an evolutionary-fuzzy dss for assessing health status in multiple sclerosis disease

10
i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254 j ourna l homepage: www.ijmijournal.com An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease Massimo Esposito, Ivanoe De Falco , Giuseppe De Pietro National Research Council of Italy, Institute for High Performance Computing and Networking (ICAR), Via P. Castellino 111, 80131 Naples, Italy a r t i c l e i n f o Article history: Received 11 February 2011 Received in revised form 1 September 2011 Accepted 1 September 2011 Keywords: Health care Clinical decision support systems Multiple sclerosis Fuzzy logic Genetic models a b s t r a c t Assisted Living provides a long-term care option that combines supportive systems and services for monitoring and assessing the health status with activities of daily living and health care. Daily monitoring of the health status in subjects characterized by chronic and/or degenerative conditions is not possible in all those cases where the disease progression has to be evaluated only by a direct interaction between the patients and the healthcare structures on a regular basis, over time and for life. In this respect, this work proposes an evolutionary-fuzzy decision support system (DSS) for assessing the health status of subjects affected by multiple sclerosis (MS) during the disease progression over time. Such a DSS has been defined and implemented exploiting a novel approach devised to facilitate the design of fuzzy DSSs for medical problems. The approach is aimed at: (i) introducing a set of design criteria to encode the medical knowledge elicited from clinical experts in terms of linguistic variables, linguistic values and fuzzy rules with the final aim of granting the interpretabil- ity; (ii) defining a fuzzy inference technique to best fit the structure of medical knowledge and the peculiarities of the medical inference; (iii) defining an evolutionary technique to tune the formalized knowledge by optimizing the shapes of the membership functions for each linguistic variable involved in the rules. An experimental session has been carried out for evaluating, first of all, the approach on five medical databases commonly diffused in literature and for comparing it with other systems. After that, the evolutionary-fuzzy DSS for assessing MS patient’s health status has been quantitatively evaluated on 120 patients affected by MS and compared with other approaches. The achieved results have shown that our approach is very effective on the five databases, since it provides, on average, the second highest accuracy when compared to eight tools. Furthermore, as far as the classification of multiple sclerosis lesions is considered, the proposed system has turned out to outperform nine popular tools. © 2011 Elsevier Ireland Ltd. All rights reserved. 1. Introduction Recent advances in biomedical science, health care technolo- gies, and public health measures are radically impacting the management and monitoring of all those diseases that, due to Corresponding author. Tel.: +39 081 6139524; fax: +39 081 6139531. E-mail address: [email protected] (I. De Falco). the lack of a cure, have led to certain disability and death in the past. Since these illnesses are transformed into chronic and/or degenerative conditions, the primary benefit is the poten- tial for individuals affected by them to live longer and with higher quality of life. In this respect, Assisted Living provides a long-term care option that combines supportive systems and 1386-5056/$ see front matter © 2011 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.ijmedinf.2011.09.003

Upload: massimo-esposito

Post on 05-Sep-2016

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease

Am

MNN

a

A

R

R

1

A

K

H

C

M

F

G

1

Rgm

1d

i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254

j ourna l homepage: www.i jmi journa l .com

n evolutionary-fuzzy DSS for assessing health status inultiple sclerosis disease

assimo Esposito, Ivanoe De Falco ∗, Giuseppe De Pietroational Research Council of Italy, Institute for High Performance Computing and Networking (ICAR), Via P. Castellino 111, 80131aples, Italy

r t i c l e i n f o

rticle history:

eceived 11 February 2011

eceived in revised form

September 2011

ccepted 1 September 2011

eywords:

ealth care

linical decision support systems

ultiple sclerosis

uzzy logic

enetic models

a b s t r a c t

Assisted Living provides a long-term care option that combines supportive systems and

services for monitoring and assessing the health status with activities of daily living and

health care. Daily monitoring of the health status in subjects characterized by chronic and/or

degenerative conditions is not possible in all those cases where the disease progression

has to be evaluated only by a direct interaction between the patients and the healthcare

structures on a regular basis, over time and for life. In this respect, this work proposes an

evolutionary-fuzzy decision support system (DSS) for assessing the health status of subjects

affected by multiple sclerosis (MS) during the disease progression over time. Such a DSS has

been defined and implemented exploiting a novel approach devised to facilitate the design

of fuzzy DSSs for medical problems. The approach is aimed at: (i) introducing a set of design

criteria to encode the medical knowledge elicited from clinical experts in terms of linguistic

variables, linguistic values and fuzzy rules with the final aim of granting the interpretabil-

ity; (ii) defining a fuzzy inference technique to best fit the structure of medical knowledge

and the peculiarities of the medical inference; (iii) defining an evolutionary technique to

tune the formalized knowledge by optimizing the shapes of the membership functions for

each linguistic variable involved in the rules. An experimental session has been carried out

for evaluating, first of all, the approach on five medical databases commonly diffused in

literature and for comparing it with other systems. After that, the evolutionary-fuzzy DSS

for assessing MS patient’s health status has been quantitatively evaluated on 120 patients

affected by MS and compared with other approaches. The achieved results have shown that

our approach is very effective on the five databases, since it provides, on average, the second

highest accuracy when compared to eight tools. Furthermore, as far as the classification of

multiple sclerosis lesions is considered, the proposed system has turned out to outperform

nine popular tools.

degenerative conditions, the primary benefit is the poten-

. Introduction

ecent advances in biomedical science, health care technolo-ies, and public health measures are radically impacting theanagement and monitoring of all those diseases that, due to

∗ Corresponding author. Tel.: +39 081 6139524; fax: +39 081 6139531.E-mail address: [email protected] (I. De Falco).

386-5056/$ – see front matter © 2011 Elsevier Ireland Ltd. All rights resoi:10.1016/j.ijmedinf.2011.09.003

© 2011 Elsevier Ireland Ltd. All rights reserved.

the lack of a cure, have led to certain disability and death in thepast. Since these illnesses are transformed into chronic and/or

tial for individuals affected by them to live longer and withhigher quality of life. In this respect, Assisted Living providesa long-term care option that combines supportive systems and

erved.

Page 2: An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease

c a l i

e246 i n t e r n a t i o n a l j o u r n a l o f m e d i

services for monitoring and assessing the health status withactivities of daily living and health care, aiming at providingassistance customized to the patients’ needs so as to enrichtheir lives and promote independence and well-being.

Monitoring the subject’s health status in the daily livingis becoming a common practice due to the increasing spreadin aging population of chronic diseases, such as obstructivepulmonary, cardiovascular diseases, etc., and has led to thedevelopment of low-cost, innovative technological systemsto daily support and assist patients, to prevent and controlpathology ongoingness, to adjust drug therapies, and to avoidhospitalization. Yet, the daily health status monitoring insubjects with chronic and/or degenerative conditions is notalways possible as it strongly depends on disease-specific fea-tures: an example is that of neurodegenerative diseases, whichrequire specific approaches for assessing the subject’s healthstatus, since their progression can be evaluated only by a directinteraction between the patients and the healthcare struc-tures on a regular basis, over time and for life.

Concerning the latter typology of pathology, the most rep-resentative example is represented by multiple sclerosis (MS),characterized by multiple demyelinated lesions, involving thebrain and spinal cord, that cause damage or destruction ofmyelin surrounding nerve fibers, and, thus, interrupt commu-nications between the nerves and the rest of the body [1]. Ithas a strong impact on the patient’s health status since it pro-duces neurological dysfunctions, such as numbness, impairedvision, loss of balance, weakness, bladder dysfunction, andpsychological changes. Many MS cases evolve over a longperiod (20–30 years) with remissions and exacerbations, but,in almost half of all cases, it relentlessly progresses to severedisability and premature death [2]. Thus, disease outcome isrepresented by chronic and/or degenerative conditions, whichare highly variable between affected individuals. Unfortu-nately, there is a lack of prognostic markers: indeed, for anindividual patient, the severity outcome and the rate of pro-gression of MS are impossible to predict.

Moreover, there is no way for monitoring the health statusof MS patients in their daily living since the only manner tocontrol the disease progression relies on clinical examinationsupported by laboratory investigations including magneticresonance imaging (MRI) to visualize lesions both in the courseof MS and in the assessment of treatment effects [3,4]. Forassessing the patient health status correctly, demonstrationof distribution of lesions in both time and space is necessary,delaying the time for an appropriate follow-up [5]. Indeed, theuse of MR images as MS marker requires the expert’s knowl-edge and intervention to identify MS lesions; nevertheless,such a task is very thorny and time-consuming due to thehuge amount of MR images to be examined and to the variablenumber, size and spatial distribution of MS lesions per image.

In the past different approaches have been proposed,and, in particular, operator-assisted methods based on localthresholding and automated methods assessing multipleparameters have been successfully employed for assess-ing the health status in MS patients by identifying lesions

[6–9]. Nevertheless, these methods suffer from one importantlimitation: they rely on mathematical models based on thresh-olding to classify MS lesions. Hence, they neither take intoaccount the fuzziness of input data nor reproduce the expert’s

n f o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254

decision-making process applied in a vague-laden domainsuch as medicine. In fact, the decision-making model everyphysician has in mind to perform heuristic diagnosis is oftenpervaded by uncertainty and vagueness. Thus expert knowl-edge abounds with imprecise formulations that do not dependon rhetorical inability, but are an intrinsic part of expertknowledge acquired through laborious experience [10]. Anyformalism disallowing uncertainty, such as crisp mathemati-cal models, is hence inapt to capture this knowledge [10] andcan represent an unrealistic oversimplification of reality, lead-ing to possible wrong interpretations when compared to adirect observation.

Fuzzy Logic [11] has widely demonstrated its capability toovercome such critical issues in medical applications, andmany decision support systems (DSSs) based on it have beenproposed in literature [12–14]. This is due to the fact that FuzzyLogic formalism is suitable to deal with the imprecision intrin-sic to many medical problems, so as to offer a more realisticinterpretation for the medical inference. Although, in theory,a DSS based on Fuzzy Logic could be proficiently used for iden-tifying lesions in the assessment of MS patient’s health status,in practice, the design of a Fuzzy DSS is a complex multi-stepprocess in which the most usual way for collecting medicalknowledge is asking the expert to write “if-then” rules. Yet,neuroradiologists usually describe their knowledge by meansof incomplete rules that typically model pieces of positiveevidence only. Moreover, after formalizing the expert’s knowl-edge under the form of rules, the designer has to choose theshape and location of membership functions for all the lin-guistic values related to all the linguistic variables involved.This requires both medical expertise and technical interven-tion along with great effort to identify which among the designchoices are suited to the given problem.

To face all these issues, this work proposes an evolutionary-fuzzy DSS for assessing the health status of subjects affectedby MS during the disease progression over time, which isaimed at supporting clinicians in the identification of MSlesions. Such a system has been defined and implementedby exploiting a novel approach devised to ease the design offuzzy DSSs so as to involve the medical practitioner in thedefinition of the domain knowledge only. Such an approach isaimed at: (i) introducing a set of design criteria to encode thehigh-level, specialized medical knowledge elicited from clin-ical experts in terms of linguistic variables, linguistic valuesand rules with the final aim of granting the interpretability;(ii) defining a fuzzy inference technique to best fit the struc-ture of medical knowledge and the peculiarities of the medicalinference; (iii) defining an adaptive technique based on an evo-lutionary algorithm, i.e. differential evolution [15], to tune theformalized knowledge by optimizing the shapes of the mem-bership functions for each linguistic variable involved in therules.

In order to address the three above reported goals, theproposed approach has been conceived as general purposeand structured and described in terms of three design stages,respectively, knowledge representation, knowledge reasoning

and knowledge tuning. Moreover, an experimental session hasbeen carried out for evaluating quantitatively its performanceon five medical databases commonly diffused in literature. Inparticular, five prototypal DSSs have been built in accordance
Page 3: An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease

i n f o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254 e247

wepobfqcn

aaampbwalfomseMaat

2

2

TowaIdlgscnsmU�

efi

isLiff

Table 1 – Design criteria description.

Criteria Description

Semantics Linguistic variables and their values havea semantic meaning well defined by theexperts.

Distinguishability The universe of discourse for eachlinguistic variable is determineddepending on the range of possible valuesin the input dataset. Since all linguisticvalues have a semantic meaning, thecorresponding fuzzy sets are well disjointwith defined ranges in the same universeof discourse.

Coverage Any element from the universe ofdiscourse belongs to at least one of thefuzzy sets defined for the linguisticvalues.

Normalization For each linguistic value at least oneelement of the universe of discourse hasa membership value equal to one.

i n t e r n a t i o n a l j o u r n a l o f m e d i c a l

ith the approach, each of them being associated to a differ-nt database, in order to compare their results against thoserovided by a set of widely used machine learning methodsn the same set of databases. Successively, the approach haseen specifically applied to build an evolutionary-fuzzy DSSor assessing MS patient’s health status. This latter has beenuantitatively evaluated on 120 patients affected by MS andompared with other existing tools for assessing its effective-ess.

To the best of our knowledge, none of existing methodsnd systems proposed in literature adopts a similar hybridpproach for assessing MS patient’s health status. Manypproaches in literature [16,17] use fuzzy logic both to seg-ent MR images and to classify MS lesions, i.e. to cluster single

ixels/voxels into homogeneous groups and identify them aselonging to a brain tissue or to an MS lesion. Differently,e are not interested in segmenting MR images, rather ourpproach is focused on the definition of a DSS that uses fuzzyogic, and, in particular, fuzzy rules provided by doctors, onlyor the classification of MS lesions by starting from the resultsf a preliminary segmentation step. Of course, the DSS perfor-ance depends on the reliability of the previous segmentation

tep, and, as a result, the proposed DSS complements existingfforts already carried out for the segmentation of MS lesions.ore specifically, if we changed the segmentation algorithm

different input dataset could be created, which would prob-bly contain a different set of possible lesions. Nonetheless,he DSS would act in exactly the same way.

. Methods

.1. Knowledge representation

he design of a fuzzy DSS for medical problems requires, firstf all, the definition of the domain knowledge in cooperationith clinical experts by means of interviews, questionnaires

nd observation of them at work, while they “think aloud”.n more detail, the structure of this domain knowledge isescribed and represented in terms of linguistic variables,

inguistic values and membership functions. The goal of lin-uistic variables is to ease a gradual transition between states,o as to naturally express vagueness in measurements, unlikerisp variables. A linguistic variable is characterized by itsame tag, a set F of linguistic values, and the member-hip functions mF of these values; linguistic values assign aembership value to elements u, within a predefined range

(known as the universe of discourse), as follows: F = {(u,

F)|u ∈ U and �F: U → [0,1]}.In order to grant the interpretability of the knowledge mod-

led via linguistic variables, linguistic values and membershipunctions, we require that the set of design criteria, describedn Table 1, be verified.

Moreover, knowledge about the medical decision-makings formalized in terms of fuzzy if-then rules relying on thetructure defined for the domain knowledge. Indeed, Fuzzy

ogic provides a procedural morphology enabling to approx-mate human reasoning capabilities of drawing conclusionsrom existing data: namely, it is possible to infer new truthsrom old ones. The fuzzy procedural morphology relies on

Orthogonality For each element of the universe ofdiscourse, the sum of all its membershipvalues is equal to one.

rules, defined as conditional statements written in the form “ifantecedent then consequent”, where antecedent is a fuzzy-logicexpression composed of one or more simple fuzzy expressionsconnected by fuzzy operators, and consequent is an expres-sion that assigns linguistic values to the output variables.

2.2. Knowledge reasoning

The definition of a specific fuzzy inference technique for med-ical problems, which require a classification task as in thecase of MS, is driven by a set of considerations about themedical knowledge. Typically, an essential part of the medicalknowledge in problems of medical classification is usually rep-resented by means of incomplete rules, i.e. that do not coverall the possible outputs. For instance, in two-class problems,the experts’ rules often collect all the positive evidence only,looking for those manifestations that could be sufficient toestablish a positive conclusion. Moreover, typically, physiciansare not used to formulating rules in which the antecedent partis false. As a result, the classical modus ponens could be notadequate to the medical inference.

Furthermore, in fuzzy theory, conjunction and disjunctionoperators are often modeled as fuzzy-set intersection andunion in the rules. This involves that, for instance, the disjunc-tion of two contiguous and orthogonal fuzzy sets (Fig. 1(a)),modeled as the fuzzy-set union, can generate a scarcely intu-itive behavior.

Indeed, in the area in which the two fuzzy sets are par-tially overlapping, the uncertainty is the greatest between thecorresponding values of the two fuzzy sets (Fig. 1(b)). Differ-ently, in the common sense, the disjunction in the overlappingarea should have an additive behavior, that means the valuesof uncertainty of both the fuzzy sets in that area should besummed (Fig. 1(c)). To overcome these issues, we introduce a

fuzzy inference technique resembling the FIRE method pro-posed in [18] to filter images. Such a technique also defines adhoc operators for aggregation, implication, accumulation anddefuzzification coherently with the Lukasiewicz Logic [19].
Page 4: An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease

e248 i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254

0.2

0.4

0.6

0.8

1

µA(X) µB(X)

0.2

0.4

0.6

0.8

1

µA(X) µB(X)

0.2

0.4

0.6

0.8

1

µA(X) µB(X)

(c)(b)(a)

Fig. 1 – Disjunction of two contiguous and orthogonal fuzzy sets.

In the following, the basic knowledge about this techniqueis described in detail.

Definition 1. Given M fuzzy rules, N continuous input vari-ables v1, v2, . . . vN and one discrete output variable s, theformulation of the fuzzy rules is defined as follows:

IF(v1, A11)AND/OR(v2, A12) . . . (vN, A1N)THEN(s, B1)IF(v1, A21)AND/OR(v2, A22) . . . (vN, A2N) THEN(s, B2)

. . . . . .

IF(v1, AM1)AND/OR(v2, AM2) . . . (vN, AMN)THEN(s, BM)ELSE(s, BE)

Aij = (aj)AND/OR(bj) . . . AND/OR(zj)

(1)

where Aij is a logical expression associated to each rule i andcalculated for the variable j in accordance with its linguisticvalues aj, bj, . . . zj, and Bi and BE are the linguistic values asso-ciated respectively to rule i and to the ELSE rule calculated forthe output variable s. This formulation models a disjunctivesystem of rules where at least one rule must be satisfied, i.e.the rules are linked by OR connectives. The ELSE rule is acti-vated when the other rules are weakly satisfied or not satisfiedat all.

Definition 2. The strength level �i of the rule i is calculatedin terms of the degrees of membership �Aj for the antecedentclause. Lukasiewicz t-norm and s-norm are used as aggrega-tion operators since they verify both the law of contradictionAij ∩ NEG(Aij) = 0 and that of excluded middle Aij ∪ NEG(Aij) = 1.

Based on Lukasiewicz norms, if the antecedents are con-nected by AND operator the strength level is defined as:

�i = max

⎧⎨⎩0,

N∑j=1

�Aij(vj) − (N − 1)

⎫⎬⎭ (2)

whereas, when the antecedents are connected by OR operator,it is defined as:

�i = min

⎧⎨⎩1,

N∑j=1

�Aij(vj)

⎫⎬⎭ (3)

The logical expression Aij associated to each rule i is calcu-lated for the variable j in terms of the degrees of membership�� for its linguistic values aj, bj, . . . zj in the antecedent clause.When those linguistic values are connected, respectively, bythe AND operator and the OR operator, the logical expressionAij is calculated as:

Aij= max

{0,

∑�

�� (vj)−(Nc − 1)

}�=aj, bj, ...zjNc=card

{aj, bj, ...zj

}(4)

Aij = min

{1,

∑�

��(vj)

}� = aj, bj, ...zj (5)

Definition 3. The strength level �E of the ELSE rule is calcu-lated by applying a NOT operator to the strength levels �i ofthe other rules, since the ELSE rule is activated when all theother ones are partially or completely unsatisfied. Moreover,since the rules are connected by OR connectives, the resultsobtained by applying the NOT operator to the strength levels�i are combined in a conjunctive system, i.e. by means of ANDconnectives. So:

�E = max

{0,

M∑i=1

�i − (M − 1)

}(6)

Definition 4. The operator of implication is defined as theLukasiewicz t-norm implication:

x → y = max{

0, x + y − 1}

(7)

This operator does not verify the properties of theclassical modus-ponens implication, yet it has beenused since in a fuzzy DSS for medical classificationwe are not interested in rules whose antecedent partis false.

Definition 5. The operations of accumulation anddefuzzification are calculated as the centre of gravityof the strength levels preventively accumulated with

Page 5: An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease

i n f

ru

wusf

Dosi

m

2

Aftliwemsndfie

fcqrafmctptasdhtsopn

i n t e r n a t i o n a l j o u r n a l o f m e d i c a l

espect to the singleton spikes defined in the outputniverse:

COG(s) =

∑p

l=1 min

{1,

Ql∑k=1

�k ∗ xBk

}

∑P

l=1 min

{1,

Ql∑k=1

�k

} xBk= xl k = 1, 2...Ql

(8)

here P is the number of singleton spikes defined in the outputniverse, Ql is the number of rules with the same singletonpike xBk as output value and the accumulation operator usedor each Bk is the Lukasiewicz s-norm.

efinition 6. The output of the centre of gravity is a continu-us appraisal value. The final discrete output is produced byimply identifying the singleton spike whose numerical values the nearest to the appraisal:

in{|COG(s) − xBi|} i = 1, 2, . . . M + 1 (9)

.3. Knowledge tuning

fter formalizing the expert’s domain knowledge under theorm of linguistic variables, linguistic values and fuzzy rules,he shape and location of membership functions for all theinguistic values related to all the linguistic variables involvedn the rules have to be opportunely defined. In such a sense,e propose an evolutionary technique based on differential

volution (DE) with the aim of optimizing the shapes of theembership functions for each linguistic variable. DE is a

tochastic evolutionary optimization strategy that presentsoticeable performance in optimizing a wide variety of multi-imensional and multimodal objective functions in terms ofnal accuracy and robustness, and overcomes many of thexisting global optimization techniques [20,21].

Given a maximization problem with q real parameters, DEaces it starting with a set, called population, of Npop randomlyhosen solution vectors, each made up by q real values. Theuality of each solution is evaluated and is represented by aeal value ˚, called fitness. At each generation new vectorsre generated by a combination of vectors randomly chosenrom the current population. Each resulting vector is then

ixed with a predetermined target vector. This operation isalled recombination and produces a trial vector. If this lat-er is better than its target vector, then replaces it in the newopulation. Ten different schemes were defined in literatureo produce the candidate trial vector [15]: we have carried out

preliminary set of experiments to investigate if one of thosechemes achieves, on average, better performance over theatabases faced in this paper. Since no clear winning schemeas been found, for brevity’s sake we have chosen to reporthroughout this paper the results related to the DE/rand/1/bin

cheme only. In it, a randomly chosen vector is perturbed byne difference vector, and binomial crossover is applied. Thearameters are set as follows: population size Npop equal to 20,umber of generations Gen equal to 100, CR = 0.8, F = 0.5.

o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254 e249

Given a database and a set of fuzzy rules, we make useof DE to find the best set of the membership functions forthe variables involved in the rules so that the highest correctclassification rate can be achieved.

In accordance with the design criteria formulated inTable 1, we have chosen the shape of the membershipfunctions as trapezoidal. This choice involves that each mem-bership function can be represented by an ordered set of fourreal values {˛1, ˛2, ˛3, ˛4}, where the first represents the start-ing value of the leading edge, the second its ending value, thethird the starting value of the trailing edge and the fourth itsfinal value. Hence, the trapezoid will be delimited by the fourpoints {(˛1; 0); (˛2; 1); (˛3; 1); (˛4; 0)}. Of course, a trapezoid canbe reduced to a triangle (˛2 = ˛3) by optimization, if this latteris the most suited shape for the value under account.

In order to tune the trapezoidal membership functions inaccordance with the orthogonality, the shapes of any couple ofadjacent fuzzy values cross at a level of 0.5, and, more impor-tantly, ˛3 and ˛4 for the first fuzzy value of a variable coincidewith ˛1 and ˛2 of next value of the same variable, and so on.Furthermore, for the first (respectively, the last) value of eachlinguistic variable the two first (two last) points coincide.

In each DE solution vector the number of real values rep-resenting the membership functions for the Ni

valvalues of

the ith variable is lower than 4 × Nival

: in general it is givenby 2 × (−Ni

val1) + 2. Each solution in the population is a vec-

tor of real numbers encoding, for each linguistic variable, thecontrol points for the trapezoid representing the shape of themembership function of any linguistic value associated to thevariable. The vector has a length equal to

∑Nvar

i=1 2 · (Nival

− 1) + 2where Nvar is the number of linguistic variables. As fitnessfunction of any solution, we have considered the percentage ofcorrectly classified examples in the training sets. In fact, as it istypical of any approach to classification, also in our approachwe firstly perform training and then we test the generalizationability achieved. More precisely, for the learning mechanism,an Nf -fold cross-validation has been carried out. This meansthat the database is divided into Nf subsets, and Nf trainingsessions are effected: in the ith session the ith subset is kept fortesting and the algorithm is shown the items in the remainingNf − 1 folds. The best solution found in each training session isthen evaluated on the testing set. Finally, the average results ofthese Nf classifiers are computed, and the best among themin terms of highest generalization ability is seen as the bestsolution to the problem, and the shapes for the membershipfunctions represented by this solution are taken into accountin the final system. For our experiments we have set Nf = 10.

2.4. Experimental evaluation for assessing theapproach

To evaluate the effectiveness of our approach, we have takeninto account five databases coming from the medical domainand widely used in literature. They are contained in the UCIMachine Learning Repository [22]. Below a very short descrip-tion and a summary of their features are provided in Table 2.

Applying the proposed approach, we have built five proto-typal DSSs, each of them associated to a different database,in order to compare their results against those provided by aset of widely used machine learning classifiers on the same

Page 6: An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease

e250 i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254

1) IF [Sph ericity is (Mod era te OR High )] AND [Compa ctness is Stron g]AND [Volume is Small] AND [Tiss ueCon tras t is Great] AND [Surr oun ding WhiteMat ter is Completely] TH EN [Tissu eStru cture is Abn ormal]

2) IF [Sph ericity is Mod era te] AND [Compa ctness is St ron g]AND [Volu me is Mediu m] AND [Tiss ueCon trast is Great] AND [S urro und ingWhit eMatt er is (Almos tCompletel y ORCompletely)] THEN [Tissu eStru ctur e is Abnorma l]

3) IF [Com pactness is Stron g]AN D [Volume is La rge] AN D [T issu eCon trast is Great] AN D [Surr ound ing Whi teMa tter is (Partiall y OR A lmost Complete ly OR Completely)] T HEN [Tiss ueStructu re is Abnorm al]

4) EL SE [Tissu eStru cture is Normal]

Fig. 2 – The complete set of rules: three if-then rules for identifying actual lesions, and the default ELSE one.

Table 2 – Database description.

Database Description Instances Attributes Classes

Cancer Breast cancer cases from University of Wisconsin atMadison Hospital.

699 9 2

Heart Patients with heart disease originally obtained fromCleveland Clinic Foundation.

303 13 2

Haberman’s survival Cases on the survival of patients who had undergonesurgery for breast cancer from University of Chicago’sBillings Hospital

306 3 2

Liver Blood tests related to liver disorders arising fromexcessive alcohol consumption.

345 6 2

Pima diabetes Patients with signs of diabetes according to the WorldHealth Organization criteria obtained from National

Kidn

768 8 2

Institute of Diabetics and Digestive and

set of databases. In particular, we have decided to take as ref-erence the results described in [23] and obtained by means ofalgorithms provided in the Waikato Environment for Knowl-edge Analysis (WEKA) system release 3.6.2 [24], i.e. C4.5, NaïveBayes (NB), support vector machine (SVM), artificial neural net-work (NN), 1-nn and One-R.

To run our DSSs we need a set of fuzzy rules express-ing expert’s knowledge about each of these medical domains.Since contacting a group of physicians for each of these dis-eases would have been quite laborious, we have decided toget those sets of rules by extracting them in an automaticway through the Ripple-Down Rule learner (Ridor) classifica-tion algorithm contained in WEKA. Ridor provides a set of

crisp if-then rules each composed by AND-connected clausesexpressing relationships between database attributes andconstant values. Then, these rules must be suitably trans-lated into fuzzy ones, e.g. a clause as (x2 > 14.57), where

Table 3 – Features describing a WMPL.

Variable Unit – range

Volume Voxel3–10,522

Sphericity Dimensionless0.01–1.23

Compactness Dimensionless0.31–1.98

Tissue contrast Percentage0.56–1.0

Surrounding white matter Percentage0.32–1.0

ey Disease.

1.05 ≤ x2 ≤ 20.42, should be translated into something like(x2 IS high). These two steps introduce imprecision in theresulting fuzzy rules, the former because the crisp rulesobtained might not be the best possible ones to achievegood knowledge, whereas the latter step needs a fine-tuningactivity for the translation from crisp to fuzzy, and highsensitivity in realizing the most suitable number of fuzzyvalues for each variable. Hence, we feel that the resultsachieved by our DSSs may be even better if a good setof rules is provided by an expert of the domain underinvestigation.

2.5. Case study of multiple sclerosis

As a specific application, the approach has been followed tobuild an evolutionary-fuzzy DSS for assessing the health sta-tus of subjects affected by MS during the disease progression

Description

Lesion volume in terms of the number of voxels.

Degree of sphericity of a lesion. The moreelongated the lesion is and the more it deviatesfrom a sphere, the lower sphericity will be.Degree of compactness of a lesion. For a givenshape, compactness is high either if the volumeis large or if the enclosing surface is small, i.e.the object is strongly compact.Minimum color contrast to detect a WML in themultiparametric space.Amount of White Matter enclosing a lesion.

Page 7: An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease

i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f

Table 4 – Linguistic variables and terms.

Variable Terms

Volume Small, medium, largeSphericity Low, moderate, highCompactness Weak, strongTissue contrast Little, greatSurrounding white Bit, partially, almost completely,

otdmUMtibtFa

ntaipsuasr

a

wis

3

3a

Wfc

end of the run. We show in Table 7 the average results achieved

matter completelyTissue structure Normal, abnormal

ver time, which is aimed at supporting clinicians in the iden-ification of MS lesions. This DSS has been evaluated on aataset, opportunely anonymized, collected at the Depart-ent of Bio-Morphological and Functional Sciences of theniversity of Naples “Federico II”. In particular, starting fromR brain images of 120 patients with clinically definite MS,

he multiparametric segmentation procedure proposed in [25]s applied to the whole data set in order to identify normalrain tissues or clusters of potentially abnormal white mat-er voxels, labeled as White Matter Potential Lesions (WMPLs).or each WMPL, the features described in Table 3 represent thectual input data for the DSS.

Starting from these features, the medical knowledgeeeded to classify WMLs has been defined in cooperation withhe team of physicians and can be stated, in natural language,s follows. The tissue composing a WMPL is abnormal if the lesion

s somewhat surrounded by WM, characterized by a strong com-actness and greatly contrasted in the multiparametric space. Thephericity is moderate or high in small lesions, whereas, as their vol-me increases, the sphericity starts decreasing progressively. Finally,s volume increases and sphericity starts lessening, a lesion can beurrounded by gradually decreasing WM and its compactness stillemains high.

In accordance with this knowledge, the linguistic variablesnd values shown in Table 4 have been identified.

These linguistic variables and values have been used torite the three “if-then” rules aimed at identifying the pos-

tive cases, i.e. when a potential lesion is an actual one. Fig. 2hows those rules and the default ELSE one.

. Results and discussion

.1. Experimental evaluation for assessing thepproach

e show in Table 5 the average results achieved over the tenolds, and those for the best fold in terms of the highest per-entage of accuracy on the test set. The results are arranged

Table 5 – Results of our systems on the five datasets.

Average results over the 10 folds

%Tr %Te Se Sp

Cancer 97.75 95.97 0.97 0.95

Heart 86.20 84.64 0.77 0.90

Hab.’s survival 77.68 76.55 0.22 0.94

Liver 75.41 68.18 0.85 0.48

Pima diabetes 77.72 75.73 0.55 0.86

o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254 e251

in terms of percentage of accuracy on the training set %Tr andon the test set %Te, sensitivity Se and specificity Sp.

Sensitivity on Haberman’s survival problem might appearquite low, yet this is due to the fact that this database is highlyunbalanced, having just 26% of positive cases. In fact, also allthe other classification methods considered in this paper showlow sensitivity, ranging from 0.20 for NB and Ridor up to 0.29for NN.

Table 6 reports the comparison among the results achievedon the five medical databases by our systems and by the sevenothers on the basis of 10-fold classification accuracy. For eachdatabase the algorithm reporting the best performance on itis shown in bold. Moreover, the last row shows the results forRidor too, as it is the starting point from which our systemstry to improve those rules. The same row shows in bracketsthe number of rules obtained by that algorithm. Similarly, therow related to our systems shows in brackets also the num-ber of fuzzy rules achieved starting from the crisp ones basedon Ridor. The penultimate column in Table 4 shows the aver-age accuracy for all the algorithms: our system is the secondbest one, and only DEPRO has slightly higher average per-formance, the difference being 0.36%. Moreover, our systemsalways perform better than their rule-provider Ridor, and onsome databases the improvement is noticeable, for exampleon Heart it is of 6.87%, on Liver of 4.13% and on Haberman’ssurvival of 8.25%. With regard to the two problems mentionedabove in obtaining the fuzzy rules for our systems, it is evi-dent that they have very good performance even if we act ina non-optimized way, and they may have a really promisingbehavior in diagnostic problems if we can provide them withgood sets of rules coming from experts.

3.2. Case study of multiple sclerosis

Binary classifications of WM lesions have been produced andcompared with the gold standard and accuracy, sensitivity,and specificity have been calculated. This gold standard hasbeen determined by a team of neuroradiologists who has clas-sified only 1905 out of 2844 detected WMPLs as proven lesions.In more detail, as soon as the system starts the adaptive phasefor each fold, better and better classification results in termsof higher accuracy values are found: while in the initial gener-ation good solutions have percentages of correct classificationof around 30–35%, as the number of generations increasesthese values get rapidly better, up to around 85–90% at the

over the ten folds, and those for the best fold in terms of thehighest percentage of accuracy on the test set. The results arearranged in terms of percentage of accuracy on the training

Results for the best fold in terms of %Te

%Tr %Te Se Sp

97.72 100 1.0 1.086.98 96.42 0.9 1.076.89 82.75 0.42 0.9575.32 84.84 0.87 0.7776.76 84.00 0.71 0.86

Page 8: An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease

e252 i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254

Table 6 – 10-Fold classification accuracy, average over the five databases, and the related ranking for each algorithm.

Cancer Heart Hab.’s survival Liver Pima diabetes Average Rank

C4.5 94.56 76.60 71.90 68.70 71.48 76.75 6NB 95.99 83.70 74.83 56.52 75.78 77.36 5SVM 96.70 84.10 73.52 58.26 77.08 77.93 4NN 95.56 78.10 72.87 71.59 75.39 78.70 31-nn 95.85 75.19 67.65 62.90 71.48 74.61 8One-R 91.55 71.11 73.02 54.78 71.35 72.36 9DEPRO 96.97 83.74 80.36 71.01 75.37 81.49 1Our DSSs 95.97 (9) 84.64 (6) 76.55(7) 68.18 (3) 75.73(4) 81.13 2Ridor 95.75(4) 77.77(6) 68.30(6) 64.05(3) 75.00(4) 76.17 7

Table 7 – Results of our system.

Average results over the 10 folds Results for the best fold in terms of %Te

%Tr %Te Se Sp %Tr %Te Se Sp

Multiple sclerosis 89.10 88.79 0.88 0.88 88.71 92.93 0.96 0.84

Table 8 – Values for the shapes of the trapezoids.

Variables Terms Values

˛1 ˛2 ˛3 ˛4

Surrounding whitematter

Bit 0.32 0.32 0.33 0.38Partially 0.33 0.38 0.40 0.41Almost completely 0.40 0.41 0.46 0.95Completely 0.46 0.95 1.00 1.00

Compactness Weak 0.31 0.31 0.36 0.74Strong 0.36 0.74 1.98 1.98

Tissue contrast Little 0.56 0.56 0.61 0.92Great 0.61 0.92 1.00 1.00

Volume Small 3 3 3177 3529Medium 3177 3529 7051 7697Large 7051 7697 10,522 10,522

Sphericity Low

Moderate

High

set %Tr and on the test set %Te, sensitivity Se and specificitySp.

To evaluate the effectiveness of our system, we have com-pared its results against those provided by a set of widely usedmachine learning classifiers provided by the WEKA system [24]and listed in Table 8. Parameter values used for any techniqueare those set as default in WEKA. Thus, Table 9 reports thecomparison among the results achieved by our system andby the other classifiers on the basis of 10-fold classificationaccuracy.

The run related to fold 7 shows the highest generalizationability, so it is the most suitable for a real situation wherenew cases arrive and should be classified. Its solution classifies

Table 9 – 10-Fold classification accuracy.

Oursystem

Bayesnet

Multilayerpercep-tronartificialneuralnetwork

Radial basisfunctionartificialneuralnetwork

K-sta

Accuracy 88.79 78.09 88.04 82.20 85.90

0.01 0.01 0.03 0.100.03 0.10 1.02 1.031.02 1.03 1.23 1.23

correctly 89.13% of all the dataset. The membership functionshapes for the linguistic values taken on by the five variablesare reported in Table 8.

As highlighted in Table 9, our system has shown thecapability to obtain the best results outperforming the otherreported tools. Moreover, it has resulted extremely simple andintuitive to be understood also by the physicians, thanks to thepresence of the fuzzy rules expressing expert’s knowledge. Onthe contrary, the lack of a simple insight as to how the othermachine learning classification methods work and the scarce

interpretability of their classification results make them toocomplex to be understood by a non-technical audience and, asa result, scarcely appealing and trustworthy by the physicians.

r Bagging Adaboost

NaiveBayestree

Rippledownrule

Votingfeatureinterval

87.86 82.34 86.81 86.91 72.39

Page 9: An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease

i n f o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254 e253

4

TaicheDtotfiSmjtiboon

mfiodnsdvad

pspstct

A

TMNstseBCsfi

Summary pointsWhat was already known on the topic

• The difficulties of assessing the daily health status insubjects with chronic and/or degenerative conditions,such as multiple sclerosis, are already known.

• They are summarized in Introduction to enlighten theinterest of developing a novel solution for assessingthe health status of subjects affected by multiple scle-rosis during the disease progression.

What this study added to our knowledge

• We propose an evolutionary-fuzzy DSS aimed at iden-tifying MS lesions for the health status assessment,based on a general purpose approach for easing thedesign of fuzzy DSSs for medical problems.

• The proposed fuzzy inference technique enables tobest fit the structure of medical knowledge and thepeculiarities of the medical decision-making.

• The adaptive technique based on differential evolutionenables to tune the formalized knowledge by optimiz-ing the shapes of the membership functions for eachlinguistic variable involved in the rules.

r

i n t e r n a t i o n a l j o u r n a l o f m e d i c a l

. Conclusions

his paper has presented an evolutionary-fuzzy DSS forssessing the health status of subjects affected by MS dur-ng the disease progression, which is aimed at supportinglinicians in the identification of MS lesions. Such a systemas been defined and implemented by exploiting a gen-ral purpose approach devised to ease the design of fuzzySSs for classification problems in medicine. A first aim of

he approach has been to involve the medical practitionernly in the definition of the fuzzy rules and not also in theime-consuming process of tuning the fuzzy membershipunctions. This has been obtained by defining and implement-ng an adaptive technique based on differential evolution.uch a technique has been devised to guarantee the linguisticeaning and interpretability of the rules and to obtain con-

unctly a good classification accuracy and a short executionime. A second aim has been to best fit the structure of med-cal knowledge and the peculiarities of the medical inferencey means of a fuzzy inference technique able to face the lackf exclusionary rules in medical classification problems with-ut forcing the medical practitioners to write also rules for theegative evidence.

The approach has been quantitatively evaluated on fiveedical databases well known in literature. In particular,

ve prototypal DSSs have been built in accordance withur approach, each of them being associated to a differentatabase to evaluate the effectiveness of our approach. In theear future we are interested in contacting physicians to get aet of fuzzy rules for each of the five data sets. Should they beifferent from the ones extracted through Ridor, in terms ofariables involved or of number of fuzzy values for each vari-ble, we plan to reevaluate our DSSs so as to investigate theependence of the quality of results on the set of rules.

Finally, the evolutionary-fuzzy DSS built exploiting the pro-osed approach has been evaluated on a real clinical datasethowing its effectiveness in assessing the health status of MSatients. As a concluding remark, since the knowledge repre-entation model, the fuzzy inference method and the adaptiveechnique have a general basis, in the future, the approachould be applied for building advanced DSSs able to supporthe assessment of other chronic or degenerative diseases.

cknowledgments

he authors are deeply grateful to the Department of Bio-orphological and Functional Sciences of the University ofaples “Federico II” for providing them with the input data

et and to all the neuroradiologists cooperating both inhe definition of the domain knowledge and in the manualegmentation of the input dataset. Finally, a special acknowl-dgement is due to Dr. Bruno Alfano, Director of the Institute of

iostructure and Bioimaging (IBB) of Italian National Researchouncil (CNR), for allowing applying his multiparametricegmentation procedure in this work and for supporting pro-ciently such a research.

e f e r e n c e s

[1] A. Compston, A. Coles, Multiple sclerosis, Lancet 372 (2008)1502–1517.

[2] P. Kidd, Multiple sclerosis, an autoimmune inflammatorydisease: prospects for its integrative management, Altern.Med. Rev. 6 (2001) 540–566.

[3] D.H. Miller, R.I. Grossman, S.C. Reingold, H.F. McFarland, Therole of magnetic resonance techniques in understandingand managing multiple sclerosis, Brain 121 (1998) 3–24.

[4] M. Filippi, M.A. Horsfield, P.S. Tofts, F. Barkhof, A.J.Thompson, D.H. Miller, Quantitative assessment of MRIlesion load in monitoring the evolution of multiple sclerosis,Brain 118 (1995) 1601–1612.

[5] W.I. McDonald, A. Compston, G. Edan, D. Goodkin, H.P.Hartung, F.D. Lublin, H.F. McFarland, D.W. Paty, C.H. Polman,S.C. Reingold, M. Sandberg-Wollheim, W. Sibley, A.Thompson, S. van den Noort, B.Y. Weinshenker, J.S.Wolinsky, Recommended diagnostic criteria for multiplesclerosis: guidelines from the International Panel on thediagnosis of multiple sclerosis, Ann. Neurol. 50 (1) (2001)121–127.

[6] M. Filippi, M. Rovaris, A. Campi, C. Pereira, G. Comi,Semi-automated thresholding technique for measuringlesion volumes in multiple sclerosis: effects of the change ofthe threshold on the computed lesion loads, Acta Neurol.Scand. 93 (1996) 30–34.

[7] A. Akselrod-Ballin, M. Galun, J.M. Gomori, M. Filippi, P.Valsasina, R. Basri, A. Brandt, Automatic segmentation andclassification of multiple sclerosis in multichannel MRI, IEEETrans. Biomed. Eng. 56 (2004) 2461–2469.

[8] B. Alfano, A. Brunetti, M. Arpaia, A. Ciarmiello, E.M. Covelli,M. Salvatore, Multiparametric display of Shin-echo datafrom MR studies of brain, J. Magn. Reson. Imaging 5 (1995)217–225.

Page 10: An evolutionary-fuzzy DSS for assessing health status in multiple sclerosis disease

c a l i

e254 i n t e r n a t i o n a l j o u r n a l o f m e d i

[9] M. Wels, M. Huber, J. Hornegger, Fully automatedsegmentation of multiple sclerosis lesions in multispectralMRI, Pattern Recognit Image Anal 18 (2008) 347–350.

[10] F. Steimann, K.P. Adlassnig, Fuzzy medical diagnosis (2000).Available from: http://www.citeseer.nj.nec.com/160037.html.

[11] L. Zadeh, Fuzzysets, Inform. Control 8 (2000) 338–353.[12] J. Estevez, S. Alayon, L. Moreno, J. Sigut, R. Aguilar,

Cytological, Images analysis with a genetic fuzzy finite statemachine, J. Comput. Methods Programs Biomed. 80 (11)(2005) 3–15.

[13] B. Anuradha, V. Reddy, Cardiac arrhythmia classificationusing fuzzy classifiers, J. Theor. Appl. Inform. Technol. 4 (4)(2008) 353–359.

[14] S. Alayón, R. Robertson, S.K. Warfield, J. Ruiz-Alzola, A fuzzysystem for helping medical diagnosis of malformations ofcortical development, J. Biomed. Inform. 40 (3) (2007)221–235.

[15] K. Price, R. Storn, Differential evolution, Dr. Dobb’s J. 22 (4)(1997) 18–24.

[16] A.O. Boudraa, S.M. Dehak, Y.M. Zhu, C. Pachai, Y.G. Bao, J.Grimaud, Automated segmentation of multiple sclerosislesions in multispectral MR imaging using fuzzy clustering,Comput. Biol. Med. 30 (1) (2000) 23–40.

[17] H. Zhu, O. Basir, Automated brain tissue segmentation andMS lesion detection using fuzzy and evidential reasoning,

in: Proceedings of the 2003 10th IEEE InternationalConference on Electronics, Circuits and Systems, Sharjah,United Arab Emirates, December 14–17. IEEE 2003; vol. 3,2003, pp. 1070–1073.

n f o r m a t i c s 8 0 ( 2 0 1 1 ) e245–e254

[18] F. Russo, G. Ramponi, A fuzzy operator for the enhancementof blurred and noisy images, IEEE Trans. Image Process. 4(1995) 1169–1174.

[19] R. Gilesa, Lukasiewicz logic and fuzzy set theory, Int. J. ManMach. Stud. 8 (3) (1975) 313–327.

[20] A. Nobakhti, H. Wang, A simple self-adaptive differentialevolution algorithm with application on the ALSTOMgasifier, Appl. Soft Comput. 8 (2008)350–370.

[21] S. Das, A. Konar, U.K. Chakraborty, A. Abraham, Differentialevolution with a neighborhood-based mutation operator: acomparative study, IEEE Trans. Evol. Comput. 13 (3) (2009)526–553.

[22] A. Frank, A. Asuncion, UCI Machine Learning Repository,University of California, School of Information andComputer Science, Irvine, CA, 2010,http://archive.ics.uci.edu/ml.

[23] F. Al-Obeidat, N. Belacel, J.A. Carretero, P. Mahanti,Differential evolution for learning the classification methodPROAFTN, Knowledge-Based Syst. 23 (2010)418–426.

[24] M. Hall, E. Frank, G. Holmes, B. Pfahringer, P. Reutemann, I.H.Witten, The WEKA data mining software: an update,SIGKDD Explorations 11 (1) (2009)10–18.

[25] B. Alfano, A. Brunetti, M. Larobina, M. Quarantelli, E.

Tedeschi, A. Ciarmiello, E.M. Covelli, M. Salvatore,Automated segmentation and measurement of global whitematter lesion volume in patients with multiple sclerosis, J.Magn. Reson. Imaging 12 (2000) 799–807.