Classification of Childhood Disfluencies Using Neural Networks

ffiELSEVIER

._--=-_II

f

CLASSIFICATION OF CHILDF{OODDISFLUEI{CIES USINGNEURAL T{ETWORKSY.V. GEETHAAll India Institute of Speech & Hearing, Mysore, India

KARANTH PRATIBHAInstitute of Speech & Hearing, Bangalore, India

RAO ASHOKCEDT, Indian Institute of Science, BangaLore, India

SHETTY K. RAVINDRALG Sofi lndia, Bangalore, lndia

Differentiating normai chilChood disf-luencies from stuttering has posed a great diagnosticchallenge to the speech and language pathologists for decades. The paper discusses an ob-jective way of making rhis differentiation. Artif icial Neural Network (ANN). a computerprogram, was used for this purpose on two groups of disfluent children below the age of 6years. Data from Group I (comprising 25 disfluent chiidren) on different variables were

used to train the ANN, whereas the data from Group II .(comprising 26 disfluent children)on the same set of variables were used for predicting the diagnoses. ANN could predict thesubclassifications of normal nonfluency and stuttering with 92Eo accuracy although the sizeof the training sample was comparatively small. The applications of ANN in general arehighlighted along with the procedural considerations for adopting ANN for the current ob-jective of classification. ANN appears to be a useful clinical tool for objectifying our diag-nostic procedures, with.regard to stutlering. O 2000 Elsevier Science Inc.

Key Words: ,Neural network: Normal nonfluency; Disfluency, Classification

INTRODUCTION

The differentiation between disfluencies of preschool children identified asstutterers and those regarded as normal speakers has been an issue of both the-oretical and practical concern for a long time. According to Dalton (1983), thedifferential diagnosis of childhood disfluencies is problematic, with. childrenmanifesting a variety of normal and abnormal behaviors. The symptomatol-.

Address correspondence to Y.V. Ceetha, 4-06-02, AIISH Layout, Bogadi II stage, Manasagango-thri P.O., Mysore-570 006, India

J. FLUENCY DISORD. 25 (2OOO),99_117

O 2000 Elsevier Science Inc. Ail rights reserved.655 Avenue of the Amer icas. New York, NY 10010

0094-730x/00/$-see front matterPrr s0094-730x(99)00029-7

100

ogyandphenomenologyofstut ter ing.atonset isnotgnlysaidtodi f fer f romthat of the typicbl uo"ri *rr" has the disorder, but also from that of the typical

,-.ilot-ug" "ttita

*fto has it (silverrnan' 1992)'

Sutteringisaheterogenouscondit ionknownforitsinterandintraindivid-ual variability' Earlf iti"t'ing is episodic in natu-reand do3s

"9t Tt{:::'::'

all situation, uno ur'utirr*.r. itr"re will be periodic fluctuationlt" ].t:j:l::::

and symptomatorog|. A variety of factors have been suggested by vanous au-

rhors to distinguisti "or*"t nonflu.ncy (NNF) and stuttering in children'

These include repetition characteristics (type, frequency, aud number of unit

repetitions)' p'ofonguiion'' t'"sit^tl"::' 'iui" of speech' presence of schwa

vowel, airflow interruptions. inappropriate articulatory postures' awareness of

t h e p r o b i e m , a n d u u o i a u n . . b e l l a v i o r s ( J o h n s o n ' Y o u n g . S a c h s ' & B e d e l i '1959; Win gur", tie:i, va'.' nlpe'' 1q7i; Bloodsi-ein ' 1914' Yairi'and !:Y1:1gg4; Silverman, tggizl.The crireria-that distir:g*ish the normal Cisfl*enctes

and stuttering, iro*.uJr, ,..-"o, uniiorn"l across -"iudies and there are several

over lappings.Thestudiesdonotspeci ty . theexact la tureois tut ter ingdis f lu .encies and how it qualitatively or quantiiutiu"ty iiifers from normal childhood

disfluencier. H""J.T;;;"; difficult to differentiate the t"r'o conditicns' es-

pecialiy in the borderiine or very mild disfluent conditicns'

AccordingtoCurlee(i993),speec}rofnran1.r., : :eschociers.i : :stbegi*-ningtosrutter often sounds very much like $iat cf cthlr ciriicien of tir"*ir ;ige *r*ch of

the rime. rnu,, ,tri-t".i"g i, episodic.and varies -sr:bsrantiallv in frequency' anci

severity fronr da'v to cial'atld one ''*uttnn ti::i-rriilli:r' ijc"i' ' li 'any' si*r:s of

Stuttering maY be observe. ciuring a clinicai evelr:ai..o:1, .I'jlu:.. i.-ier.:ii:;i*'= stut.

t e r i ngamong thesech i l d renCan .poseas ign i f i can td iagn r : s l i ccha } lenge to

"u"ri"^p"rienced, expert cli nicians'

ln the indian .onrJut also, this probiem in the differentiai ii'''g*csl:' ui NNF

and stuttering was hig1,ugnteo ly a pitol st':dv u*derraken by :i;e f:Est a:'ii.hor'

In this pilot stuoy, .ui" ril analysis of 64 disfluent children tlel{:r*" the ag* of 6

yea rs reg i s te reda t theA l l l nd ia ins t i t u teo fSpeechandHear ing (A I iSF I ) ,Mysore, during lggZ *u'undertaken * t93i'ifV'in-:

::i:1iot"s crucial to the cli-

niciantomakethedist inctionbetweenNNFandstuttering.Thevar. iablessuchaSage'sex,familyhistoryofstuttering,speechandnonspeechcharactel ist icsof disfluen.i.' (;;p;' i"qu"nt'' uno #uion of disfluencies' presence of sec-

ondaries and avoidance), and rat.e of speech \A,ere co.mpared between the chil-

drendiagnosedasstuttering.asaguinstNNF.Itwasobservedingeneral.thattheage was ,rr. *or, i*port*iuuriuule foliowed by the sex factor for the differen-

t ia ld iagnosis , i .e . ,youngerchi ldrenandfemaiechi ldrenwerel ike ly toget- thediagnosis of NNF,'wheieas older and male childfen were often diagnosed as

Sutterers'rrr is*asconfirmedinthechangesinthediagnosisofsomechildrenwhomadefoi low-upvisitswhoweresubsequentiyrediagnosedasStutterers.

In view of these findings *"d ;";;p"ut"a *nfusiJns and problems ob-

served in rhe .ri"i."i olrgiori, of childhood disfluencies' the present study

was undertaken with the foilowing objectives:

CHILDHOOD DISFLUENCIES CLASSIFICATION

lVlaterials

DisfluencyAssessmentProcedureforChildren(DAPC)wasformulatedtoas-sess the disfluent children using the following materials:

l .Ques t i onna i re :Ade ta i l edques t i onna i rewaspreparedbasedonSS l 'SPl,stutteringchronicitypreoict ionchecklist(Cooper,L973)andotherfactors associated with stuttering onset and development (see Appen-

10i

1. To classify and subgroup young disfluent children based on family his-

tot of stuttering, ,f,"""i, and nonspeech characteristics, and concomi-

tant'speech and language problems'

2. Todevelop a.o.pi"rr"nsive procedure to facilitate early, easy' and ob-

jective differential diagnosis of disfluent children'

Twogroupsofd is f luentch i ldrenbelowtheageof6yearsweretaken.Theda taon tenva r iab les (age ,sex , f requency 'du ra t i on 'phys i ca i concomihn tscores of disfluenclrr, i-riltori.a1, attiiudinal, and behavioral scores based on

the questionnaire, along with the presence of family history and associated

problems) *"r" ur"J i .ornpure disfluenr children diagnosed as NNF as

against stuttering. Artificial Neural Network (ANN), a computer program'

was used to classify the groups. ANN was trained on the data of Group I

(comprising 25 disfluent crriloren) on the 10 variables, whereas the data from

c.oup iI (Jompris ing26 disfluent children) on rhe same set of variables were

r,rsed for predicting tie diagnosis. The procedural considerations for adopting

ANN for the classification of NNF and stuttering subgroups are detailed.

METHODOLOGY

SubjectsTwo groups of disfluent children who had registered at AIISH, Myscre' were

chosen as subjects fbr the study. These subjects were in the age range of 2:3 to

6years (a t the t imeo f in i t i a l reg i s t ra t i on ) 'w i thgood-know ledgeo fKannada 'either as mother tongue o. u, -"dium of instruction in school (Kannada is one

of the south Indian ilnguug"s of Dravidian origin). Group I consisted of 25 of

the 64 disfluent children (18 males and 7 females), registered during 1992,

who responded for follow-up. Group II comprised 26 new dist-luent cliildren

( 19 males and T females) takln form those registered at AIISH in 1995 ' Group

II children were foilowed up after 6 months to confirm the diagnosis and to

check for the predictive u..,,u.y of the ANN diagnosis' of the 26, only 21

( l6ma leand5 fema le ) responded fo r fo l l ow-up .Becausea l l t hech i l d ren inGroup I had crossed the age of 6 years and their diagnosis were confirmed,

they were not followed up"further. Al1 the children were screened to rr-rle out

associated audiologic problems and mental retardation'

102 Y. V. GEETHA ET AL.

dix). These questions were framed to elicit historical, attitudinal, and

behavioral indicators of disfluency, motor, speech and language devel-

opment. and scholastic history. The scoring was designed to give more

weight ro the items having more predictive values as per the literature

ie'g., oisnuencv types, presence of avoidance, secondaries, and family

hisiory). Ir yieldei a total score of 50 (15, 15, and 20 each under histori-

cal, attitudinal and behavioral indicators of disfluency, respectively)'

2. The Diagnostic Kannada Picture Articulation Test by Babu, Ratna, and

Be t tage r i ( l L )12 )andKannadaLanguageTes tdeve lopedands tandard -ized b,y Regronal Rehabilitation Training Center (RRTC), Madras and

Ali Yavar Jung National Institute for the Hearing Handicapped

(AYJNIHti), Bombay, were used to test the often repcrted associated

articulator-v and.language disorders in stuttering children.

3" Picture stoi-v sequena$ of six common Panchatantra stories $'ere used

to obtain narrative sampies from the'sub.jects'

4. Stutterint Scveil ty Instrument (Riley, i980) u'as used io aSSeSS severlty

6 .

of s tu t te t ' t t tS

Natiopal Pairasontc Portabie tape recorder with bui i t - in microphone and

headsei .u,its useil for audio recording of speech sampies rvhich were

t ransc l ibcd i ( ' i l na l1 lS is .

A N N ' a C o m p u i e r p r o g r a m , w a s u s e d f o r t h e a n a l y s i s o i r e s u l t s .

Procedure

Detailed history regarding the age and nature cf onset of stuttering, family his-

tory, status of the problern, attitudes of self and oihers with regard to child's

disfluencies, handeiiness, motor and language develonlnent, and schclastic

history were circl i fc ir-om parents in response ic i tems in the questionnaire'

Later, the child r,,'as seated iomfortably and after a brief conversation to buiid

rapport, art iculation and language testing was carried out using appropriate

test materials. Recrtation, narration, and conversation san-rples were el icited

using pictures and story sequences by using questioning and repeat after tech-

niqu"s. The whoie ,.rrion was audio recorded using the portabie tape recorde;:

that was kept about 5" away from the child. Except for one child who insisted

on switching off the tape recorder, all the children enjoyed being recorded and

often wished to hear their speech from it. Proper verbai and tangibie reinforc-

ers were used to marntain motivation and cooperation from each child' In a

couple of instanccs when the chiid refused to cooperate, parents were Ie-

quested to elicil ar]d record the response from the child in the absence of the

exanr iner ,especia i iy ' lor thespeechsamples. inoneor twoinstanceswhenthesample was not adequate and when the paients reported greater disfluency at

home, speech saInpies were recorded by parents ai home and biought for anal-

ysis. Recitation was eiicited by asking chilciren to reciie numbers from I to 25'

5 .

' : - : - - : 4 .

CHILDHOOD DISFLUENCIES CLASSIFTCATION

"/ week days, months, and nursery rhymes or poems. Narration samples wereobtained using picture stories (sequences of common Panchathantra stories)that were narrared to them initially by the examiner and later elicited using

cueing or questioning whenever necessary. Conversation samples were elic-i ted using a set of 15 questions pertaining to their school, home, daily rou-t ines, and hobbies. The recorded samples of al l the 51 children (25 in Group Iand26 in Group tl) were transcribed and analyzed.

Art iculation and language test results were compared with the availablenorms. Speech and nonspeech disfluencies were assessed using the Children'sSSI (nonreader's scale), which yielded frequency, duration, ahd physical con-comitant scores. SSI also provided severity ratings based on percentile scoresas very mild, mild, moderate, severe, and very severe.

All iitese scofes were compared between children diagnosed as NNF andstuttering in Group I and II along with the presence of concomitant articula-

tion and language problems and family history of stuttering.AII the children were routinely assessed in the Department of Speech Pa-

thology at AIISH before assessment on the DAPC. The diagnosis arrived at inthese routine irssessments were compared with those on the DAPC.

RESUL"S AND DISCUSSION

The daia on difl-erent variables were compared across NNF and Stutteringgroups by taking each variable separately. Further, using ANN analysis. clas-sitication of NNF and stuttering subgroups was attempted by taking 10 of theinrportar-rt variables, which wil l be discussed in the later part of this section.

. . .

Descriptive .A.'nalysis .Age and sex. The data on the age and sex distribution of cases at first

consultation, during first follow-up for Group I and Group II and their diag-noses as per routine clinical assessment in the department, were comparedwith those using the DAPC. Table l .provides this comparison along with thatfor the follow-up evaluation for Group IL

As mentioned earlier and as revealed in the pilot study, during routine as-sessment in the department, age and sex were the main criteria for differentiat-ing children with NNF and stuttering. As shown in Table I, there were dis-crepancies in the diagnosis between the routine assessment I (first

consultation) and second assessment (fol low-up) and between routine assess-menr and DAPC. However, the agreements were consistently better during thesecond evaluations, i.e., between RA-2 and DAPC for Groups I and II, high-lighting the age tactor in the assessment. There was only 70Vo agreement be-tween the two RAs for Group II where it was 95Vo on DAPC, indicating highpredictive accuracv of the latter.

t04 Y. V. GEETHA ET AL.

Table 1. Comparison of Age, Sex, and Diagnosis for Group I and Group II Between

Routine Assessment (RA) and DAPC

Group I Group Ii Group II

R A - 1 T(A.Z DAPC RA- 1 DAPC-1 RA-2 DAPC-2

AgE SCX NNF STG NNF STG NNF STG NNF STG NNF STG NNF STG NNF STG

z-3

34

A <a - J

5-6

6 +

Total

Total

It was observed in general that the mean age ol onset of disfluencies was 3

years with a gradual onset in majority 0AV;S of children in bath groups. As of-

ten reported in the literature the male-to-female sex raiic fcrr the stuttering

group in the present study also rvas 4:1. The faniiy history cf ':luttei:ing for

both Croup I and iI was about 707o, rvhich is torvards the higirest range of

those reported in the literature i.e.,3Ac/o '.o 59aic (Blco<Jstein, 19i31).

Speech Characteristics Observed in Strigteriilg araet fiFiF ChiFdres?

Types, fiequency, duration of disfluencies, physical ccncomitanis (secondar-

ies), rate of speech, total SSI Scoles with their corresponding percentiles and

severity ratings, historical, attitudinal, and behavioral indicator scores (based

on the questionnaire data) and presence of associated problems were coln-

pared berween NNF and stuttering children in Group I and II. Results of this

comparison are given below.Type of disfluency. Types of disfluencies. especially the prolongations

and articulatory fixations were present only in stuttering children. Sound/syl-

lable repetitions were present in both ihe groups but their frequency' number

of unit repeiitions, presence of schwa, and rate of repetitions were useful in

clifferentiating stuttering children. Pauses, both audible and inaudible, were

present in both NNF and stuttering groups, specially the audible ones.

Frequency of disfluency. Frequency of disfluencies in NNF and stuttel,-

ing children in both Groups I and II is given in Table 2. There was slight ovei-

t l J

l t l l

- 3 3, - t

i { t \: J

t t ' ' ,

1 Al - T

; ; ; 1 ;J ) I +

8 1 3 2 1 9

- 1 l -- 1 i *

J - - J

l - - lA 1 ? 5

| ' 2

1 - 35 5 , r 2 - 5i 3 i i . - 2l 72 l6 1 2 1 5 4 . 4 1 53 5 5 ? 1 69 t j 2 0 6 , i 2 r

Il -

1 tt l

3 i

a r nA Aa t

1 1 l Al - l T

I

')

3

3

M 4F 5M 1F 2M 2

I / ' '

l -

lVt

F _M 1 5F 8

?3

/ .a . : ' . ' , i '

CHILDHOOD DISFLUENCTES CLASSIFICATION 105

d!4'

lap in the frequency of disfluencies among the NNF and stuttering childrenwhen sound syllable repetitions and prolongations were considered. NNFchildren had 2-37o disfluencies while it ranged from 3-30Vo for the stutteringchildren, with majority showing around 107o disfluencies.

Duration of disfluency. As seen in Table 3, all NNF children had fleet-ing disfluencies while the stuttering children exhibited disfluencies rangingfrom half a second to as long as 60 seconds with majority having disfluenciesof l-9 seconds duration.

Fhysicalconcomitants. Physical concomitants or secondary behaviorswere observed oniy in stuttering children (in about 60-707o), the most com-mon ones being eye blinking, nose flaring, grimaces, or frowning. Thesescores based on SSI are given in Table 4.

in some of the children in whom these secondaries were observed, theywere mild and could be missed unless carefuliy looked for. In about l5-207oof mild stutterers they were also present, whereas they were not observed insone severe stutterers. This indicates that the presence of secondaries was notcorrelated with severity although their presence contributed to the severity rat-irrgs. About I07o of these children with severe stuttering demonstrated someadult-like secondary behaviors like stamping, abnormal lip, hand, and headrnovcments.

Some stuttering children also demonstrated certain avoidance behaviorssuch as poor eye contact, looking away, and low volume of voice associated'*,ith their stuttering. These observations are in contrasi to the traditional viewpoilt t that secondaries and avoidance behaviors are learned reacLions to copewrth or compensate foi stuttering (Conture, 1990) and are assumed to emergeduring later stages in qh9 development of stuttering. However, i t 'snpports therecent findings that these behaviors are frequentiy associated with child:s stut-l;nng ryjlrlja a feri' monrbs of reporttd on-ser eld r.h:rr rlrel' can show-qualitative and quantitative changes over time (Schwartz,Zebrowski, & Con-ture, 1990) .

Table 2. Frequency of Disfluenci es (7o) in NNF and Stuttering Children in Group I and tr

Group I Group II

Fq. in 7o Score NNF' STG NNF STG

l1 1: - -)4

1-9r 0 - r 4r.5-2829

BO20

46I

r0l z

t 4t oIB

44.455.6 12.5

12.56 .3

+- ' r . I

t 2 .46.36.3

23.89.5

23.84.8

28.64.8

106Y. V. GEETHA ET AL'

Table 3. Duration of Disfluenc ies (%)in NNF and Stuttering Children in Group I and iI

Group I Group II

Durationin Sec Score NNF STG ryNF STG

Fleetingl12I2-9

l 0-3030-60>60

l23A

561

100i 8 .7

a 1 aJ I . J

6.3

1001 928.638.49.5

Ra teo fspeech .Thera teo fspeechwassub jec t i ve l y ra tedass low ,ave r -age,andfast ' l twasver i f iedagaindur inStranscr ip t ions. Ingenera i therateofspeech was fasrer in stutterinjchildren (o*-laqr), compared with NNF chil-

dren. There were variations oiserved in their rates in different communicative

contexts. That 1s, .r,itd..n often used faster rates during narrative tasks than

during conversatio,r. ri it not known whether faster rates were used by these

chiidren in nn attei.npt to avoid stuttering (as reported by VanRiper' 1982) or

the use of lasrer rates precipitated stutieri'g (Kloth" 'ianssett' Kraaimaat' &

Brutten, 199-5 ).Seve r i t y .W i th respec t to theseve r i t yo f thep rob len rbasedonSS l , t he re

was slight cvertap ;";g the NNF and siuuering children, NNF children had

yery milcl ro *rild J.sr"J severi.ty rvhereas stuuering children rvere rated b*-

tween mild to very severe degree'

H o w e v e r ' a s s h o w n i n T a b l e 5 , t h e r e w a s n o o v e r l a p i n t h e i r s c o r e s ( a n d

o";;;;i.'",ni.i, ,."g"0 t o* 0_8 (0-l 1 percentilei for i{NF and945 (t2-

100 percentile) for stultering children. r;aiiabiiir,v in stuttering severit'; end

,yrnpro*o,ology' was reported and observed ta a large exleni'

H i s ta r i ca l , a t t i t ud ina l ,andbehav io ra l f ac to rs . c r i t heh i s to r . i ca ianda t t i t ud lna i tac to rs the rewasover lap insco res inabou t l0Too fNNFandverymi ldstut ter ingchi idren ' Ingenera l 'NNFchi idrenhadscoresbeiow5whereas

Table 4. Physicai Concomitant (PC) in E:t" NXF ""d

St"t"dt

Group I Group II

NNF STG NNF STGPC Score

01 4) - l u

I 1 - 1 51 6-20

100 1 2 . 556.23 1 . 3

100 1928.64'7.64.8

. - .: -;...


Table 5. Total SSI Scores, Percentiles, and Severity for NNF and Stuttering Children(Vo) in Group I and II

Group I Group II

Score Percentile Severity NNF STG NNF STG

SSI

6-89- 13

l4 -15l 6 - 1 92C_l324-3728-303 r -45

0+5-l I

12-23244041-6061*1778-8990-r0097-r 00

V. mi ldMildMildMildModerateModerateSevereSevereV. se.rere

80zo

44.455.6

:

-:

rgz25.6251 218.7

;l 914.423.89.59.54.8

most stuttering children's range was between 5-15. Attitudinal factors (of self

and significanr others).are important in determining stuttering disfluencies as

negative attitudes have important repercussions.As can be seen from Tabie 6, behavioral factors clearly differentiated stut-

tering and NNF groups without overlap and scores were directly related to the

severity of the problem. NNF children had zero scores whereas it ranged from

3-20 for stuttering children.Associated communication problerns. Articulation and language prob-

Iems were the most common associated problems found in stuttering children

of both the groups. These were more predominant in Group II stuttering chil-

dren (nearly 80Vo) compared with Croup I children. This supports St. Louis,

Murray, and Ashworth (1991), who reported that the associated problems are

more frequent in younger Stutterers compared to older ones. AlSo, delayed

Table 6. Historical, Attitudinal, and Behavioral Indicators (HI, AI, and BIrcspectively) fbr NNF and Stuttering Children (7o) in Group I and II

Group I Group II

Variable Score NNF STG NNF STG

802A

H]

A I

N TD I

0-45-9

l0-150-45-9

r 0-15045-9

10-15tG20

8 8 . 91 l . l

88 .9I l . r

100

1 8 . 88r.2

6.393.7

ez.sJ t . z

6.3

100

95.2.1.8

23.871.4 '4 .89.5

28.638. i23.8

ed

108

: - .

Y. V: GEETHA ET AL.

motor milestones and delayed speech and language development were re-

ported in about 50Vo of stuttering children and none in NNF children.

Analysis of results using.A'NN

An attempt was made to classify or subgroup disfluent children based on the

scores for the above variables using ANN, a computer based program. In one

of the recent studies by Howell, Sackin, Glenn, and Au-Yeung (1997), an at-

tempt was made to recognize and locate disfluencies in young children using

ANN. Hor.vever, because the details of the study are not available in the said

reference, comparison could not be made-with the present study. As the appli-

carion ol ANN is new to this field, a brief introduction and methodologic con-

siderations lor using ANN is given below.

ANN, commoniy referred to as Neural Network, are also termed aS neuro

computers, connectionist networks, and parailel distributed processors. In the

most generai form, an ANN is a machine designed to model the rvay in which

the brain performs a particular task or function of interest by using electronic

components or simulated in software on a digital computer. It resembles the

brain'in trvo iespects: knowledge is acquired by the network through a learn-

i lg process and inter-neuron connecticjn strength, known as weights are used

ro srore knou,iet lge. In other words, ANNs are biological ly inspired networks

having tl.re apparent ability to imitate the brain's activity to make decisions

and draw conclusions when presented with complex and noisy information

( H a y k i n . 1 9 9 5 ) .ANN d6rives its computing power ihrough its inassively paraliei distrib-

ured structure and ifs ability to learn and to generalize. Generalization refers to

the ANN producing reasonabie outputs for inputs not encountered during

training (learning). These two informaiion proce.ssing capacities of ANN

make it possible to solve compiex (large scale) problerns that are cuffently in-

tractable.Einpirical classification techniques, such as cluster analysis, provide methods

fbr grouping individuals who show similar pattern or response on a given set of

variables. However, they do not ensure that they (the clusters) are psychologi-

cally or educationally meaningfui or predictive (McKinney & Speece, 1985).

The basic structure of ANN used in the present study is shown in Fig. i .

Frocedure used for classification of disfluencies using ANN. The mul-

rilayer perceprr6n classifier was used in the present attempt at classifying dis-

f luept behavior in chi idren. A mult i layer perceptron has wide practical appli-

cation such as pattern recognit ion, speech recognit ion, and fault diagnosis.

They are most appropriate when bir-rary representations are possible and can

be used aS an associative memory or to solve optimization problems' This

kind of r]etr,vork has input, butput, and hidden layers with three units or nod-es

in the hidden layer. Back-Propagation (BP) algorithm, one of the supervised

CHiLDHOOD DISFLUENCIES CLASSIFICATION

HIDDEN LAYER

VARIABLE I

VARIABLE 2

VARIABLE 3

VAzuABLE 4

LAYER

VARIABLE 1O

ADruSTABLE COIINECTTON WEIGHTS

Figure 1. Basic structure of ANN.

learning algorithms that involves a multilayered interconnections of neurons,was used. The BP algorithm derives its name from the fact that error terms inthe algorithm are back-propagated through the netr,vork on a layer-by-layerbasis and naturaliy is more powerfui in applications (Lippman, i9g9).

ln the present study eight classes of disfluencies were differentiated basecion 10 variables. Eight classes were chosen because values to the power twoare easy for binary coding essentiai for these networks. The 10 input variableswere age, sex, frequency, duration, physical concomitant scores of disfluen-cies, historicai, attitudinal, and behavioral scbres based on the SSI and ques-tionnaire data along with the presence of family history and associated prob-lems. The values obtained on these variables were normalized or decoded toget scores within zero and one (relative scaling of numbers between maximumand minimum to be mapped ro numbers between 0 and 1). This is required asanaiysis is on premise of binary valued variables obtained by selecting a suit-able threshold. The eight classes of output were as follows:

Normal Nonfluency (NNF) INNF with associated problems - 2NNF with family history and associated probiems 3Stuttering 4Stuttering with associated problems 5Stuttering with family hisrory 6Stuttering with secondaries and famiiy history 7stuttering with secondaries, family history and associated problems g

l l 0".

U. OUU'HA ET AL.

For the purpose of classification, multilayer perceptron was adopted with

ten input features corresponding to ten variables, with three binary output

unirs i000,001,... . . ,111) representing eight classes. Generalized Delta rule'

also known aS elror Correction learning rule for supervised learning tech-

niques, was used (see Lippman, 1987 for details)'

After training the ANN with the experimental data of Group I (25 disfluent

children) and data of Group II (26 disfluent children) was taken for output

generation. As the sample size in Group i was very smali, Some items ran-

dcmly chosen were repeated for training ANN'

Table 7 shows the comparison and DAPC diagnosis for Group I and II and

ANN prediction for Group II. As indicated in Table 7 (with asterisks), there

were four misclassif ications altogether (sl.no. N8,N9,Nl1 and N22)' Of these,

two wefe sirlple, in which stutleiirlg w'ith associaled problems was ciassit-ied

as stuitering with family history (si.no.N8) and I'{NF with family history and

as-sociated problem was classif ied as NNF (sl.no.N9). The other two v"ere of a

sericus narure where a stuttering chiid Vrith family history was classified un-

cler NNF with associated probiems (sl.no.Ni i) and another child (sl'no'N22)

having stuttering with family history and associated problem was classified as

NNF rvirh family history and associated problems. Both these latter chrldren

ntisciassified across groups had mild and very rnild stuttering symptoms

where there were symptom overlap with NNF group in some of the variables,

which could have effected their misclassif ications'

If rve exclude the two children who were misclassif ied within NNF or stut- '

renitg caiegories; the total overal i agreement in classifying chlldren in fte.

rnain NNF (classes' l . to 3) versus stuttering categories (classes 4 to 8) is92ah

which is reasonably high considering the reiatively smali sample size used for

training the ANN.C.o"up II chiidren were follorved up aiter 5 months to check for the predic-

iive accuracy of the procedure. Of the 26 children, cnly 21 responded for fol-

iow-up. In only onsinstance tir* diagnosis was changed from NNF to miid

srule;ing. In this child the speech samples obtained in the first evaiuation was

inadequate aS the child was very uncooperative and hardly any long utterances

were obtained.Intra-j udge reliabilitY

each other) for GrouP II

more objective scores, it

between the two DAPC assessments (indeoendent of

was found to be high (95Vo). As DAPC is based on

is expected to yield better inter and intra individual

re l iab i l i t l , . . '

The et-ficacy of ANN could be increased in classifying or subgrouping dis-

iluerrr children by increasing tire training sample which plays a cruciai roie in

actual prediction of groups. Hower{er, as reported often, subjectivity influ-

ences wharever objeciive procedures we adopt in the field of speech pathol-

ogy. This is more so with regard to stuttering as it is known for its inter- and

intra-individual variability. Tlie difticulties involved in identifying and qudn-


Table 7. DAPC Diagnosis for Group I and II Along Witirl ANN Prediction for Group II

Sl . No. Age Sex DAPC Sl. No. Age Sex DAPC ANN Prediction

o l

L-r -1

o4o5o601o8o9

o 1 0o l ro t 2o r 3o l 4o r 5o 1 6o t lo r 8o r 9o200 2 ro22o23o24o25

l : 36:35:57:A5 :8A . <

6:l

1 :37:57:25 :2

5:55:65:'1a . A'1 .1

5 :05 :26:05 :25 :06 :3

N IN2N3N4N5N6N7NBN9N 1 0N 1 rN l 2N l 3N 1 4N 1 5N 1 6N l 7N r 8N r 9N20N 2 rN22N23N24N25N26

/ 1 . A

6:06:04:3+'.33:43:3A . A

A . O

4:95 :3A . Aa - +

J:.{.

6:04:06:06:0

4:83.45:03:3

3:02 :82 :8

FMMMMMMMMMt\/{l Y l

MMFMtrt-

MMMMtrMn

E

I3656l3ox

h

56a

5z4

+511i8l5l

MFMMMMMFMMI

MMMMMMF

MFFMMMMF

o

845

85

J

o

t)

56

885I5I55857ii

6

8435oo

6*1 { <

82*J

888585I553 *57I,I

tifying the stuttering behaviors are often debated and ii depends on the type of

sample (audio/audio-visual), the person who is making the judgments and the

size of the sample.'A representative sample covefing different situations of

sufficient duration (preferably video recorded) will facilitate proper identifica-rron or differentiation of the disorders. This is crucial because misdiagnosiscan have serious repercussions on the child's speech and behavior in general.

S{JMM,A.RY AND CONCLUSIONS

Based on the observations of results of this ANN a{alysis. it can be concludedthat the disfluent chiidren can be subgrouped into ['{NF and stuttering groups

on the basis of DAPC. They can be further grouped based on associated com-munication problems and family history of stuttering. The eighth group is the

most severe of all, with secondary symptoms, family history, and associated

112 Y. V. GEETHA ET AL.

problems. These groupings have important prognostic and management impii-

cations. Thus ANN can be developed as a useful clinical tool for objectifying

our diagnostic procedures.

WeIegethat

are thankful to Prof. N.S. Vishwanath, Dept. of Neurology,

of Medicine, Houston, Texas for his useful cornments and

greatly improved the earlier version of our paper'

Baylor Col-suggestions

REFBRE,NCES

Babu. P.R.M.. Ratna, N. , & Bet tager i , R. ( i972) . Test o f ar t icu lat ion in

Kannada. Journal of AtL lndia Institure oj Speech and Hearing IIl,7-24'

Bloodstein, O. (1974). The rules of early' stuttering. Journal of Speech and

H e aring D i s o rders 3 9,'37 9-3.9 4.

Bloodstein, O. (19S7). A hattdbook of srurtering (4th Ed.). Chicago: National

Easter Seal Society.

Conture, E.G. (1990). Stuttering: Assessment and e'taluation (Chapter 2, pp'

35-84) 2nd ed. Englewood Cliffs, NJ: Prentice Haii'

Cooper, E B. ( 1g73). The development of stuttering chronicity prediction

checklist for schooi-age stuttefers. Jountal of Speech and Hearing Disor'

de rs 38 .215 -233 .

Curlee, R.F. (1993). Stuttering and reLatecl disorders ol fluencp'New York:

George Thieme Verlag Stuttgart.

Dalron, P. (1983). Approaches to the treatftxerfi of stutteritzg. Kent, UK:

Croom Helm Ltd.

Haykin, S. (i995). Neural nefi1)orks: A comprehensive foundation.l-trew York:

MacMillan College Publishing Company'

Howell, P., sackin, S., Glenn, K., & Au-Yeung, I. (1997). Automatic recogni-

tion of stuttering. Joumal of Fluency Disorders 22(2)' 83 '

Johnson, w., Young, M., Sachs , J.L., & Bedell, G.N. (1959). Effects of super-

ventilation and tetany on the speech fluency of stutterers and nonstutterels'

JottrnaL of Speeclz and Hearing Research 2' 203-215 '

Kloth. S.A.M., Janssen, P., Kraaimaat, F.w.R., & Brutten, G.J. (1995).

Speech- motor and linguistic skilis of young stutterers prior to onset' Jour'

nal of FLuency Disorders 20' 151-LIO'

Lippman, R.P. (1987). An introduction to computing with neural nets, IEEE,

ASSP Mapazine 4-22.


Lippman, R.P. (1989). Pattern classification using neural networks. IEEEC ommunication M agazine 47 -64.

Riley, G.D. (1980). Stuttering Severity Instrument for young children andaduLts. Austin, TX: Pro-Ed.

Riley, G.D. (1984). Stuttering prediction instrumentfor youtxg children. Aus-t in, TX: Pro-Ed.

Riley, G.D. & Riley, J. (1979). A component model for diagnosing and treat-ing children who stutter. Journal of Fluency Disorders 4,219-293.

Schwartz, H.D., Zebrowski, P.M., & Conture, E.G. (1990). Behaviors at theonset of stuttering. Journal of Fluency Disorders I5, 17-86.

Silverman, F.H. (1992). Stuttering and other fluency disorders. Englewood.Cliffs, New Jersey, Prentice-Hall.

St. Louis, K.O., Murray, C.D., & Ashworth, M.S. (1991). Co-exist ing com-niunication disorders in a random sample of school-aged stutterers. Journalo.t''FLuency Disorders 16, 13-23.

Van Riper, C. (i971). The Nature of stutrering. Englewood Cliffs, NJ: Pren-t ice Hal i .

Van Riper, C. (1982). The Narure of sturtering.Znd ed. Englewood Clif fs, NJ:Prentice Hall.

Wingate, M.E. ( 1962). Evaiuation of stuttering, Part I: Speech characteristicsof young children. Journal of Speech and Hearing Disorders 27, 106-115.

Yair i , E.H. & Lewis, B. ( 1984). Disf luencies at the onset of stutterin g. Journalo.f Speech and Hearing Research 27, 154-159.

APPENDIX

Disfluency Assessment Procedure for Children

i . HISTORICAL INDICATORS: (Max. Score l5)l. When did the child first exhibit disfluencies?

Within 6 monthsGreaier than 6 months

2. What are the related circumstances?Any illness / injuryEmotional i PsychologicalExposure to stutterers

i) Close relativesii) Far relatives

(DAPC)

Score obtained

IZ

II2

1 1 4I l J

t t 4 Y. V. GEETHA ET AL.

(No score)

(No score)

2I

6L

i

Score obtainedas 'severe '?

II.

iii) Friends / neighbours:Any other reason (specify):

3. Has any family member ever stuttered?Parents F / MSibl ingsUncle / Aunt / Cousins p / MFar relatives (specify)

4. Does anybody in the family have or had any speech /hearing problems? Please specifiy:a) Relation to the childb) Associated problemc) Present status

5. Did the child's problem starr suddenly / graduallyintermitrently / NA

6. Is the severity of the problemIncrediing./. Same I Fluctu atingDecreasing

7. Did the disf luencies besin rvithBlockingsEasy repetitions / hesitarions

ATTITUDINAL INDICATORS (Max. Score 15)8. Is the stuttering now or ever considered by the child

No 0Yes I

9. Does the child indicate that he perceives himself to be a stutterer?Never 0uccaslonally IOften ?

10. Does the child indicate that he experiences communication fearbecause of stuttering?NeverOccasionallyOften

I l . Does the child indicare that heworse?NoYes

0i

. a

believes stuuering to be gettin g

U

I

0I

I2. Does the stuttering vary periodically? If so; when and how?NoYes

13. Does the child get frustrared when he cannot utter a word?(e.g. cries, stamps foot, bites self etc.,)

*d


NeverOccasionallyOftenDo either of the parents consider the child to be a stutterer?

NoYesDoes either of the parents indicate that they believe the child

will not out-grow stuttering?NoYesDoes the child's disfluency make the parents feel:

UnconcernedConcernedVery concerned

17. Has the child been teased about his stuttering?

NeverOccasionallyOften

III. BEHAVIORAL INDICATORS (Max' Score 20)

(To be scored after analyzing the transcript)

18. Do syllable repetitions occur more than twice on the same word?0No

Yes 1

19. Is the rapidity of repetitions faster than normal?' N o 0

Yes I

20. Is the schwa vowel inappropriately inserted in the syl lable repetit ion?

NoYes

2l . is the airflow during repetitions often interrupted?

No

22.YesIs vocal tension often apparent during the repetitions?

NoYesDo prolongations last longer than 1 second?

NoYesDo prolongations occur on more than one word in 100?

NoYes

02

1 1 5

012

0I

I AI T .

r 5 .

l 6

L J .

1 A

0I

0t

. ' )L

Score obtained

0I

2

02

01

0I

0 , /1I .

i l6

Are the prolongations uneven or intem.rpted rather than beingsmooth?NoYesIs there observable tension during the prolongations?NoYesAre the terminations of prolongations sudden?NoYesDuring prolongations of voiced sounds is the airf low inten-unted?NoYesAre the silent pauses prior to the speech attempt unusuaiiy iong?NoYesAre the inflection patterns restricted and monotonous?NoYesIs there a loss of eye contact during the moment of disfluency?NoYesAre there observable and / or.distracting extraneous fae ial or bodymovements during the moment-of disfluency?NoYesDoes the child sometimes change a word because of lear of '

stuttering?NoYesDoes the child actively avoid speaking situations?NoYes

ry. GENERAL OVER-ALL DEVELOPMENT (No scoring)35. Wbre the general motor milestones of the child normal?

If not, specify.36. Does the chiid show specific hand preference?

a) Right lleft I ambidextrousb) Age of establishment of hand preferencec) Was thcre a forced change in handedness?

37. Is there any associated speech / hearing / language probiem?If yes, please specify.

25.

26.

27.

28.

0I

0I

0I

0I

0I

I

U

I

?oL J .

30.

3 1 .

J Z .

0aL

a aJ J .

J a .

0i

01

*d


Problems Age when notedHLDSMRDSDSMAOthers (Specify)

Treatment any Present status

38. Is the scholastic performance: Poor / Average I Good

39. Please specify the areas of weakness if any:

40. Has the child-been given any treatment for his stuttering problem?

a) The type of treatment

b) Treatment byc) Duration of treatmeni

d) Present staius41 . Rate of speech42. Speech mechanism43. Reading44. Writing45. Remarks

S l o w / A v e r a g e / F a s tAdequate / InadequatePoor /Average/GoodPoor /Average/Good

Classification of Childhood Disfluencies Using Neural Networks

Documents