1 Constructing Fuzzy Signature Based on Medical Data Student: Bai Qifeng Client: Prof. Tom Gedeon.

1

Constructing Fuzzy Signature Based on Medical Data

Student: Bai Qifeng

Client: Prof. Tom Gedeon

2

Proposal

Explore an approach to automatic construct Fuzzy signature based on medical database

It contains three questions:1. How to identify SARS suspect patients

group?2. How to explore the relationships among

symptoms?3. How to construct fuzzy signature based on

above analysis?

3

Fuzzy Logic Theory

Fuzzy logic uses linguistic rules which reflect uncertainty or vagueness of concepts in natural in natural language.

If 50m/h is the boundary of “slow” and “fast” , Conventional bivalent sets regards 50.1m/h as fast.

What if current speed is 49.9m/h?

In real world, it should be a smooth shift.

4

Fuzzy Set

Now, assume there are three temperatures

We can get the fuzzy sets:

A fuzzy set is a set whose elements have degrees of membership.

78.0,22.0,0,08.39 A

0,0,0,83.08.37 A

0,0,71.0,29.04.38 A

8.39,4.38,8.37U

38.4 39.837.8

37.3 37.9 38.6 39.1 400

0.2

0.4

0.6

0.8

1

Slight Moderate Sever

eExtreme

5

Why use fuzzy sets

Assume:1. IF Fever = Slight THEN

dose = Low.

2. IF Fever = Moderate THEN dose = Ave.

Fuzzy value of fever is slight = 0.29 and moderate = 0.71

Value of dose will share properties of both Low and Ave range.

IN

OUT

0

0. 2

0. 4

0. 6

0. 8

1

1. 2

37 38 39 40 41

Low Ave

38. 4

37. 9 38. 6 39. 1 40

0

0. 2

0. 4

0. 6

0. 8

1

1. 2

37 38 39 40 41

0.71

0.29

6

Problem Definition

A Major issue in fuzzy applications is how to create fuzzy rules

the number of rules have an exponential increase with the number of inputs and terms.

At least one activated rule for every input.

e.g. 5 terms, 2 inputs => 25 rules

5 terms, 5 inputs => 3,125 rules

)(|| kTOR

7

Sketch of Solution

Three possible solutions Decrease T :

Sparse Fuzzy System

Decrease K:

Hierarchical Fuzzy System

Decrease both simultaneously :

Sparse Hierarchical Fuzzy Rule Bases

8

Hierarchical Fuzzy Systems

Hierarchical fuzzy systems reduce to the dimension of the sub-rule bases k by using meta – levels

9

Fuzzy Signatures

Fuzzy signatures structure data into vectors of fuzzy values, each of which can be a further vector.

Each signature corresponds to a nested vector structures or, equivalently, to a tree graph.

10

Fuzzy Signatures

The relationship between higher and lower levels is govern by fuzzy aggregations.

Fuzzy aggregation contains union, average, intersection etc.Examples: Union: AUB = max [A, B] = A or B Intersection: A∩B = min [A,B] = A and B

11

Clustering

The aim of cluster analysis is to classify objects based on similarities among them.

Definition of cluster is a group of objects that are more similar to one another than to members of other clusters.

Clustering is unsupervised classification: no predefined classes

12

Clustering: Similarity

How to evaluate the similarities of data?

Cluster analysis adapts the distance between two points as the criterion of similarity.

Distance-type measure has Euclidean distance and City block distance.

13

Clustering: hard/fuzzy methods

14

Clustering: Procedure

15

Clustering: Fuzzy C-Means

Bezdek define objective function as :

represents the deviation of data with centre. The number m governs the influence of membership grades. uij represents the degree of membership of the data point xj belonging to v .

c

i

n

jij

mijm mvxXVUJ

1 1

21,;,

2ik vx

16

Clustering: Cluster Valid Index

Xie and Beni Index

The numerator calculates the compactness of data in the same cluster and the denominator computes the separateness of data in different clusters.

Smaller value of numerator validity index indicates that the clusters are more compact and larger values of denominator denotes the clusters are well separated.

min

),,( 1 1

2

jivvn

xvu

cVUVji

c

i

n

jji

mij

xie

17

Factor Analysis

Factor analyses are performed by examining the pattern of correlations between the observed measures.

. X is a vector of variables, where

is a vector of r<p latent variables called factors, is a (p*r) matrix of coefficients (loadings),

is a vector of random errors.

X

Tr ,,, 21

Tp ),,,( 21

18

Factor Analysis

19

Factor Analysis: Principal component analysis

Principal component analysis aims to reduce the dimension of variables and these new variables can interpret most of cases.

20

Factor Analysis: Principal component analysis

. x is the p dimensional variables, where U is an orthogonal matrix. 1. The loading of matrix U and vector Z( ) ,

which correspond to the variance and vector of the principal components respectively.

2. The value represents the contribution ratio which indicates how much percentage the principal component represents of the total tendency of the variables.

3. Usually, an accumulative contribution ratio of 70 - 80 percent can effectively represent the major variations in the original data.

xUy '

jjj yz

21

Factor Analysis: PCA vs FA

Direction is reversed: the measured responses are based on the underlying factors while in PCA the principal components are based on the measured responses

22

Factor Analysis: Factor Rotation

For identify some variables having similar factor loading, we could rotate the factor coordinates in any direction without changing the relative locations of the points to each other.

23

Experiment: Scatter of Raw Data

kinaseaspartatedehydrogena

selymphopeniachestdyspneacoughmalaisefever

feve

rm

alai

seco

ugh

dysp

nea

ches

tly

mph

open

iade

hydr

ogen

ase

aspa

rtat

eki

nase

Gravities of components are deviated by the noise or outliers.

24

Experiment: Scatter After Clustering

Collected data can represent the pattern of the disease more accurately.

kinaseaspartatedehydrogena

selymphopeniachestdyspneacoughmalaisefever

feve

rm

alai

seco

ugh

dysp

nea

ches

tly

mph

open

iade

hydr

ogen

ase

aspa

rtat

eki

nase

25

Experiment: KMO and Bartlett’s Test

KMO test indicates the possibility of containing underlying factors. KMO < .50, factor analysis is not useful.

Bartlett's test indicate whether variables are unrelated. significance level < .05 significant relationships

Kaiser-Meyer-Olkin Measure of Sampling Adequacy.

.608

Bartlett's Test of Sphericity

Approx. Chi-Square 191.238

df 36

Sig. .000

26

Experiment: PCA Model

Initial Eigenvalues Sums of Squared Loadings

Total % of Variance Cumulative % Total % Variance Cumulative %

1 2.581 28.679 28.679 2.581 28.679 28.679

2 2.024 22.485 51.163 2.024 22.485 51.163

3 1.096 12.179 63.342 1.096 12.179 63.342

4 .902 10.022 73.365

5 .825 9.168 82.533

6 .644 7.153 89.686

7 .465 5.171 94.858

8 .249 2.770 97.627

9 .214 2.373 100.0

Accumulative contribution ratio = 63%

27

Experiment: PCA Model

It denotes that variables could be divided into 3 factors

987654321

Component Number

3.0

2.5

2.0

1.5

1.0

0.5

0.0

Eig

enva

lue

28

Experiment: Results

Factors

1 2 3

fever -.079 -.121 .600

malaise -.474 .487 -.035

cough -.054 -.164 .653

dyspnea .289 .868 -.037

chest .120 .290 .613

lymphopenia -.252 .813 -.095

dehydrogenase .857 .146 -.127

aspartate .861 -.014 -.014

kinase -.649 .121 -.035

29

Experiment: Results after rotation

Factors

1 2 3

fever -.082 -.275 .547

malaise -.551 .398 -.039

cough -.053 -.324 .590

dyspnea .132 .886 -.190

chest .040 .148 .672

lymphopenia -.388 .760 -.075

dehydrogenase .823 .309 -.016

aspartate .849 -.130 -.056

kinase -.658 .019 -.061

30

Experiment: Constructed fuzzy signature

Hierarchical clustering or K-means can be used to cluster each factor

Weighted aggregation method in this fuzzy signature had higher performance 3 weights & 3

aggregations

malaise

kinaseasedehydrogen

aspartate

alymphopeni

dyspnea

chest

cough

fever

31

Experiment: Possible rule bases

Aggregations: Min (fever, cough, chest) Min (dyspnes, lymphopenia) Max (Min (kinase, malaise), Min(aspartate, de

hydrogenase) )

Rules If a patient has fever, cough and chest. If a patient has dyspnes and lymphopenia. If patient has kinase and malaise or has aspart

ate and dehydrogenase

32

Experiment: Possible rule bases

Further assumption: If a patient has fever, cough and chest, he/sh

e would has 64% possibility to get SARS If he/she has kinase and malaise or has aspa

rtate and dehydrogenase simultaneously, the possibility is increasing to 93%

If he/she has dyspnes and lymphopenia, he/she can be diagnosed as a SARS Patient

33

Conclusion

Advantages:1. Fuzzy signatures are capable of improving

the applicability of fuzzy systems.

2. Fuzzy signatures have the ability to cope with complex structured data and interdependent features problems.

3. With weighted aggregated, fuzzy signatures can assist experts to make decision by removing redundant information

34

Further Work

Further research can be focused on evaluating underlying relationships between the structures of fuzzy signatures, aggregation functions and weights of each vector.

35

Thank you

---- Bai Qifeng

36

Appendix

Demo of Fuzzy Control Sparse Fuzzy System Automatic Constructing Fuzzy Signature Fuzzy c-Means

37

Fuzzy Control

Fuzzy control is the most important current application in fuzzy theory.

Usually, three steps in Fuzzy control:

1. Fuzzification

2. Rule evaluation

3. Defuzzification

38

Demo of Fuzzy Control

The most common one is the centre of gravity

39

Demo of Fuzzy Control

Use a procedure originated by Ebrahim Mamdani as demo. The application is to balance a pole on a mobile platform that can move

in only two direction, to the left or the right. The angle between the platform and the pendulum and the angular velocity of this angle are chosen as the inputs of the system. Output is corresponding to the speed of the platform.

40

Fuzzification

First of all, the different levels of input and output are defined by specifying the membership functions for the fuzzy sets.

For similarity, it is assumed that all membership functions are spread equally. Hence, this explains why no actual scale is included in the graphs

41

Fuzzification

Input Angel

Input Angel Velocity

Output Speed

42

Rule Evaluation

The next step is to define the fuzzy rules. The fuzzy rules are a series of if-then statements.

For example: If angle is zero and angular velocity is zero

then speed is also zero.

If angle is zero and angular velocity is low

then the speed shall be low.

43

Rule Evaluation

The full set of rules are listed in table

Speed AngleNeg. high Neg. Low Zero Pos. Low Pos. High

V Neg. High Neg. High

E Neg. Low Neg. Low Zero

L Zero Neg. High Neg. Low Zero Pos. Low Pos. High

O Pos. Low Zero Pos. Low

C Pos. High Pos. High

44

Rule Evaluation

Suppose an example has 1. 0.75 and 0.25 for zero and positive low angles

2. 0.4 and 0.6 for zero and negative low angular velocities.

45

Rule Evaluation

Consider the rule

"if angle is zero and angular velocity is zero, the speed is zero".

46

Rule Evaluation

Consider the rule

"if angle is zero and angular velocity is negative low, the speed is negative low".

47

Rule Evaluation

Consider the rule

"if angle is positive low and angular velocity is zero, the speed is positive low".

48

Rule Evaluation

The Results overlap and are reduced to the following figure

49

Defuzzification

Defuzzification is used to choose an appropriate representative value as the final output.

The most common one is the centre of gravity

50

Sparse Fuzzy Systems

Sparse fuzzy systems can be used in situations where full knowledge of the problem domain is not available. Problem domain experts often work with only important fuzzy rules.

Self learning algorithms to tune the parameters of a fuzzy system for accuracy improvement can also lead to sparse fuzzy systems.

In most cases, parameter tuning involves the reshaping of the fuzzy sets in the rule antecedents. It can happen that the shrinking of the fuzzy sets leads to gaps between neighboring fuzzy sets.

Generating a sparse fuzzy system benefits from the reduced number of rules. (Chong 2004)

51

Sparse Fuzzy Systems

Sparse system can reduce T. The essential idea is based on the omission of less important fuzzy rules to form sparse fuzzy systems.

In sparse systems, it would be possible that inputs do not match any of the rule antecedents.

Fuzzy rule interpolation is used to infer these rules for the inputs from existing fuzzy rules in the system.

52

Interpolation overview

Tomato colours: back

IF colour = Red THEN it is Ripe

IF colour = Green THEN it is Unripe What about a yellow tomato? Potential tomato colours:

53

Automatic Construct Fuzzy Signature

Sub-Structure may be hidden in large data set.

More separable the elements in subspace, the easier sub-rule base selection is.

Finding suitable Π and Z0 affect each other.

54

Sugeno and Yasukawa Approach

Sugeno and Yasukawa (1991) introduced a solution for sparse rule-base generation.

It clusters output data sample and induces the

rules by projecting clusters of output to input domains.

Cons: it only produces necessary rules for the input-output sample data

55

Projection-based Fuzzy Rule Extraction

1. Perform c-Means to cluster data along output space. The FS index of Fuzzy c-Means can be used to get a optimal number of clusters.

2. For each fuzzy output cluster, all points contained in the cluster are projected back to input dimensions.

3. The projected points in each dimension are clustered again. In this procedure, the FS index is used in conjunction with the merging index. This process will produce multiple fuzzy clusters in each dimension.

4. Each of the clusters in the input dimension is a projection of the multi-dimensional input cluster to that input dimension. Then, the clusters from the individual dimensions are combined to form the multi-dimensional input cluster.

5. For each of the multi-dimensional clusters identified, a rule can be created.

56

Fuzzy c-Means

Let as a fuzzy partition C

cnc

n

n

ncU

1

221

11211

),.....,( 121 c

57

Fuzzy c-Means

Dunn defined a fuzzy objective function:

vi is cluster center of i set

Bezdek extended it to:

represents the deviation of data with . The

number m governs the influence of membership grades.

c

i

n

jijijD vxVUJ

1 1

22,

c

i

n

jij

mijm mvxXVUJ

1 1

21,;,

2ik vx

58

Fuzzy c-Means

Limitation: it needs to know the number of clusters.

How to find an optimal number of clusters.

A cluster validity index proposed by Fukuyama and Sugeno (FS):

ncxvvxUSc

i

n

jiik

mijc

1 1

22)( 2),(

59

Finding Suitable Subspace

Rules: age & experience to salary

IF Age = young & Exp = Little Then $ =Low

IF Age = young & Exp = Moderate Then $ =Low

IF Age = young & Exp = Good Then $ =High

IF Age = Middle & Exp = Little Then $ =Low

IF Age = Middle & Exp = Moderate Then $ =Ave

IF Age = Middle & Exp = Good Then $ =High

IF Age = Older & Exp = Little Then $ =Ave

IF Age = Older & Exp = Moderate Then $ =Ave

IF Age = Older & Exp = Good Then $ =High

60

Finding Suitable Subspace

Rule in a tree (Age/Exp/Con)

Age

Exp

Y M O

lM

G

L L H L A H A A H

Age

Exp

Y M O

G

L H L A H A H

Prune tree

61

Finding Suitable Subspace

Rule in a tree (Exp/Age/Con) back

Age

Exp

Y

M

O

l

M

G

L L A L A A H H H

Age

ExpM

O

l

M

G

L A L A H

Prune rule tree

62

Fuzzy Signatures in SARS Diagnosis

The following scheme is of some daily symptom signatures of patients:

Sore

Nauseapm

amCough

pm

pm

pm

am

fever

AS

9

128

4

12

8

63

Fuzzy Signatures in SARS Diagnosis

Two examples with linguistic values and fuzzy signatures.

25.0

25.0

5.02.0

2.0

0.0

0.0

1

slight

slight

normalslight

slight

none

none

A

0

25.09.0

7.04.0

4.0

2

none

slightsevere

highmoderate

moderate

A

64

Fuzzy Signatures in SARS Diagnosis

An aggregation method can compare components regardless of the different numbers of sub-components.

25.0

25.05.0

5.02.0

1 fA

0

25.08.0

6.04.0

2 fA

65

Fuzzy Signatures in SARS Diagnosis

Aggregation methods for different symptoms here are different with that of signatures of same symptoms.

Here, we define weights are:

fever = 1, Cough = 0.9, Nausea = 0.4, Sore = 0.25

back

4.0,25.0 21 oo AA