multi æstep prediction of pathological tremor with ... · layer 2, eac h ratio of fir i ring...

Multi step Prediction of Pathological TremorWith Adaptive Neuro Fuzzy Inference System

(ANFIS)

Sivanagaraja Tatinati and Kalyana C. Veluvolu

School of Electronics Engineering, College of IT Engineering, Kyungpook NationalUniversity, Daegu, South Korea

Abstract In this paper, to overcome the inevitable phase delay due to pre filtering techniques andelectromechanical delay in the procedure of pathological tremor compensation, a machine learningtechnique based on fuzzy logic and neural networks (adaptive neuro fuzzy inference system (ANFIS)) isemployed. Experimental evaluation on tremor data is performed with data acquired from eight patients.Results show that ANFIS provides good performance in comparison with existing methods.

I. Introduction

Pathological tremor is defined as an involuntary and approximately rhythmic oscillations of a body part [1].Pathological is one of the serious disabling forms of tremors and often accompanies with aging. Althoughpathological tremor is not a life threatening disorder, but it hampers the ability of individual to performdaily living activities like drinking water with a glass, unlock the door with a key etc. Further, the inabilitydue to tremor leads to social embarrassment [1]. To compensate the functional disability (tremor) therebyto improve the quality of life, several pharmaceutical and surgical therapies are developed. The methodshave their limitations and side effects. So, an effective treatment that can compensate tremor accurately yetto be developed.

Over the last decade, considerable research has been focussed on developing active pathological tremorcompensation techniques in real time [3], [4], [6], [7]. One such approach is compensating pathologicaltremor by stimulating the relevant muscle groups with appropriate electric pulses thru functional electricalstimulator (FES) [3]. For example, consider hand supination pronation muscle group, if tremor is due to thesupination muscle group necessary electric pulses will be provided to pronation muscle group to counteractthe involuntary movement, and viceversa. To compensate the tremor, a wearable device named as WOTAS(wearable orthosis for tremor assessment and suppression) was developed by using a feedback controlledFES.

The efficacy of the wearable devices depends on the accurate and zero phase delay filtering of tremulousmotion from the whole motion. In the tremor compensation procedure with FES, a delay of approximately80 ms was identified. The major source for the delay is electromechanical delay (EMD), defined as the delaybetween the onset of activation of muscle and the detection of movement [2]. As a result of several studies,the EMD was identified in the range of 50 60 ms [2]. The other source for delay is the pre filtering stagewhich is used to separate the voluntary motion and tremulous motion from the whole sensed motion [7].On a whole, a latency of 80 ms was inevitable in the procedure and this delay adversely affects the tremorcompensation accuracy.

To overcome the phase delay limitation, multi step prediction with harmonic method and Autoregressivemethod was proposed in [4]. The developed methods yield accurate prediction with periodic signals. Owingto the quasi periodic nature of the tremor signals, prediction accuracy with the above methods is ratherlimited. To further enhance the prediction capabilities, in this work, we employed a machine learningtechnique, adaptive neuro fuzzy inference system (ANFIS). ANFIS has been recently popular as an effectivemethod for time series prediction, function estimation, system identification and classification problems.

Proceedings of The National Conference on Man Machine Interaction 2014 [NCMMI 2014] 22

NCMMI '14 ISBN : 978-81-925233-4-8 www.edlib.asdf.res.in

Downloaded fro

m w

ww.e

dlib

.asd

f.re

s.in

The goalthe targefuzzy ruemploystremor dresults ob

In this sdescriptiANFIS ar

Fig. 1: SpronatioFFT of th

l of ANFIS iset (predictedule describeshybrid learn

data collectionbtained with

section, weion regardingre presented.

ubject #1, An and supinahe whole mot

to find a modvalues). Thea local beha

ning rules ton, delay in thANFIS are p

first discussg the delay as

Action tremoation movemetion

del or mappie fuzzy infereaviour of thtrain is call

he procedureresented. The

s about thessociated in t

r (a) Parkinents of the le

ng that will cence systeme system. Thled ANFIS. Tand the mete last section

II. Method

collection otremor compe

nsons patienteft hand; (b) Z

correctly asso(FIS) is a knhe network sThe paper isthod employen presents the

ds

f pathologicensation proc

t tremor whZoomed porti

ociate the inpnowledge repstructure thaorganized ased are discuse conclusions

al tremor dacedure and a

hile executingion of the wh

puts (initial vresentation wat implemens follows: Inssed. In the las.

ata. Later tha brief descrip

g voluntaryhole motion (

values) withwhere eachnts FIS andsection II,

ater section

he detailedption about

alternating(0 – 5 s); (c)



Downloaded fro

m w

ww.e

dlib

.asd

f.re

s.in

Tremorwithout othe Ethipathologwork formotion)parkinsoshown in

The phasfilteringfilter/Butidentifiedwith pasaccuracythe appaacquired

is acquired wobjecting freeics Committgical tremor dr analysis, fouand four sub

onian patientn Fig. 1. FFT o

B

se delay in westage, to fitterworth bad for fifth ords band 2 20y was adversearent peak ind form the flex

with a portae flow of anyee of localdata set contur subjects’ abjects posturawhile volun

of the signal i

B. Latency i

earable devicilter the trend pass filterder ButterwoHz a delay oely effected. En EMG [2]. Inxion and exte

Fig. 2: D

A. Patho

able Motus sy motion [8]. Ahospital. Thtains posturaaction tremoal tremor arentarily performs also shown

in Tremor C

ces with FES iemulous motr was employorth low passof 20ms was iEMD, in genen [2], a delayension muscl

Delay in trem

ological Tre

system compAll the subjehe samplingal tremor, par (tremulousemployed. Fming pronatiin Fig. 1(c).

Compensat

is due to thetion from wyed. From ths filter, whereidentified [7]eral a delay ofy of 40 60 msle group in fo

mor suppressio

emor data

prises a gyroscts sign the irate emplo

arkinsonian ts motion collFor an illustraion and supi

tion Proced

two sources:whole sensede phase respe as with fifth]. In the presf 80§20 ms exs was identifiorearm.

on procedure

scope placednformed conyed is 100tremor and elected whileation, actionination move

dure With F

pre filteringd motion, aonse, a averah order buttesence of eithexists betweenfied between

e with FES

d on the patinsent forms adeg/sec. Theessential tremperforming atremor collecement with le

FES

stage and EMButterworth

age delay of 3erworth bander delays then peak musclsurface EMG

ient’s wristpproved bye collectedmor. In thisa voluntarycted from aeft hand as

MD. In preh low pass300 ms wasd pass filtere predictionle force andG and force



Downloaded fro

m w

ww.e

dlib

.asd

f.re

s.in

Thus a tfrequencto aboutdecrease

A neurointelligen

A briefimpleme

In generaand linguinput nolayers noare eithedescribed

Lp

total delay ocy, 100Hz, fort 90± phased. To overcom

Fuzzy systemnt hybrid syst

descriptionentation refer

Fig

al, ANFIS struuistic variabldes represenodes functioner fixed or add as below:

Layer 1: All nparameters.

of 60 80 msr a signal withe difference.me this delay

C. Adapt

m is inspiredtem enjoys b

of ANFIS isr to [5].

g. 3: Sugeno t

ucture followles) and mulnts the traininning as membdaptive, as sh

nodes in this l

is inevitableh frequency rWith this

y, multi step

tive Neruo F

from the supoth the metr

s provided b

type fuzzy mo

ws the takagilti layers (5 lng values andbership functhown in Fig.

layer are ada

e in the procranges fromphase differprediction ba

Fuzzy Infer

pervisory humics of neural

below. For

odel illustrati

sugeno fuzzylayers), as shd output nodetions (MF’s)3 [5]. Differe

aptive. Furthe

cedure, as sh6 12 Hz, therence tremorased on ANFI

rence Syste

man brain’s fsystem and f

detailed des

ion and ANFI

y sets with sehown in Fig.es representsand rules. Alent layers an

er, parameter

hown in Fig.delay of 6 tor suppressioIS is develope

em (ANFIS)

feed forwardfuzzy logic ap

scription of

IS architectur

et of rules (co3. In this co

s predicted vall nodes in dd their corre

rs in this laye

. 2. With theo 8 samples con ability ised in this pap

activities thepproaches.

the algorith

re [5]

ombination ofonnected strualues, and inifferent layeresponding no

er are named

e samplingorrespondsdrastically

per.

erefore this

hm and its

f numericalucture, thethe hiddenrs of ANFISodes can be

as premise



Downloaded fro

m w

ww.e

dlib

.asd

f.re

s.in

Ltt

Lcn

La

Lc

The paraLeast squThe gradbackward^ fk areoutputs)faster anbe expresin Fig. 3.

The procstep predwas conscontaine

ANFIS wmemberssquare eepochs to

Layer 2: All nthe product othis layer.

Layer 3: Likecomputes thenormalized fi

Layer 4: In thare referred a

Layer 5: Thiscomputes the

ameters of ANuares methoddient descend pass. The ethe k th desof data in th

nd reduce thessed as a line

D

cedure emplodiction, a twostructed as shed one data po

was trained wship functionrror). The opo perform the

nodes in thisof all the inpu

layer 2, eache ratio of firifiring strength

his layer, allas consequen

s layer contaie overall outp

NFIS are optd is employent method iserror measuresired and esthe training see search spacear combinati

. Multistep

oyed to perforo dimensionhown in (1). Doints:

with the first 5ns will be opptimized pare multi step

Fig. 4: Mult

layer are labeuts. Further,

h node in thiing strengthshs.

nodes are adnt parameters

ins only oneput of the sys

timized withed to updates to updatee to train thetimated outpet.Due to thece dimensionion of the con

p prediction

rm multi stepal matrix ofData points i

500 epochs. Wptimized forrameters obtaprediction.

tistep predict

elled as ¦ andthe firing stre

is layer is alsos of ith rules.

daptive. The ps.

node and thastem by summ

combinationthe parametthe premiseabove menti

put, respective implications due to backnsequent par

n of patholo

p prediction wdimensions 1n each row w

With the traiaccurate preained with th

tion of patho

d are fixed. Thength of each

o fixed and laThereby, the

parameters c

at node is alsmations of all

n of gradientters iterativelparameters

ioned ANFISvely, and n iof two updak propagatiorameters, as s

ogical trem

with ANFIS i1000£5 (eachwere chosen t

ining epochsediction (withe training e

logical tremo

he output obth rule is repre

abelled as N.e outputs of

corresponds t

so fixed andl inputs to th

t descent andly in the forwwith the ob

S is E = Pn k=1s the total nate algorithmn algorithm.shown

mor with AN

is shown in Fh row being rto be six step

s, weights of nth cost functepochs are em

or with ANFIS

tained from tesented by ea

. In this layerthis layer are

to all nodes i

labelled as Pe node.

d least squarward pass unbtained erro1(fk¡ ^ fk)2 whnumber of pams, the conver

The overall

NFIS

Fig. 4. To perfreferred to asps apart and e

neural netwotion as minimmployed for

S

this layer isach node in

r, I th nodee named as

in the layer

P. The node

es method.ntil layer 4.or signal inhere fk andairs (inputsrgence of isoutput can

form multis an epoch)each epoch

ork and themum meanthe testing



Downloaded fro

m w

ww.e

dlib

.asd

f.re

s.in

In thissimulatio

where RMk and e is

tremulou4.

For illusperformiwhole mshown. I

section, weons. To quant

MS(s) =s the predicti

us motion for

tration, theing multi ste

motion is shown Fig. 5(c), p

analyze thetify the perfo

ion error.

rm a zero ph

prediction erep predictionwn, where asrediction err

e multi stepormance, we e

hase bandpass

rror due to tn for the samin Fig. 5 (b),ror obtained d

III. Result

ahead trememploy %Acc

m is the nu

s filter with s

the delay ofme delay for sfiltered tremdue to 80 ms

ts

mor predictiocuracy, define

umber of sam

FiAacmfilobwob

Thmshthprtonefuthgyfiffilthsepr8LSasanpr

same specific

f 80 ms andsubject #1 are

mulous motions is shown tog

on performaed as:

mples, sk is the

A. Perfor

ig. 5: MultANFIS (a) Whction tremmotion obtaltering ;btained due t

with multistbtained with

he proceduremulti step prehown in Fig.he collectedrovided for oo obtain theeural netwounctions. Forhe wholeyroscope is pfth orderlter with pashe tremulousensed motiorediction forsamples (80S SVM tossociated wind EMD. Trediction wations is emp

the predictioe shown in Fn obtained afgether with t

ance with A

e input signa

rmance An

tistep predichole motion (mor); (b)ained after(c) Predictto 80 ms delatep predictANFIS

e employedediction with4. First 500d tremoroffline traininoptimized w

ork and mr multi stepmotion senrovided as anButterworths band 2 20motion from

on. Further,6 samples (6

0ms) is perfoovercomeith bandpasTo validatewith ANFISployed, as sh

on error obtFig. 5. In Figfter bandpassthe multistep

ANFIS thru

al at instant

nalysis

ction with(Subject #1,Tremulousbandpass

tion erroray togethertion error

to performh ANFIS issamples ofdata was

ng of ANFISweights formembershipprediction,nsed withn input to a

bandpassHz to filter

m the wholemulti step60 ms) andormed withthe delayss filteringmult istep

S, filteredown in Fig.

tained afterg. 5 (a), thes filtering isp prediction



Downloaded fro

m w

ww.e

dlib

.asd

f.re

s.in

error obtained with ANFIS for comparison. The prediction accuracy obtained for subject #1 is 49:21 § 1:21%.The average prediction accuracy obtained over all subjects and trials for 60 ms prediction length is 50 § 2%and 80 ms prediction length is 45:25 § 5%.

IV. Conclusions

In this paper, ANFIS is employed for multi step prediction of pathological tremor to overcome the phaseddelay and to improve the performance. The performance of ANFIS was experimentally assessed with thetremor data collected from eight patients. An average prediciton accuracy of 50§2% is obtained with themulti step prediction with ANFIS for 60 ms ahead prediction and an average accuracy of 45:25 § 5% isobtained for prediction length 80 ms.

References

1. R. J. Elbe and W. C. Koller, Tremor, U. MD, Ed. John Hopkins University Press: Baltimore, MD,USA, 1985.

2. P. R. Cavanagh, and P. V. Komi, “Electromechanical delay in human skeletal muscle underconcentric and eccentric contractions,” Eur. J. Appl. Physiol., vol. 42, pp. 159–163, 1979.

3. F. Widjaja, C. Y. Shee, W. L. Au, P. Poignet, and W. T. Ang, “ Using electromechanical delay forreal time anti phase tremor attenuation system using functional electrical stimulation,” Proc. Int.Conf. IEEE on Robo. Auto., Shanghai, 2011.

4. A. P. L. Bo, P. Poignet, F. Widjaja, and W. T. Ang, “Online pathological tremor characterizationusing extended kalman filtering,” in Proc. 30th Annu. Int. Conf. IEEE Eng. in Med. Bio. Soc.,Vancouver, Canada, 2008, pp. 1753–1756.

5. K. Manish, H. Nystrom, L. R. Aarup, T. J. Nottrup, and D. R. Olsen, “Respiratory motion predictionby using the adaptive neuro fuzzy inference system (ANFIS),” in Phy. in Med. Bio., vol. 50, pp. 47214728, 2005.

6. S. Tatinati, K. C. Veluvolu, S. M. Hong, W. T. Latt, and W. T. Ang, “Physiological tremor estimationwith autoregressive (AR) model and Kalman filter for robotic applications,” in IEEE Sensors J., vol.13, pp. 4977 4985, 2013.

7. K. C. Veluvolu, S. Tatinati, S. M. Hong and W. T. Ang, ” Multistep prediction of physiologicaltremor for surgical robotic applications,” in IEEE Tran. on Biomed. Eng., vol.60, pp. 3074 3082 ,2013.

8. G. N. Reeke, Modeling in the Neurosciences, www.MotusBioengineering.com.



Downloaded fro

m w

ww.e

dlib

.asd

f.re

s.in

multi æstep prediction of pathological tremor with ... · layer 2, eac h ratio of fir i ring...

Documents