse Embedded Markov chainEmbedded Markov chain
7
Medical Decision SciencesMedical Decision Sciences
•• Medical decision makingMedical decision making•• Data, Information, & KnowledgeData, Information, & Knowledge•• Knowledge representationKnowledge representation•• BayesBayes rulerule•• ROC curveROC curve•• Clinical decision analysisClinical decision analysis•• Machine learning: Artificial Intelligence in MedicineMachine learning: Artificial Intelligence in Medicine•• Clinical decision support systemClinical decision support system•• Grade CGrade C
Machine Learning Approach in Bioinformatics
Classifications
Non-exclusive Exclusive
Supervised Unsupervised
Hierarchical Partitional
Artificial Intelligence in MedicineArtificial Intelligence in Medicine
•• RuleRule--based system: MYCINbased system: MYCIN•• Classification treeClassification tree•• Bayes Bayes netnet•• Artificial neural netArtificial neural net•• Rough setRough set•• Reinforcement learningReinforcement learning
RuleRule--based systembased systemMYCINMYCIN
>> What is the patients name? John Doe.
>> Male or Female? Male.
>> Age? 55.
>> Let's call the most recent positive culture C1. From what site was C1 taken?
From the blood. ... >> My recommendation is as follows: give gentamycinusing a dose of 119 mg (1.7 mg/kg) q8h IV [or IM] for 10 days. Modify dose in renal failure. Also, giveclindamycin using a dose of 595 mg (8.5 mg/kg) q6h IV [or IM] for 14 days.
MYCINMYCIN
IF: 미생물이 그림 양성으로 염색된다.
미생물이 형태학적으로 막대균이다.
환자가 감염되었을 위험이 높다.
THEN: 감염체는 슈도모나스로 추정된다. (확신도 = 0.6)
RuleRule--based systembased system
•• RulesRulesA A →→ BBB and C B and C →→ DDD and E D and E →→ GG
•• Backward chainingBackward chaining
•• Forward chainingForward chaining
A B
D
G
E
C
MYCINMYCIN
RuleRule--based systembased system
8
Classification TreeClassification TreeSunburn at the BeachSunburn at the Beach
Classification TreeClassification Tree
Classification TreeClassification Tree Classification TreeClassification Tree
Classification TreeClassification Tree Classification TreeClassification Tree
9
Bayesian belief netBayesian belief net Artificial Neural NetworkArtificial Neural NetworkA Universal Function A Universal Function ApproximatorApproximator
Reinforcement LearningReinforcement Learning
Markov Chains and Hidden Markov Model Markov Chains and Hidden Markov Model Medical Decision SciencesMedical Decision Sciences
•• Medical decision makingMedical decision making•• Data, Information, & KnowledgeData, Information, & Knowledge•• Knowledge representationKnowledge representation•• BayesBayes rulerule•• ROC curveROC curve•• Clinical decision analysisClinical decision analysis•• Machine learningMachine learning•• Clinical decision support systemClinical decision support system•• Grade CGrade C
“Our integrated approach to medicine skillfully combines an array of holistic alternative treatments with a sophisticated computerized billing services”
Clinical Information System
• OCS - LIS - RIS - PACS - EMR
• Departmental Information Systems
• Library / Factual Information Systems
• Research Database
Hospital Information System is Dead!
Intelligent Integration of
10
Decision support systemDecision support system
•• Passive systemPassive systemConsultant systemConsultant systemCritiquing systemCritiquing system
•• SemiSemi--active system active system ReminderReminderAlertAlert
•• Active systemActive system
Your answeris...
X >1
Sex=F
UTI314:
Y > 1
Hello!...
k>1Go 314
Go 1200
Go 354
Hx of hos..
adolescentSub R34
freqency
urgency pain
Age > 30
For your..
Mx. We are...
Allergy?
Hx.?
Admin?
Help?
UTI:RTS Protocol 3110
Go 399Sub R34
Sub R44
Sub R55
Go 1200
Flow chart systemFlow chart system
Flow chart systemFlow chart system•• simple, easy to buildsimple, easy to build•• hard to manage / maintainhard to manage / maintain
Diagnostic strategy
Decision support systemDecision support system
Knowledge baseFactual
Information
Inference Engine
•• flowchartflowchart•• rulerule--based systembased system•• frameframe--based systembased system•• sequential sequential BayesBayes•• modelmodel--based system based system
11
Decision support componentsDecision support components
•• GLIF (Guide Line Interchange Format)GLIF (Guide Line Interchange Format)•• Arden syntaxArden syntax•• MLM (Medical Logic Module)MLM (Medical Logic Module)•• OneOne--rule expert systemrule expert system
GLIF : GLIF : GuideLineGuideLine Interchange FormatInterchange Format
Data modelData model
GLIF : GLIF : GuideLineGuideLine Interchange FormatInterchange Formatimplementationimplementation Medical Decision SciencesMedical Decision Sciences
•• Medical decision makingMedical decision making•• Data, Information, & KnowledgeData, Information, & Knowledge•• Knowledge representationKnowledge representation•• BayesBayes rulerule•• ROC curveROC curve•• Clinical decision analysisClinical decision analysis•• Machine learningMachine learning•• Clinical decision support systemClinical decision support system•• Grade CGrade C
Theory and PracticeTheory and Practice
In theory, theory and practice In theory, theory and practice should be the same. should be the same.
But in practice, theory and practice But in practice, theory and practice are different.are different.
Grade CGrade C
NEJM 1994NEJM 1994