A MULTI-CRITERIA SORTING APPROACH FOR DIAGNOSING

MENTAL DISABILITIES

Paulo Freitas

, Carlos Henggeler Antunes

and Jorge Dias

Institute of Systems and Robotics, University of Coimbra, Coimbra, Portugal

INESC Coimbra, University of Coimbra, Coimbra, Portugal

Keywords: Multi-criteria models, Outranking methods, ELECTRE TRI, Mental disabilities, Alzheimer disease.

Abstract: A multi-criteria model tackled by an outranking method devoted to the sorting problem is presented to

support decision making in assessing individual mental disabilities using information required in the

Clinical Dementia Rating scale. This diagnosis process is a critical factor for adapting treatments to the

current stage of the disease and improving health care and quality of life. The criteria required in the

Clinical Dementia Rating scale have been considered as an input for developing our multi-criteria model,

the output of which is the classification of each individual under evaluation in a pre-defined ordered class

(category) as an indicator of the revealed level of mental disabilities. A method based on the exploitation of

an outranking relation for the sorting problem is used to compare the individual information according to

multiple evaluation criteria with reference profiles (specified standards) that define the boundaries of the

classes. This methodological approach is substantially different from the ones based on the aggregation of

the different criteria using weighted-sums to produce a “common value” measure. The method requires

meaningful technical parameters, such as weights (herein perceived as true importance coefficients of the

multiple evaluation aspects), distinct thresholds to ascertain the outranking classification, and a cutting level

establishing the exigency of the classification. A realistic example using the decision support system Iris is

presented to illustrate the results.

1 INTRODUCTION

Decision making processes are daily tasks associated

with several contexts in people's life. In health care

decisions, inadequate evaluations may lead to bad

judgments and consequently result in inappropriate

treatments and negative health effects. Therefore,

sound models and methodologies shall be developed

to support making the best decisions when handling

with situations concerning people's health care in

face of multiple, often conflicting, evaluation

aspects. A specific area that requires feasible and

reliable diagnosis is associated with dementia

assessment and treatments. In these cases the

diagnosis decision support process is generally

divided in several stages and is based on multiple

criteria to reach a comprehensive evaluation. This

type of decisions is increasingly important due to the

growth of life expectancy, which is accompanied by

an increasing prevalence of health impairments and

mental-health problems such as dementia (Hendrie,

1998). Early and accurate identification of

individuals who are at a high risk of developing

dementia is regarded as a research priority. This

identification followed by effective interventions

may significantly contribute to reducing the

prevalence and incidence of dementia diseases,

improving the quality of life both of the patients and

their caregivers, and making a more efficient use of

the resources needed to provide adequate

institutional and home health care. The process of

early identification assumes even greater importance

knowing that there are already treatments to help

slowing the disease progression and prevention

strategies including lifestyle changes (Roberson and

Mucke, 2006). The DESCRIPA Study (Vissera et

al., 2008) presents an evaluation over three years of

a set of clinical criteria for further analysis of which

variables best predict dementia, in particular

Alzheimer’s disease. Functional impairment in

people in risk of dementia has been studied to

understand what are the indicators associated with

the disease’s progress (Wilkins et al., 2007).

392

Freitas P., Henggeler Antunes C. and Dias J..

A MULTI-CRITERIA SORTING APPROACH FOR DIAGNOSING MENTAL DISABILITIES.

DOI: 10.5220/0003752303920398

In Proceedings of the 1st International Conference on Operations Research and Enterprise Systems (ICORES-2012), pages 392-398

ISBN: 978-989-8425-97-3

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

The aim of this work is to develop a multi-

criteria model and use a multi-criteria method based

on an outranking relation to provide decision support

in the diagnosis of dementia related diseases of

individuals according to the assessment of their

mental status. Taking as a basis the Clinical

Dementia Rating (CDR) scale, a tree of criteria has

been developed to encompass all the relevant aspects

for a comprehensive assessment. A multi-criteria

method devoted to the sorting problem and a

computer package for helping the analysis have been

selected. This decision aid approach expects to

receive data (the performance of each individual

according to each criterion) from an external

assessment system. Other inputs include the

technical parameters required by the methodological

component. The output consists in the assignment of

each individual to a pre-defined ordered category

associated with the physical/mental status of the

individual.

Furthermore, this enables to carry out a long

term analysis of the individual historical data, which

may establish a correlation between the model's

output and the actual situation. Our approach is

aimed at providing a tool for helping technical staff

in charge of diagnosing dementia diseases to support

making the best decisions, increasing the accuracy

and reliability of the evaluation.

The assessment of dementia diseases has been

traditionally made using the aggregation of different

perspectives of evaluation (criteria) by means of

some type of scoring and weighted-sum approaches

thus transforming the performances of the entities

under evaluation according to the different criteria

into a “common value” score (Robert et al., 2010).

We believe that in most cases, as in the one under

study herein, it is sufficient for analysis and provides

more confidence on the results the assigning of the

entities under evaluation (individuals) to pre-defined

ordered categories of merit rather than producing a

single numerical figure. Furthermore, a more

detailed analysis within each category is possible

whenever it is considered useful to improve the

discrimination of the evaluation model.

In this setting it is important to appraise the

entities using known standards or profiles.

Moreover, it is also convenient to evaluate entities

on an “as they come” basis. This capability of

evaluating each individual in absolute terms

according to reference profiles, and not just in

comparison with their peers, as well as the need to

include evaluation aspects expressed in different

units using different types of scales (also

qualitative), can be accomplished using the

ELECTRE TRI method (Mousseau, Slowinski,

Zielniewicz, 2000). The ELECTRE TRI method is

adequate for our assessment problem because it does

not require controversial (scale-dependent) weight

specification in order to obtain a single score for

each individual and allows the definition of standard

profiles (establishing the frontiers between the

categories) with which each individual is compared.

Those profiles may be updated for further

adjustments as required by distinct practical

situations. This type of evaluation model can bring

advantages to support medical staff in the

assessment of dementia diseases in comparison with

some traditional processes of applying pre-defined

scoring scales.

Other approaches based on multi-criteria models

and methods have been proposed in the literature,

specifically to assist the process of diagnosing

Alzheimer's disease. The work presented in (Castro,

Pinheiro, Pinheiro, Tamanini, 2011) proposes a

hybrid model combining influence diagrams and

multi-criteria methods to compare the values for

each entity in a set and then perform a rank within

the group. Bayesian Networks are used in (Pinheiro,

Castro, Pinheiro, 2008) to serve as a modelling tool

for aiding in decision making for the diagnosis of

Alzheimer’s disease. In (Castro et al., 2011; Pinheiro

et al., 2008; Filho, Pinheiro, Coelho, Costa, 2009)

good overviews about related works for this area are

presented, proposing approaches for decision aiding

models applied to medical activities. Moreover, the

performance achieved applying methods based on

ELECTRE IV and a genetic algorithm is presented

in (Filho, Pinheiro, Coelho, 2009; Filho, Pinheiro,

Coelho, Costa, 2010). Experiments have been made

(Costa, Filho, Coelho, Pinheiro, 2009; Filho et al.,

2009) to compare the accuracy and effectiveness of

different multi-criteria decision aid methods with

different data sets, the conclusions pointing out to

some changes in the results for the same models.

The paper is organized as follows. Section 1

provides the interest and motivation of this study.

Section 2 presents a brief overview about multi-

criteria outranking methods and the software

package used in this work to accomplish the goals

previously defined. In section 3 the multi-criteria

model and the design of the overall approach are

presented. Section 4 describes some illustrative

results obtained using a case study. Finally,

conclusions are drawn and future work is outlined in

section 5.

A MULTI-CRITERIA SORTING APPROACH FOR DIAGNOSING MENTAL DISABILITIES

393

2 THE ELECTRE TRI METHOD

The ELECTRE TRI method is a member of the

ELECTRE (Elimination and Choice Translating

Reality) family of multi-criteria methods (Roy,

1996). ELECTRE methods are based on the

construction and exploitation of an outranking

relation (“outranking” having the meaning of “is at

least as good as”). ELECTRE TRI is devoted to the

sorting (classification) problem, which consists in

assigning each entity under evaluation to one of a

pre-defined set of ordered categories (C

,…,C

according to several evaluation criteria g

(j=1,…,n).

Each entity object of evaluation (individual) is

described through a vector of multi-criteria

performances. The categories are defined by

specifying reference profile vectors (b

,…,b

), being

each reference profile b

(h=1,…,k-1) the upper

bound of category C

and the lower bound of

category C

h+1

The assignment of each entity a

to a category C

is done by comparing its value in each criterion to

the corresponding reference profiles. The method

assigns each entity to the highest category such that

its lower bound is outranked by a

. The outranking

relation is verified by comparing a credibility index,

computed by using the differences in performance

and the criterion weights, with a cutting level

(

∈

[0.5,1]), which defines the “majority requirement”

and consequently the exigency of the classification.

In ELECTRE methods weights shall be

perceived as true coefficients of importance of the

criteria (their “voting power”), which are scale-

independent, i.e. not linked to the scales in which

each criterion is measured. This is a totally different

interpretation of weights with respect to weights

used as technical parameters for translating the

scores in each criterion into a global score. For

further details about ELECTRE TRI see (Mousseau

et al., 2000).

The ELECTRE TRI method requires the

specification of a set of technical parameters (which

convey meaningful preference information): the

reference profiles defining the categories (b

,…,b

the criterion weights w

, the cutting level (

), a set of

indifference (q

), preference (p

) and veto (v

)

thresholds for each criterion g

and reference

profiles. Indifference and preference thresholds

characterize the acceptance of imprecision in the

judgment by considering two entities as indifferent

when their individual performances in each criterion

differ less than a specified amount q

. The

transition from indifference to preference is not

sharp but changes linearly from q

to p

, this being

captured by the criterion concordance index c

Figure 1 illustrates the computation of the criterion

concordance index when comparing alternative a,

and the profile b

for criterion g

Figure 1: Criterion concordance index.

The veto thresholds are quite relevant in our case

study because they capture “non-compensatory”

situations in which a very bad performance in any

criterion prevents an entity of being classified in the

best category or even force it to be classified in the

worst category independently of having very good

performances in all other criteria. In general,

practical evaluation models require the consideration

of a certain level of non-compensation, at least for

some criteria.

The assignment classification provided as result

of ELECTRE TRI is the desired indicator of the

disease progress level for each individual. In this

paper, we assume that the input data (the

performance of each individual in any criterion

expressed in a quantitative or qualitative scale) must

be supplied by an external system and/or medical

inputs. It is advisable that an expert panel, for

instance a medical board, could supply the technical

parameters (reference profiles, weights and

thresholds) referred to above. The overall decision

support approach will be discussed in detail in

section 3.

3 AN APPROACH BASED ON AN

OUTRANKING METHOD

Multi-criteria analysis methods are largely

unexploited so far on health care, particularly in

classification processes using medical information.

In our work, a multi-criteria approach using the

ELECTRE TRI method is proposed to deal with a

model that can use behavioural and cognitive data to

infer about the mental and physical state of

individuals.

The keystone of this study has been the

development of a comprehensive multi-criteria

model encompassing all the fundamental axis of

evaluation. The structuring phase is an essential step

to reach a stable multi-criteria evaluation model

from a generally “messy situation” through a process

of unveiling and refinement procedure of a

ICORES 2012 - 1st International Conference on Operations Research and Enterprise Systems

394

consistent family of criteria. The fundamental

criteria that we have selected (tree of criteria) to

assess any individual are depicted in Figure 2, in

which the operational criteria are inside the boxes.

The fundamental criteria, which are dealt with by the

method, are derived from a set of sub-criteria

identified in the structuring phase. The score of any

individual in each fundamental criterion results from

a weighted aggregation of the scores in the

corresponding sub-criteria. Typically, each sub-

criterion is measured in a qualitative scale according

to the frequency of occurrences using three levels

with the associated meaning of: “it never happens”,

“it happens sometimes”, and “it happens very often”.

An external system is responsible for providing

the performance of each individual according to

each evaluation criterion in a given measurement

scale. A group of experts, or medical board, provides

the technical parameters required by the method.

This can be done in an iterative way for model

calibration purposes.

The establishment of a pattern of behaviours that

is indicative of the disease status, which can be

parameterized, conveys the information to define the

reference profiles that define the boundary of the

classes. These profiles have been specified by means

of scales used in Neuroscience, Psychology, and

other related areas, to select which type of

behaviours should be used in the evaluation of the

mental state of the individual. The CDR scale is one

of the most well-known scales used in the process of

Alzheimer’s disease diagnosis. This type of scales is

applied as questionnaires, which raise the

subjectivity issue when the patient or the caregiver is

answering to them. The subjectivity inherent to

answering the questionnaires is somehow mitigated

by means of the use of the indifference, preference

and veto thresholds in the operational framework of

ELECTRE TRI.

Four categories (classes) have been identified to

which any individual will be assigned based on the

information about his/her mental and physical

performances:

• Urgent (earlier medical intervention is

required).

• Disturbing (medical accompanying is needed).

• Mild Impairment (attention shall be paid to

evolution).

• Normal (no need to be followed on a regular

basis).

Figure 2: Criteria tree.

Figure 3: Categories and reference profiles.

If judged necessary more classes can be

considered to increase discrimination between

statuses. This would imply to define more technical

parameters associated with new reference profiles.

Figure 3 outlines graphically the categories

bounded by the reference profiles, which each

individual is compared with. The reference profiles

are displayed in Figure 3 just for illustrative

purposes of how categories are defined and they do

not represent the actual values used in the

experiments whose results are presented in section 4.

Table 1 presents each reference profile (b

, b

and b

) and the corresponding criterion performance

(by this order in each criteria), which may be

considered as “reference individuals” establishing

the frontier between the categories.

A MULTI-CRITERIA SORTING APPROACH FOR DIAGNOSING MENTAL DISABILITIES

395

Table 1: Criteria and Reference Profiles.

Criteria Reference Profiles (b

, b

)

Temporal

Orientation

- Do not know the current day

- Often does not know the day of the

week

- Do not know the current day of the

week

- Often does not know the month

- Does not know the month

- Often does not know the year

Space

Orientation

- Often gets lost away from home

- Lose yourself away from home

- Often lost in familiar locations

- Often confused where currently is

- Confused where currently is

- Often gets lost inside of the house

Financial

Management

- Cannot give / make accounts of large

amounts

- Cannot give / make accounts of small

amounts

- Lose track of the money

Social

Activity

- Often does not drive car

- Left driving car

- Decrease in activities outside home

- Lack of activities outside home

Similarities

and

differences

- Slight decrease in the level of

abstraction

- Cannot get basic abstraction

- Often does not know

Memory

- Failure at the address details

- Cannot say all the words

- Often misses the address

Physical

Activity

- Decrease in motor activity

- Sedentary lifestyle

- Begins to detect hand tremor

Knowing

personal data

- Does not know the birthplace

- Does not know the name and location of

the last school

- Does not know / incomplete date of

birth

- Does not know his education level

- Does not know the last job

- Forgot the main profession

Sleep Analysis

- Issues falling asleep

- Waking up too early

- Sleeping excessively during the day

- Getting up in the night

In order to classify the validity of the outranking

relation between the individual and the reference

profiles is assessed, thus determining the assignment

of the individual to one of the categories. In Table 1

the reference profiles are presented in terms of the

meaning associated with the scores in each criterion

and not actual values, using the CDR scale as a

guideline.

4 SOME ILLUSTRATIVE

RESULTS

In order to test our model, data of 20 individuals

have been used associated with persons in various

stages. We have used the Iris software (Dias and

Mousseau, 2003) to implement our method and to

test it in distinct scenarios. The data contain

information about selected individuals with a very

good mental state, others displaying intermediate

disease indicators, and others with a bad diagnosis.

That is, the entities under evaluation have been

selected to span a wide set of conditions to illustrate

the operation of the multi-criteria model coupled

with the Iris package implementing a version of

ELECTRE TRI.

Initially, all the technical parameters required by

the method (see sections 2 and 3) have been

specified: definition of the categories in which the

individuals will be classified; preference,

indifference and veto thresholds for each criterion

and reference profile; criterion weights; cutting

level. Although this specification process may seem

to impose a significant burden on the decision

makers, these parameters are essential to bear their

experience and insightful information into the

sorting procedure. Usually these parameters are

elicited from decision makers with the aid of an

analyst with expertise on the methodological

component to ease the elicitation process. Moreover,

some of these parameters can be preset (according to

experience in previous studies). For instance,

indifference and preference thresholds may be

established as percentages (e.g., 2% and 10%,

respectively) of the value ranges in each class.

Figure 4 presents the results obtained using the

Iris package for the cutting level λ=0.5. This means

that a “simple majority” of criteria supporting the

outranking relation is required. The left column

displays the 20 entities under evaluation (individual

0-19) and each column C1–C4 is associated with a

category (ordered from the worst to the best one):

C1 – Urgent; C2 – Disturbing; C3 – Mild

Impairment; C4 – Normal. The partially coloured

matrix displays the assignment of each entity to a

category. The darker cells represent the assignment

proposed by Iris for each individual, which is

associated with a central combination of parameters.

The lighter cells represent the other possible

assignments, which are obtained for other feasible

combinations of parameters under certain constraints

(in this case, criterion weights that may vary within

intervals).

ICORES 2012 - 1st International Conference on Operations Research and Enterprise Systems

396

Figure 4: Assignments with λ=0.5.

Analyzing Figure 4, we can conclude, for

example, that individual 0 is restricted to category

C2 – Disturbing (all sets of weights lead to this

result). Individual 1 is assigned to category C3 –

Mild Impairment according to the central parameters

but he/she can also be sorted into category C2 (since

this is feasible for other parameter combinations).

Figure 5 displays the results obtained when using

a cutting level λ=0.85, that is, increasing

significantly the exigency of the classification

derived from the outranking relation verification.

Therefore, the assignments proposed by Iris for each

entity are now “less favourable”. For example,

individual 14 was previously classified in C2, using

central parameters, and after the increase of λ his/her

assignment is restricted to C1.

More experiments have been done to analyze the

effects of changing the criterion weights for testing

scenarios in which the contribution for the

classification is not equal for all the criteria. That is,

we assume that there are criteria more important

than others, so they are given a higher “voting

power” in ascertaining the outranking relation.

Figure 6 illustrates the results obtained using a

cutting level λ=0.5 and different criterion weights

reflecting the following importance rank (in

decreasing order): “Space Orientation” and “Sleep

Analysis”; “Financial Management”; “Temporal

Orientation” and “Social Activities”; all the other

criteria. Comparing the results obtained in Figure 4

and Figure 6 we conclude that individuals 2 and 13

can now attain classes C1 and C2, respectively, and

individual 8 can just be assigned to class C2, thus

displaying the impact that criterion weight changes

can have on the final classification.

Figure 5: Assignments with λ=0.85.

Figure 6: Assignments with λ=0.5 and with different

weights.

The interactive environment provided by Iris

may be used, for instance, to assess the trend of the

progression of the disease for a given individual.

The same individual in different points in time may

be considered as different entities under evaluation

and the relative assignment is easily assessed.

Lighter coloured cells are also indicative that

individuals may be close to change to the next

categories.

5 CONCLUSIONS

This paper presents a multi-criteria model to provide

A MULTI-CRITERIA SORTING APPROACH FOR DIAGNOSING MENTAL DISABILITIES

397

decision support in the diagnosis process of

dementia cases. The model is tackled using a

methodology based on an outranking relation, which

is exploited for a sorting problem in which

individuals are assigned to categories associated

with the perceived status of the disease. Categories

are defined using references profiles with which

individuals are compared to check the outranking

relation.

This approach offers the possibility to adjust the

reference profiles as well as other technical

parameters required by the method to better suit the

different usage scenarios. The use of ELECTRE

TRI’s technical parameters may help to cope with

subjectivity issues that are present in the traditional

processes of applying pre-defined scales. This multi-

criteria approach offers a flexible methodology

capable of being adjusted according to the objectives

of medical staff in the assessment of dementia

diseases.

Future developments include using previous

examples of classified individuals to infer new

reference profiles and classification assignments,

analyze thoroughly the reliability of considering

“non-central” classifications in the output results as

tendency indicators, perform tests with other multi-

criteria decision aid methodologies to conclude

about the reliability achieved when using different

approaches.

REFERENCES

Castro, A., Pinheiro, P., Pinheiro, M., Tamanini, I. (2011).

Towards the Applied Hybrid Model in Decision

Making: A Neuropsychological Diagnosis of

Alzheimer's Disease Study Case. International

Journal of Computational Intelligence Systems, 4, (1),

652-656.

Costa, N., Filho, A., Coelho, A., Pinheiro, P. (2009).

Selecting prototypes for two multicriteria

classification methods: A comparative study. World

Congress on Nature & Biologically Inspired

Computing. 1702 – 1707.

Dias, L., Mousseau, V. (2003). IRIS – Interactive

Robustness Analysis and Parameters Interference for

Multicriteria Sorting Problems (Version 2.0). User

Manual, Document 1/2003 INESC Coimbra.

http://www.inescc.pt/documentos/DocInt_1_2003.pdf

Dias, L., Mousseau, V.( 2003). IRIS: A DSS for Multiple

Criteria Sorting Problems. Journal of Multi-Criteria

Decision Analysis,12, 285-298.

Filho, A., Pinheiro, P., Coelho, A. (2009). Towards the

Early Diagnosis of Alzheimer’s Disease via a

Multicriteria Classification Model. Proceedings of the

5th International Conference on Evolutionary Multi-

Criterion Optimization. Springer, 393-406.

Filho, A., Pinheiro, P., Coelho, A., Costa, N. (2009).

Comparison of two prototype-based multicriteria

classification methods. IEEE Symposium on

Computational Intelligence in Multi-Criteria

Decision-Making, 133 – 140.

Filho, A., Pinheiro, P., Coelho, A., Costa, N. (2010).

Comparison of Two MCDA Classification Methods

over the Diagnosis of Alzheimer's Disease.

Proceedings of the 4th International Conference on

Rough Sets and Knowledge Technology. Springer,

334-341.

Hendrie, H. (1998). Epidemiology of dementia and

Alzheimer’s disease. American Journal Geriatric

Psychiatry, 6, S3–S18.

Mousseau, V., Slowinski, R., Zielniewicz, P. (2000). A

user-oriented implementation of the ELECTRE-TRI

method integrating preference elicitation support.

Computers & Operations Research, 27, 757-777.

Pinheiro, P., Castro, A., Pinheiro, M. (2008). A

Multicriteria Model Applied in the Diagnosis of

Alzheimer's Disease: A Bayesian Network. IEEE

International Conference on Computational Science

and Engineering, 15–22.

Roberson, E., Mucke, L., 2006. 100 Years and Counting:

Prospects for Defeating Alzheimer’s Disease. Science,

314, 781-784.

Robert, P., Ferris, S., Gauthier, S., Ihl, R., Winblad, B.,

Tennigkeit, F. (2010). Review of Alzheimer’s disease

scales: is there a need for a new multi-domain scale for

therapy evaluation in medical practice?. Alzheimer’s

Research & Therapy, 2(24).

Roy, B. (1996). Multicriteria Methodology for Decision

Analysis. In Kluwer Academic Publishers, Dordrecht.

Vissera, P., et al. and DESCRIPA study group. (2008).

Development of Screening Guidelines and Clinical

Criteria for Predementia Alzheimer’s Disease.

Neuroepidemiology, 30 (4), 254-265.

Wilkins, C., Wilkins, K., Meisel, M., Depke, M.,

Williams, J., Edwards, D. (2007). Dementia

Undiagnosed in Poor Older Adults with Functional

Impairment. The American Geriatrics Society, 55 (11),

1771–1776.

ICORES 2012 - 1st International Conference on Operations Research and Enterprise Systems

398