Methodology to Weight Evaluation Areas from Autism Spectrum Disorder ADOS-G Test with ...

18 Pages·2014·2.35 MB·English

ARTÍCULO DE INVESTIGACIÓN
Vol. 35, No. 3, Diciembre 2014, pp. 223-240
Methodology to Weight Evaluation Areas from Autism
Spectrum Disorder ADOS-G Test with Artificial
Neural Networks and Taguchi Method
M. Reyes ∗ ABSTRACT
P. Ponce ∗
Autism diagnosis requires validated diagnostic tools employed by
D. Grammatikou ∗ mentalhealthprofessionalswithexpertiseinautismspectrumdisorders.
A. Molina ∗ This conventionally requires lengthy information processing and
technical understanding of each of the areas evaluated in the tools.
Classifying the impact of these areas and proposing a system that can
∗ Tec de Monterrey, CCM. aidexpertsinthediagnosisisacomplextask. Thispaperpresentsthe
methodology used to find the most significant items from the ADOS-
G tool to detect Autism Spectrum Disorders through Feed-forward
ArtificialNeuralNetworkswithback-propagationtraining. Thenumber
of cases for the network training data was determined by using the
TaguchimethodwithOrthogonalArraysreducingthesamplesizefrom
531,441 to only 27. The trained network provides an accuracy of
100% with 11 different cases used only for validation, which provides
a specificity and sensitivity of 1. The network was used to classify the
12 items from the ADOS-G tool algorithm into three levels of impact
for Autism diagnosis: High, Medium and Low. It was found that the
items “Showing”, “Shared enjoyment in Interaction” and “Frequency
of vocalization directed to others”, are the areas of highest impact for
Autism diagnosis. The methodology here presented can be replicated
todifferentAutismdiagnosisteststoclassifytheirimpactareasaswell.
Keywords: Autism Spectrum Disorder (ASD), diagnosis, screening,
ADOS-G, Artificial Neural Networks, Feed-forward networks,
Taguchi Method, Orthogonal Arrays, classify.
224
RevistaMexicanadeIngenieríaBiomédica·volumen35·número3·Diciembre,2014
Correspondencia: RESUMEN
Mayra Reyes
El diagnóstico del autismo requiere del uso de herramientas de
Calle del Puente # 222, Col.
diagnóstico validadas internacionalmente que son utilizadas por los
Ejidos de Huipulco, Tlalpan,
profesionales de la salud expertos en trastornos del espectro autista,
C.P. 14380
lo cual requiere de procesamiento de mucha información y un
Correo electrónico:
entendimiento técnico de cada una de las áreas evaluadas en ellas.
[email protected] La clasificación del impacto que tienen cada una de estas áreas, así
como la propuesta de un sistema que pueda ayudar a los expertos en
el diagnóstico, es una tarea compleja, por lo que en este artículo se
Fecha de recepción:
presenta una metodología utilizada para encontrar los elementos más
30 de abril de 2014
significativos de la herramienta de diagnóstico de autismo ADOS-G a
Fecha de aceptación: través de redes neuronales artificiales entrenadas con retropropagación
29 de agosto de 2014 del error. El número de casos para entrenamiento de la red se
seleccionó utilizando el método de Taguchi con arreglos ortogonales,
reduciendo el tamaño de la muestra de 531,441 a solo 27 casos. La
red entrenada tiene una exactitud del 100% validada con 11 casos
diferentesdeniñosevaluadosparadiagnósticodetrastornodelespectro
autista con lo que se obtuvo una especificidad y sensibilidad de 1.
La red neuronal artificial se utilizó para clasificar los 12 elementos
del algoritmo de la herramienta ADOS-G en tres niveles de impacto:
Alto, Medio y Bajo. Se encontró que los elementos “Mostrar”, “Placer
compartido durante la interacción” y “Frecuencia de vocalizaciones
dirigidasaotros”sonlasáreasdemayorimpactoparaeldiagnósticode
autismo. Lametodologíapresentadapuedeserreplicadaparadiferentes
herramientas de diagnóstico de autismo para clasificar sus áreas de
mayor impacto también.
Palabras clave: Trastorno del Espectro Autista (TEA), diagnostico,
detección, ADOS-G, Redes neuronales artificiales, Método de
Taguchi, arreglos ortogonales, clasificación.
INTRODUCTION about 64% from 2006 to 2010 in the U.S. [3].
In Mexico there is not a national study that
can provide the Autism prevalence [4], but some
Autism Spectrum Disorder (ASD) is the group
nongovernmental associations estimate that 1 in
of developmental disorders whose clinical profile
300 children has been diagnosed with Autism
includesarangeofdisordersinsocialinteraction,
in Mexico [5]. The main characteristics of
communication, imagination and reduced and
ASD are disorders in social communication and
restricted behavior [1]. ASD is a world health
interaction such as lack of emotional reciprocity,
problem described for the first time in 1943 by
non verbal communication, development and
Kanner [2]. It usually begins during the first 24
management of relationships [6].
months of life; this period is defined as crucial
for the maturation of human neural circuits. As Although the causes of ASD remain
a result it affects, in varying degrees, normal unknown, all recent clinical data of
brain development in social and communication neuroanatomical, biochemical, neurophysiologic,
skills. According to the Centers for Disease genetic and immunological characters indicate
ControlandPrevention,1in68childrenhasbeen that autism is a neurodevelopmental disorder
diagnosed with ASD; this number has increased with a clear neurobiological basis. Currently
225
Reyeset al.MethodologytoweightevaluationareasfromADOS-GtestwithartificialneuralnetworksandTaguchimethod.
there is no biological test for the diagnosis of health specialists in areas such as Psychiatry or
autism. Diagnosis is achieved by behavioral Psychologywhocancarryoutaclinicaldiagnosis
evaluations specifically designed to identify based on the fifth edition of the Diagnostic and
and measure the presence and severity of the Statistical Manual also known as the DSM-V
disorder. The evaluations, made by trained [6] and the tenth revision of the International
and experienced health care professionals, are ClassificationofDiseasesalsoknownastheICD-
very important in order to assess strengths 10 [10] ; or even use screening and diagnostic
and weaknesses in the child and associated tools validated internationally. A summary of
developmental impairments. The diagnostic some of these tools is presented in Table 1.
criteria has been derived through consensus
among specialists and the diagnostic cut-offs
are hard to define. It is considered a spectrum THE AUTISM DIAGNOSTIC
because the core impairments in communication
OBSERVATIONAL SCHEDULE
and social interaction vary greatly.
GENERIC (ADOS -G) INSTRUMENT
In 2012, the Ministry of Health of
Mexico published the guide “Diagnosis and
The ADOS-G scale is a semi-structured
Treatment of Autism Spectrum Disorders” with
instrument based on observation that consists
recommendationsorientedtoearlydiagnosisand
of 4 modules that are managed in accordance
intervention algorithms, recognizing that timely
with the age and language skills of the child.
care is a crucial factor in order for these children
Faced with the challenge of characterizing or
to achieve the maximum functioning level
measuring symptoms and locate a patient at
and independence, and facilitate educational
a functioning level, the ADOS -G has the
planning,healthcareandfamilyassistance. The
advantage, with its variety of tasks, to make
manual includes the diagnostic procedure for
a diagnosis on observational basis. And through
the Autism Diagnostic Observational Schedule
the development of tasks it manages to make
Generic (ADOS-G) tool among others tools [7].
a representation of deficits and the level of
impairment of the patient. It usually takes
AUTISM SPECTRUM DISORDER between 30 to 60 minutes to be applied and
the test consists of activities performed by the
(ASD) DETECTION AND DIAGNOSIS
child in interaction of the expert who observes
According to the Clinical guide of Generalized him and assigns a grade [18].
Development disorders of the Infant Psychiatry Different modules and tasks of the test are
Hospital “Dr. Juan Navarro” in Mexico City mainly oriented towards evaluating the level of
[8], which was based on the multidisciplinary communication and specific behaviors in social
Consensus Panel described by Filipek et al. in interactions. Module 1 is used for toddlers that
1999 [9], ASD diagnosis can be separated in do not use language to communicate. Module
two levels. The first level corresponds to the 2 is used for children that communicate with
detection of development disorders by parents flexible phrases composed of 3 words. Module
or health professionals in the first contact clinic. 3 is used for children with fluid language and
It includes red flags for activities that the child Module 4 for adolescents and adults. The
had not developed at specific ages as well as objective of the instrument is not to evaluate
screening tools such as questionnaires. The knowledge abilities in the subject but rather to
second level corresponds to the evaluation and evaluate if the subject wants to participate in a
diagnostic of ASD that should be performed by social exchange [19].
226
RevistaMexicanadeIngenieríaBiomédica·volumen35·número3·Diciembre,2014
Table 1. ASD screening and diagnostic tools
Tool Type of tool Age range Advantages Disadvantages
Checklist for Autism Screening test 18 months Quick application Low detection
in Toddlers (CHAT) capacity
[11]
Modified Checklist Screening test 16-30months The predictive Low positive
for Autism in capacity increases predictivecapacity.
Toddlers Revised when used with Large number of
with Follow-Up a clinician’s false positives.
(MCHAT-R/F) [12] interview
Screening Tool for Screening test 12-23 months High sensibility Sensitivity and
Autism in Two-Year- Specificity based
Olds (STAT) [13] only on 12 cases.
Infant Toddler Screening test Less than 18 High sensibility Does not
Checklist (ITC) [14] months differentiate
between ASD
and any other
developmental
disorder.
Childhood Autism Diagnostic test Starting from Quantitative tool Can misdiagnose
Rating Scale 24 months that evaluates ASD in children
(CARS) [15] the severity of with intellectual
the symptoms. disabilities.
Also useful to
control evolution
of the patient after
treatment.
Autism Diagnostic Diagnostic test It can be used Direct observation Requires clinical
Observation for children of the child training and
Schedule - Generic over 2 years of interaction practice to observe
(ADOS-G) [16] mental age or through specified and evaluate.
in adults activities. Takes around 30
minutes to perform
the activities and
then some more
time to evaluate
the algorithm.
Autism Diagnostic Diagnostic test Starting from Interviewanswered Takes from 1 to
Interview-Revised 18 months by parents that 2 hours to apply
(ADI-R) [17] help distinguish because it has 93
ASD from other questions with
disorders. multiple options.
227
Reyeset al.MethodologytoweightevaluationareasfromADOS-GtestwithartificialneuralnetworksandTaguchimethod.
For this study only Module 1 was used, this sumsreachthethresholdorcutoff,thenthechild
module consists of 8 communication items, 12 can be diagnosed with Autism. The problem
social interaction items, 2 game quality items, 4 withthisevaluationisthatallareasareweighted
stereotypedbehaviorsitemsand3itemsforother equally;aslongasthesumsachievethesetpoints
abnormal behaviors. From a total of 29 items, Autism is diagnosed. For example it would be
the evaluation algorithm only takes into account the same for the ADOS-G algorithm to have a
12 items, 5 items that evaluate the child’s ability value of 2 (definitely abnormal) assigned to the
to communicate which are: how frequent the item “Pointing” and a value of 2 assigned to
child vocalizes directed to others (A2), if he/she the item “Gestures”, than having a value of 1,
uses words or phrases in a stereotyped way (A5), which means mildly abnormal, assigned to the
if he/she uses other people’s body as a tool four items “Frequency of vocalization directed
to communicate (A6), if he/she can point to to others”, “Stereotyped used of words”, “Use of
an object of interest (A7) and the emotional other’s body to communicate”, and “Pointing”.
gestures he/she normally employs (A8). The Since both sums (“2 +2” and “1+1+1+1”)
algorithm also counts the following 7 items to would be equal to 4, the threshold is met for
evaluate the child’s social interaction: if the the Communication area for either case. It is
child makes unusual eye contact (B1), if his/her clear that “definitely abnormal” in two areas
facial expressions directed to others attempt to is not exactly the same as “mildly abnormal”
communicate emotions (B3), if he/she enjoys in four areas since mildly abnormal could be
interaction with others (B5), if he/she shows easier to overcome than a definitely abnormal.
objects to others without asking for a specific Unfortunately this type of evaluation based on
need (B9), if the child wants to obtain attention sums is not focusing on the main aspects that
ofanadulttowardsobjectsthatnoneistouching determine Autism diagnosis, therefore there are
(B10), if he/she responds to the adults attention many aspects that are believed to be relevant
towards a specific unreachable object (B11) symptomsforAutismbuttherealimpactfactors
and finally the quality of social interaction have not been determined according to their
attempts(B12). severity or impact.
ADOS-G possible scores are 0, 1,2,3,7 and 8.
Zero means no evidence of abnormality related
ARTIFICIAL NEURAL NETWORKS
to autism, 1 means mildly abnormal, 2 means
definitely abnormal/severity varies, 3 means Artificial Neural Networks (ANN) are
markedly abnormal/interferes with interview, 7 computational models based on a simplified
means abnormal behavior not included from 1 version of biological neural networks with which
to 3 and 8 means that the behavior could not be they share some characteristics like adaptability
evaluated [16]. to learn, generalization, data organization and
The evaluation of the ADOS-G algorithm parallel processing. An ANN is composed of
consists of three sums. For autism cutoff, layers, one input layer, one output layer and
the sum of the five communication items one or several hidden layers. Inside each layer
(Frequency of vocalization directed to others, there are several neurons which are processing
Stereotyped used of words, Use of other´s body units that send information through weighted
tocommunicate,PointingandGestures)mustbe signals to each other and an activation function
greater or equal to 4; the sum of the seven Social determines the output as shown in Figure 1.
Interaction items (Unusual eye contact, Facial Weights have to be trained and many neurons
expression directed to others, Shared enjoyment canperformtheirtasksatthesametime(parallel
in interaction, Showing, Spontaneous initiation processing) [20].
of joint attention, Response to joint attention The input of a neuron would be the weighted
and Quality of social overture) must be greater sum of its entire input links plus a bias or offset.
or equal to 7 and the sum of all 12 areas should It will be activated only if the sum reaches the
be greater or equal to 12. Only when the three activationfunctionlevel. Theactivationfunction
5

Revista Mexicana de Ingeniería Biomédica

5

228 Revista Mexicana de Ingeniería Biomédica
RevistaMexicanadeIngenieríaBiomédica·volumen35·número3·Diciembre,2014

Figure 1. Multi-layered Artificial Neural Network

The input of a neuron would be the weighted sum of its entire input links plus a bias or offset. It
will be activated only if the sum reaches the activation function level. The activation function is a sigmoid
or “S” shaped function because it is bounded and always has a continuous derivative. ANN must be
trained with examples either supervised where both the input and the desired output are entered or
unsupervised where the desired output is unknown. The more examples it is trained with, the higher
precision should be achieved to solve new cases. ANN can be classified depending on their learning
process as presented in Figure 2. Learning can be supervised, where both inputs and desired outputs are

well known and the ANN must infer the input-output relationship. Unsupervised learning is used when
Figure 1o.nMly uthlet ii-nlpauytesr aerde kAnrotwifin cainadl tNhee uArNaNl oNrgeatwnizoersk by itself in clusters of patterns.
Figure 1. Multi-layered Artificial Neural Network

Unsupervised Kohonen Self Organized Maps
The input of a neuron would be the weighted sum of its entire input links plus a bias or offset. It
Only inputs are
will be activated only if the sum reaches thkneo awcnt. i vation function level. The activation function is a sigmoid
Adaptive Resonance Theory
or “S” shaped function because it isA NbNo uanred teradi naedn dto always has a continuous derivative. ANN must be
trained with examples either supervisseedlf -owrghaenrizee bino th the input and the desired output are entered or
Sparse Distributed Associative
clusters of patterns.
unsupervised where the desired output is unknown. The more exaMmemploersy it is trained with, the higher
ANN
precision should be achieved to solve new cases. ANN can be classified depending on their learning
learning
process as presented in tFypigesu re 2. Learning can be supervised, where both inputs and desired outputs are
well known and the ANN must infer theS uinpperuvti-soeudt put relationship. PUenrcseupptreornv ised learning is used when
Inputs and desired
only the inputs are known and the ANN organizes by itself in clusters of patterns.
outputs are known.
Back-propagation
ANN are trained to
infer the I-O
Implementing Genetic
relationship.
Algorithms in a network
Unsupervised Kohonen Self Organized Map s
OnFlyig uinrep u2.t sA NaNre l earning types [22], [23]
Figure 2. ANN learning types [21], [22].
known.
Adaptive Resonance Theory
is a sigmoid or “S” shaped fuAnNcNti oanre btreacianuesde to TherearetwomaintopologiesforANN:feed-
self-organize in
it is bounded and always has a continuous forwSaprdarsnee tDwisotrrkibsu,tewdh Aicshsoccoiantniveec t information
clusters of patterns.
derivative. ANN must be trained with examples only in one diMreecmtioonry from the input towards
ANN
either supervised where both the input and the output through any number of layers,
learning
the desired output are entered or unsupervised and recurrent networks, which contain feedback
types
where the desired output is unknown. The more information that may flow in both directions
Supervised Perceptron
examples it is trained with, the higher precision from input to output and vice versa where the
should be achieved to solve newInpcuatsse as.nd AdeNsNired weights can be automatically trained to obtain
outputs are known.
can be classified depending on their learning an optimaBl bacehka-pvrioorpoafgtahtieonne twork by reducing
ANN are trained to
process as presented in Figure 2. Learning can the error between the desired output and the
infer the I-O
be supervised, where both inputs and desired obtainedImouptlpeumtefnrotimng tGheenneettiwc ork, see Figure 3.
relationship.
outputs are well known and the ANN must infer Algorithms in a network
Back-propagation training method consists
the input-output relationship. Unsupervised
on minimizing the error with respect to the
learning is used when only the inputs are known
weights through gradient descent. The error is
Figure 2. ANN learning types [22], [23]
and the ANN organizes by itself in clusters of
thedifferencebetweenthedesiredoutputandthe
patterns.
real output delivered by the ANN. Minimizing
6

Revista Mexicana de Ingeniería Biomédica

There are two main topologies for ANN: feed-forward networks, which connect information only in
one direction from the input towards the output through any number of layers, and recurrent networks, 229
Reyeset al.MethodologytoweightevaluationareasfromADOS-GtestwithartificialneuralnetworksandTaguchimethod.
which contain feedback information that may flow in both directions from input to output and vice versa
where the weights can be automatically trained to obtain an optimal behavior of the neTthweosrku mbym reeddsuqcuinagre d error is then E given by
the error between the desired output and the obtained output from the network, see Figure 3.
E = XEp (5)
p
where Ep is the error on pattern p. Applying the
chain rule
δEp δEp δsp
= k (6)
δw δsp δw
jk k jk
Where
δsp
k = yp (7)
δw j
jk
Defining an update rule

Figure 3. Main Artificial Neural Network Topologies ∆ w = γδpyp (8)
p jk k j
Figure 3. Main Artificial Neural Network
Topologies. Where
δEp
Back-propagation training method consists on minimizing the error with respect to the weightsδ p = (9)
k δsp
through gradient descent. The etrhroer eirsr otrheis daicffheireevnecdeb byestewnedeinng thbea cdketshireeder roourtput and the real output k
delivered by the ANN. Minimizingf rthoem etrhroero iust apcuhtielavyeedr tboy aslelnthdeinign vboalcvke dthhei dedreronr fromA thpep loyuintpgutt hlaeycehra tion rule for the change in the
all the involved hidden layers whiclahy egrest aw hpircohpogretitonaal ppraorpt oorft itohnaat lerproarr,t eoafcht hnaeturon taerkreosr aa sprfoupnoctrtiioonnaolf the output and the change
part and retrains the weight for ereraocrh, eaincphunt eduurorinngta kthees anperxot peoprtoiocnha l[2p4a]r. tTahned activaintiothne fouuntcptiuotna sisa afu nction of the changes in the
differentiable function of the inputsre gtrivaeinns btyh e weight for each input during the input,
δEp δEpδyp
next epoch [𝑦2!3]=. 𝐹T(h𝑠e!)a ctiva tion fun ction is a δp =(−1) = − k (10)
! ! k δsp δyp δsp
Where differentiable function of the inputs given by k k k
𝑦!!  is  the  output  value  for  each  output  unit yP = F(sp) (1) Using the chain rule, when k is a hidden unit it
k k is called h
𝐹  is  the  activation  function
𝑠!!  is  the  input  of  the  𝑘  unit  in  patteWyrknPh  𝑝ies rethe output value for each output unit. δδEyhpp = oXN=o1 δδEspopδδyshpop = −oXN=o1δopwho (11)
F is the activation function
Where sp is the input of the k unit in pattern p. These yields to
k
Where
𝑠!!sp== X!𝑤!w"𝑦!!y+p+𝜃!θ ( 2) δhp =(2)F 0(sph)XNo δopwho (12)
𝑤  is  the  weighting  factor  for  input  j  and  output  k k jk j k o=1
!" j
𝜃  is  the  threshold Increasing the number of hidden neurons can
!
w istheweightingfactorforinputj andoutput
jk prevent from falling in a local minimum and
k.
Where diminish the error, but it might consist of a long
θ is the threshold.
k training process [23].
Where 𝑗 = layer   j (3)
j = layer(j) (3)
The error is defined as the quadratic error ($E^p$) at the output units for pattern $p$ between the desired
ARTIFICIAL NEURAL NETWORKS
The error is defined as the quadratic error (Ep)
output and the real output
at the output units for pattern p between the IN AUTISM
𝐸! = ! !! (𝑑! −𝑦!)! (4)
desired output and! th!!e!rea!l out!put
Feed-forward Networks have been used for a
𝑑!!  is  the  desired  Eoupt=pu1t  fXoNro  u(dnpit−  𝑜  yinp  )p2attern  𝑝 (4) great variety of medical applications such as
2 o o diagnosis of appendicitis, dementia, myocardial
o=1
The summed squared error is then E given by infarction, pulmonary embolism, back pain and
dp is the desired output for unit o in pattern p. skin disorders among others [24]. Autism
o
230
RevistaMexicanadeIngenieríaBiomédica·volumen35·número3·Diciembre,2014
diagnosis should be early, exact, cost effective parents to detect autism. The sample size was
and easy to use for health specialists so that of 966 individuals and an accuracy of 99.9% was
they can design the best intervention and offer achieved.
thechildmoreresourcestointegrateintosociety. Table 2 summarizes the current methods
Unfortunately this is not an easy task and for Autism diagnosis using Artificial Intelligence
requires plenty of knowledge and experience (AI). It can be observed that most of these
of the clinicians at first and second level of methodshaveusedalargesamplesizeinorderto
intervention. Artificial Neural Networks may train their models and none of them have tried
be able to provide the approach needed to to minimize the sample size. Although a large
detect Autism Spectrum Disorders (ASD) by sample used for training AI algorithms such as
identifying the highest impact factors that could ANN, usually provide better results, the quality
help detecting it at early stages of children’s of the samples for training data and possible
development. computational problems when training it due to
time consumption and machine resources used
Cohen&Sudhalter[25]usedArtificialNeural
must be taken into consideration too.
Networks (ANN) to create pattern recognition
in order to discriminate if a patient had autism
or mental retardation. After training the ANN
DESIGN OF EXPERIMENTS WITH
withinformationobtained frominterviewing 138
TAGUCHI METHOD
parents whose children had autism or mental
retardation, the system could predict 97% of
Designofexperiments(DOE)isthemethodology
testing cases for autism and 86% of cases with
that defines several conditions for an experiment
mental retardation. The general accuracy from
with multiple variables. The most common
the discriminant function analysis was 85%
technique is the factorial design which considers
increasing to 92% with the ANN. Veeraraghavan
all the possible combinations for the variables
& Srinivasan [26] created a Knowledge Based
and their states or levels [30]. The full factorial
Screener (KBS) and an intelligent trainer
design is given by
system that can detect different categories of
developmentaldisordersusingruledbasedexpert N = Lm (13)
systems with factual and heuristic information
through internet. Arthi & Tamilarasi [27] Where m is the number of factors and L is the
created a Neuro-Fuzzy system to diagnose number of levels for each factor or the possible
autism depending on few questions focused values each factor can have. For example, the
on the patient’s communication and linguistic ADOS-G tool algorithm is composed of 12 items
abilities. The questionnaire is answered by (A2, A5, A6, A7, A8, B1, B3, B5, B9, B10, B11,
the children’s parents. The system had B12) with 3 possible states for each item (0, 1
an overall performance of 85-90%. Wall, and 2). A complete factorial design would need
Kosmicki, DeLuca, Harstad & Fusaro [28] used N = 312= 531,441 possible cases to train the
an Alternating Decision Tree (ADT) to reduce network which would be costly, time consuming
from 29 to 8 items in the Autism Diagnostic and might generate an over fitted network. The
Observation Schedule-Generic (ADOS) in order goal is to look for a reliable number of cases that
to classify autism. The population used to train could be used to train an ANN to generate a
the system consisted of 612 individuals with reliable diagnosis of ASD [31].
autism and 15 individual without autism, 446 The Taguchi method proposes a fractional
cases were used to verify it reaching an accuracy factorial design based on orthogonal arrays
of 99.8%. Wall, Dally, Luyster, Jung & DeLuca (OA) which are tables of significant population
[29] used an Alternating Decision Tree (ADT) distribution. All the trials from the OA include
to decrease the number of items of the Autism all combinations with independent relationships
Diagnostic Interview-Revised tool (ADI-R) from among variables. It provides the least number of
93 to 7 questions which are asked to the child’s testcombinationsforasetofvariableslinear/non
231
Reyeset al.MethodologytoweightevaluationareasfromADOS-GtestwithartificialneuralnetworksandTaguchimethod.
linearanddependent/independenttoeachother. tool are the 12 parameters and since they have
This Orthogonal array is used as the selected 3 possible states, then the OA corresponds to
data to train the ANN. The selection of the OA the L orthogonal array which is presented in
27
is made depending on the number of parameters Table 3 and it contains the most representative
and the number of levels for the parameters. combinations for the 12 items at different levels.
In this case, the 12 items from the ADOS-G
Table 2. State of the Art of Artificial Intelligence methods for Autism diagnosis.
Tool name Reference Tool Advantage Disadvantage Minimization
description of
training
sample
Artificial Neural Cohen & ANN Increased Based on No
Network (ANN) Sudhalter. created to accuracy from the DSM-
trained with (1993) A discriminate discriminant III which
Backpropagation Neural between function evaluates
of the error Network Autism analysis with differently
Approach and Mental 85% to 92% Autism
to the retardation with ANN. compared
Classification based on to the most
of Autism [25] the Autism current
Behavior version DSM-
Interview V. Selection
(ABI) of 11 from 28
questions of
the ABI tool
which is not
validated as a
gold standard
for Autism
diagnosis.
138 samples
needed.
Knowledge Veeraraghavan Ruled based Available Does not No
Based Screener & Srinivasan expert system through mention if the
(KBS) (2007). with factual internet knowledge is
Exploration and heuristic obtained
of Autism knowledge from a
using Expert to analyze standardized
Systems [26] children tool or only
development from clinical
and identify experience.
developmental Does not
disorders. mention if it
was tested.
232
RevistaMexicanadeIngenieríaBiomédica·volumen35·número3·Diciembre,2014
Neuro fuzzy Arthi & Neurofuzzy Helps Not based No
system Tamilarasi system converts diagnosing on a certified
(2008). inputs from a autism with test, depends
Prediction parent answered an overall on the
of autistic questionnaire performance expertise of
disorder using which is of 85-90% the clinicians
neuro- fuzzy converted to that help
system by fuzzy membership to construct
applyingANN values. Those the system.
technique [27] values are Started with
evaluated with 40 samples
if-then rules and and needed
the fuzzy output to increase
becomes the input to 194 to
for the artificial increase
neural network training
trained with performance.
backpropagation
method.
Alternating Wall, ADTree classifier Reduction Large sample No
Decision tree Kosmicki, consisting of 8 from 29 needed
(ADT) DeLuca, questions from the to 8 items for system
Harstad ADOS Module1 to classify training (623
& Fusaro tool. autism individuals)
(2012). Use with 99.8%
of machine accuracy with
learning False positive
to shorten rate of 0 and
observation- True positive
based rate of 1
screening
and diagnosis
of autism [28]
Alternating Wall, Dally, Decision tree Reduction Large sample No
Decision tree Luyster, classifier to from 93 to needed
(ADT) Jung, DeLuca detect autism 7 questions for system
(2012). Use rapidly through 7 to classify training (966
of Artificial questions from the autism individuals)
Intelligence to ADI-R tool with 99.9%
shorten the accuracy with
behavioral False positive
Diagnosis of rate of .013
Autism [29] and True
positive rate
of 1