ClassificationofGaitPatternsinPatientswithNeurodegenerativ...

Research ArticleClassification of Gait Patterns in Patients withNeurodegenerativeDisease Using Adaptive Neuro-Fuzzy Inference System

Qiang Ye1 Yi Xia 2 and Zhiming Yao 3

1Department of Sport and Health Science Nanjing Sport Institute Nanjing 210014 China2School of Electrical Engineering and Automation Anhui University Hefei 230601 China3Institute of Intelligent Machines Chinese Academy of Sciences Hefei 230031 China

Correspondence should be addressed to Yi Xia xiayiustcedu

Received 14 May 2018 Revised 26 August 2018 Accepted 29 August 2018 Published 30 September 2018

Academic Editor Po-Hsiang Tsui

Copyright copy 2018 Qiang Ye et al-is is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

A common feature that is typical of the patients with neurodegenerative (ND) disease is the impairment of motor function whichcan interrupt the pathway from cerebrum to themuscle and thus causemovement disorders For patients with amyotrophic lateralsclerosis disease (ALS) the impairment is caused by the loss of motor neurons While for patients with Parkinsonrsquos disease (PD)and Huntingtonrsquos disease (HD) it is related to the basal ganglia dysfunction Previously studies have demonstrated the usage ofgait analysis in characterizing the ND patients for the purpose of disease management However most studies focus on extractingcharacteristic features that can differentiate ND gait from normal gait Few studies have demonstrated the feasibility of modellingthe nonlinear gait dynamics in characterizing the ND gait -erefore in this study a novel approach based on an adaptive neuro-fuzzy inference system (ANFIS) is presented for identification of the gait of patients with ND disease -e proposed ANFIS modelcombines neural network adaptive capabilities and the fuzzy logic qualitative approach Gait dynamics such as stride intervalsstance intervals and double support intervals were used as the input variables to the model -e particle swarm optimization(PSO) algorithm was utilized to learn the parameters of the ANFIS model -e performance of the system was evaluated in termsof sensitivity specificity and accuracy using the leave-one-out cross-validation method -e competitive classification results ona dataset of 13 ALS patients 15 PD patients 20 HD patients and 16 healthy control subjects indicated the effectiveness of ourapproach in representing the gait characteristics of ND patients

1 Introduction

Neurodegeneration is the progressive loss of structure orfunction of neurons including death of neurons Manyneurodegenerative diseases such as amyotrophic lateralsclerosis (ALS) Parkinsonrsquos disease (PD) and Huntingtonrsquosdisease (HD) occur as a result of neurodegenerative pro-cesses Such diseases are incurable resulting in progressivedegeneration andor death of neuron cells and producingchanges in altered neuromuscular control Since flexion andextension motions of two lower limbs are regulated by thecentral nervous system the gait of a patient with a neuro-degenerative disorder would become abnormal due to de-terioration of motor neurons [1] PD and HD are twoneurodegenerative disorders of the basal ganglia and the

ability to maintain a steady walk with small stride-to-stridefluctuations is dramatically impaired [2] Amyotrophiclateral sclerosis (ALS) which is also referred to as Charcotrsquosdisease is a progressive neuromuscular disease caused by thedestruction of motor neurons in the brain and spinal cord[3] -is causes loss of nervous control of the voluntarymuscles resulting in the degeneration and atrophy of themuscles At the early stage of ALS weakness of lower motorneurons would affect the movement function and may leadto balanced impairment or altered gait rhythm [2 4]-erefore analysis of gait cadence may help us to produceautomatic noninvasive detection of movement disorderscaused by the degeneration of motor neurons

Several studies have analysed the impact of ND diseaseson the gait dynamics of the subjects Hausdorff et al [5]

HindawiComputational and Mathematical Methods in MedicineVolume 2018 Article ID 9831252 8 pageshttpsdoiorg10115520189831252

found that the ability to maintain a steady gait with lowstride-to-stride variability of gait cycle timing and its sub-phases was diminished for both PD and HD two typicaldisorders of basal ganglia Furthermore they found that gaitspeed was significantly lower in PD but not in HD andaverage gait cycle duration and the time spent in manysubphases of the gait cycle were similar in control (CO)subjects HD subjects and PD subjects In another similarstudy Hausdorff et al [2] studied the magnitude of thestride-to-stride fluctuations and perturbations of the gaitrhythm in subjects with ALS disease in comparison withhealthy CO subjects and PD and HD patients -eir studyrevealed a longer stride interval for ALS patients than all theother three compared groups -ey also found that patientswith ALS have less steady gait between successive strideintervals in comparison with control subjects using thedetrended fluctuation analysis technique

In order to facilitate automated identification of neuro-degenerative diseases like PD HD and ALS for the clinicalassistant purpose by using gait signals different machinelearning tools have also been applied in previous studiesBaratin et al [6] introduced a wavelet-based scheme for ef-fective characterization of gait associated with ALS patients-ey extracted features such as entropy that reflects theregularity of gait and coherence between left and right limbsfrom the wavelet approximation part of the raw gait signal-en the classification of ALS gaits and normal gaits wasimplemented with a linear discriminant analysis (LDA)classifier Wu et al [7] quantified the gait variability in pa-tients with PD by counting the signal turns -ey found thatthe signal turns in the stride interval time series presenteda significant difference between the healthy control subjectsand the PD patients In another similar study by the sameauthors [8] the nonparametric Parzen window method wasutilized to estimate the probability density functions (PDFs)of gait stride interval time series and then by computing themean of left foot stride interval time series and the modifiedKullbackndashLeibler divergence of the PDFs between left footand right foot as the two features they realized the classifi-cation of gait in subjects with ALS and normal subjects byusing the least squares support vector machines (LS-SVM)

-ese studies suggest that gait cadence exhibits complexand nonlinear behaviour in both patients with ND diseasesand control (CO) subjects because of the nonlinear dy-namics of the human system [9 10] However due to thepresence of neurological disorders the ND patientsrsquo gaitrhythm manifests significant differences than that of thehealthy CO subjects such as the magnitude of stride-to-stride gait variability gait speed [2] and the degree ofasymmetry in gait rhythm [11] Previous evidence has shownthat discrimination between normal and abnormal gaitcould be implemented by means of the statistical nonlinearcomputational method [1 6 8 12] ANFIS harnesses thepower of two paradigms fuzzy logic and artificial neuralnetworks which have shown significant results in modellingnonlinear functions [13 14] In the ANFIS the membershipfunction parameters are learned from a dataset that de-scribes the system behaviour under the constraints of a givenerror criterion [15 16] As far as we know ANFIS has been

successfully applied in biomedical engineering for differenttasks such as classification [17 18] and data analysis [19]

In this study a novel approach based on ANFIS ispresented for the classification of gait cadence between thepatients with ND disease and the healthy subjects -eANFIS model combined with the particle swarm optimi-zation (PSO) algorithm is used to learn the nonlinear gaitdynamics by using five gait cadence signals as the inputvariables and then the classification between patients withALS disease and normal subjects is implemented witha simple distance-based classifier by using the outputs of theANFIS model -is paper has been organized as follows thedescription of the gait dataset and an overview of ourproposed method are given in Section 2 -e experimentalresults and some discussions are presented in Section 3Finally some conclusions are given in Section 4

2 Methods

21 Description of Dataset -e gait dataset used in this studywas contributed by Hausdorff et al [2] and it has been usedby several different studies [1 6 8] to investigate the gaitcharacteristics of the patients with neurodegenerative dis-eases To measure the gait signals force-sensitive switcheswere placed in the subjectrsquos shoe -e gait signals weresampled at 300Hz via an onboard analog-to-digital converter(12 bit) -e experimental protocol required that each subjectshould walk at their normal pace along a 77m-long hallwayfor 5min According to the descriptions in [2] the first 20 s ofthe recorded data were excluded to eliminate the start-upeffects and a median filter was applied to remove data pointsthat had a standard deviation (SD) greater than 3 or were lessthan the median value As we had introduced in our previousstudy [20] the gait dataset used in this study provides thefollowing seven gait parameters left and right stride interval(time from initial contact of one foot to the immediatesubsequent contact) left and right stance interval (amount oftime when one foot is on the ground) left and right swinginterval (amount of time when one foot is in the air) anddouble support interval (time of bilateral foot contact)

-irteen ALS patients fifteen PD patients and twentyHD patients were recruited from the Neurology OutpatientClinic at Massachusetts General Hospital Boston MA USAAlso another sixteen healthy CO subjects were enrolled toconduct the comparative study Each subject provided in-formed consent as approved by the Institutional ReviewBoard of the Massachusetts General Hospital -ere was nosignificant difference between the weights and heights of theCO subjects and the ND patients -e ND patients were notusing a wheelchair or assistive device for mobility and werefree of other ailments that might affect lower extremityweakness -e presence or absence of symptoms that mightaffect the gait was determined by a qualified physician -eseverity of ALS is represented by the time since the onset ofthe disease -e Hoehn and Yahr (HampY) scale [21] and thetotal functional capacity (TFC) [22] are used to assess thedegree of neurologic impairment in PD and HD patientsrespectively Basic information of all the subjects partici-pated in this study has been listed in Table 1

2 Computational and Mathematical Methods in Medicine

22 Adaptive Neuro-Fuzzy Inference System (ANFIS)

211 Architecture of ANFIS -e ANFIS is a useful neuralnetwork method for the solution of function approximationproblems [23] An ANFIS produces the mapping relation be-tween the input and output data by using the hybrid learningmethod to determine the optimal distribution of membershipfunctions It combines artificial neural network (ANN) andfuzzy logic Such a framework makes ANFIS modelling moresystematic and less reliant on expert knowledge To present theANFIS architecture for simplicity two fuzzy if-then rules basedon a first-order Sugeno model [15] can be stated as

rule1 if x is A1 and y is B1

then z is f1(x y) p1x + q1y + r1



(1)

where x and y are the inputs of ANFIS Ai and Bi are thefuzzy sets fi(x y) is a first-order polynomial and representsthe outputs of the first-order Sugeno fuzzy inference systemand pi qi and ri are the design parameters that are de-termined during the training process

-e architecture of ANFIS is shown in Figure 1 and thenode function in each layer is described below Adaptivenodes denoted by squares represent the parameter sets thatare adjustable in these nodes whereas fixed nodes denotedby circles represent the parameter sets that are fixed in thesystem [23]

In the first layer all the nodes are adaptive nodes -eoutput can be specified as

Oi1 μAi

(x) i 1 2

Oi1 μBiminus 2

(y) i 3 4(2)

where μAi(x) and μBiminus 2

(y) can adopt any fuzzy membershipfunction For instance if the bell-shaped membershipfunction is applied μAi

(x) is given by

μAi(x)

1

1 + xminus ci( 1113857ai1113858 11138592

1113966 1113967bi

(3)

where ai bi and ci are the parameters set governing the bell-shaped functions accordingly -ese parameters are namedas premise parameters [23]

In the second layer the nodes are fixed nodes marked bya circle -ey are labeled Π indicating that they serve asmultipliers -e output of this layer can be given as

Oi2 ωi μAi

(x) middot μBi(y) for i 1 2 (4)

where the output ωi represents the firing strength of a ruleEvery node in the third layer is also a fixed node marked

by a circle and labeled N with the node function of nor-malizing the firing strengths of the previous layer -eoutputs of this layer can be represented as

Oi3 ωi

ωi

1113936ωi

ωi

ω1 + ω2for i 1 2 (5)

In the fourth layer the nodes are adaptive nodes -eoutput is simply calculated as

Oi4 ωi middot fi for i 1 2 (6)

where f1 andf2 are the fuzzy if-then rules mentioned beforeand parameters pi qi and ri are also referred to as theconsequent parameters

-e single node in fifth layer is a fixed node -is nodeperforms the summation of all the incoming signals -eoverall output of ANFIS is given as

Oi5 fout 1113944

i

ωi middot fi ω1f1 + ω2f2

ω1 + ω2 (7)

212 Learning Algorithm of ANFIS -e aim of the learningalgorithm for this architecture is to tune all the modifiableparameters including ai bi ci1113864 1113865 and pi qi ri1113864 1113865 so as tomatch the ANFIS output with the training data In this studyANFIS employs particle swarm optimization (PSO) algo-rithm to adjust the parameters of the membership functions[24] -e PSO techniques have the advantage of being less

Table 1 Some information about the subjects participated in the experiments

Statistical parameterMean plusmn SD

CO ALS HD PDAge (year) 393 plusmn 185 556 plusmn 128 474 plusmn 125 668 plusmn 109Height (m) 183 plusmn 008 174 plusmn 010 184 plusmn 009 187 plusmn 015Weight (kg) 6681 plusmn 1108 7711 plusmn 2115 7347 plusmn 1623 7507 plusmn 169Gait speed (ms) 135 plusmn 016 105 plusmn 022 115 plusmn 035 10 plusmn 02Disease severity 0 183 plusmn 178 679 plusmn 39 28 plusmn 086

A1x

prod

A2

B1

B2

prod

N

N

x

y

y

x y

ω1

ω2 ω2

ω1 ω1middotf1

ω2middotf2

fout

Layer 1 Layer 2 Layer 3 Layer 4 Layer 5

Σ

Figure 1 Architecture of the ANFIS with five layers

Computational and Mathematical Methods in Medicine 3

computationally expensive for a given size of network to-pology [25]

PSO is a swarm intelligence technique first introducedby Eberhart and Kennedy [26] that imitates the socialbehaviour of groups of insects and animals such as swarmsof bees flocks of birds and shoals of fish [27] Empiricalevidence has been accumulated to show that the PSOalgorithm is a useful tool for optimization and it has beenapplied to many optimization problems in engineering[28 29]

Suppose that the searching space is D-dimensional Aswarm of N particles are initialized in which each particle isassigned a random position in the D-dimensional hyper-space Let x denotes a particlersquos position and v denotes theparticlersquos flight velocity over a solution space

-e best historical position of a particle is Pbest -eglobal best historical position among all particles in theswarm is Gbest Velocity and position of a particle areupdated by the following rules

]i(t) ω]i(tminus 1) + c1 times rand1(middot) times xPbestiminusxi(t)( 1113857

+ c2 times rand2(middot) times xGbest minus xi(t)( 1113857

xi(t) xi(tminus 1) + ]i(t)

(8)

where ω is the inertia weight rand1(middot) and rand2(middot) are theuniformly distributed random numbers between 0 and 1and c1 and c2 are the learning rates In this work constants c1and c2 are both set at 20 which was suggested in [24] Alsoas suggested in the same work [24] an inertia correctionfunction called ldquoinertia weight approach (IWA)rdquo is appliedDuring the IWA the inertia weight ω is modified accordingto the following equation

ω ωmax minusωmax minusωmin

ItrmaxItr (9)

where ωmax and ωmin are the initial and final inertia weightsItrmax is the maximum number of iteration and Itr is thecurrent number of iteration

-e particlersquos fitness is calculated by inputting its po-sition into a designed objective function In this study theoptimization function to be minimized is root mean squareerror (RMSE) defined by the following equation

RMSE

1113936Nk1 z(k)minus zM(k)1113858 1113859

2

N

1113971

(10)

where z(k) is the kth actual sample value zM(k) is the targetoutput of the ANFIS model for the input features of the kthsample and N is the amount of training samples

23 Performance Evaluation In the classification stage fora test subject a segment of gait dynamics is inputted to theANFIS model which will produce the same length ofoutputs -e average value of these outputs will be used topredict the final label of the test subject which is determinedaccording to the minimum distance of the average value tothe label of either patients or healthy CO subjects In ourexperiments first each group of ND patients against the

healthy CO subjects was evaluated and then we evaluated allgroups of neurodegenerative diseases against the healthy COsubjects To validate the performance of the proposedmethod the leave-one-out cross-validation (LOOCV)method over the entire dataset was done During LOOCVone subject was left out at a time and used for testing andother remaining subjects were used as training data -e testperformance of the ANFIS model was evaluated by thefollowing statistical measures specificity (Sp) sensitivity(Sn) and classification accuracy (Ca)

Sn TP

TP + FN

Sp TN

TN + FP

Ca TN + TP

TN + TP + FN + FP

(11)

where TP TN FP and FN are true positives true negativesfalse positives and false negatives respectively

3 Results and Discussion

In our experiments the following five different gait pa-rameters are chosen as the inputs to the ANFIS system leftand right stride interval left and right stance interval anddouble support interval Also as mentioned before thereare a few of walking turns that happened when the subjectsreached the end of the hallway In this study only thelongest time series segment between two adjacent walkingturns is utilized And every gait record in such segment istreated as an independent observation to be used as aninput for training the ANFIS system -e statistics of eachfeature for different category of subjects is presented inTable 2

-e final target outputs of ANFIS are designated as0 and 4 indicating healthy CO subjects and ND patientsrespectively With five input variables and two mem-bership functions each the number of rules reaches25 32 As an example the mean square error curve oftraining the ANFIS model for classification between theALS patients and the CO subjects is shown in Figure 2-e initial and final membership functions are shown inFigure 3 Based on the analysis of membership functionsof each input feature it was found that most of them haveconsiderable impact on the final identification of ALSpatient

-e classification results between the ALS patients andthe CO subjects are shown in Table 3 It can be observedthat only one ALS patient and one CO subject werewrongly classified -us the statistical measures of theclassification performance evaluated by the LOOCVmethod are as follows a specificity of 9375 a sensitivityof 9231 and an accuracy of 9310 -e classificationperformances of PD patients vs CO subjects and HDpatients vs CO subjects are listed in Tables 4 and 5 re-spectively -e accuracy for differentiating PD patientsfrom CO subjects is 9032 with two PD patients and one


CO subjects are wrongly classified Similarly in the ex-periments of diagnosing HD patients 19 out of 20 HDpatients and 15 out of 16 CO subjects are correctly

recognized leading to an accuracy of 9444 Finally theresults of all groups of neurodegenerative patients vs COsubjects are reported in Table 6 where a specificity of

Table 2 -e statistics of each feature for different groups of subjects

Feature descriptionMean plusmn SD

ALS PD HD COLeft stride interval (s) 1520 plusmn 0093 1173 plusmn 0052 1146 plusmn 0093 1112 plusmn 0083Right stride interval (s) 1520 plusmn 0095 1173 plusmn 0046 1146 plusmn 0093 1112 plusmn 0077Left stance interval (s) 1043 plusmn 0092 0788 plusmn 0047 0763 plusmn 0072 0737 plusmn 0073Right stance interval (s) 1023 plusmn 0078 0814 plusmn 0052 0800 plusmn 0066 0720 plusmn 0060Double support interval (s) 0546 plusmn 0080 0429 plusmn 0066 0416 plusmn 0071 0345 plusmn 0057

0 100 200 300 400 500Epochs

06

065

07

075

08

085M

ean

squa

re er

ror

Figure 2 -e curve of network error convergence of the ANFIS

08 1 12 14 16 18 2Left stride interval before training

005

1

05 1 15 2 25Right stride interval before training

005

1

05 1 15Left stance interval before training

005

1

06 08 1 12 14 16Right stance interval before training

005

1

0 05 1 15 2 25 3Double support interval before training

005

1

08 1 12 14 16 18 2Left stride interval after training

005

1

08 1 12 14 16 18 2Right stride interval after training

005

1

06 08 1 12 14Left stance interval after training

005

1

06 08 1 12 14 16Right stance interval after training

005

1

0 05 1 15 2 25 3Double support interval after training

005

1

Figure 3 Generalized bell-shaped membership before and after training


8750 a sensitivity of 9167 and an accuracy of 9063are reported

Furthermore another experiment is performed toexplore the performance of the proposed method indiscriminating ND patients with different severity levelsAccording to the dividing approach proposed in [2] theND gait data available in this study are broadly catego-rized into two groups mild lower extremity functionalimpairment and more advanced functional impairment-e severe group consists of 9 PD patients with the Hoehnand Yahr score greater than or equal to 3 9 HD patientswith their total functional capacity score less than or equalto 5 and 9 ALS patients with stride time greater than 12 s-e rest belongs to the group with mild impairmentDiscrimination accuracies were calculated for classifica-tion between the CO group and the two ND groups -eclassification results are shown in Table 7 -e best per-formance (an accuracy of 100) was obtained for the

classification between the CO group and the severe NDgroup -e accuracy for classifying the CO subjects andthe mild ND group is 8649 However for the classifi-cation between the two ND groups ie mild ND group vssevere ND group the statistical measures of the perfor-mance of a sensitivity of 8519 a specificity of 8095and an accuracy of 8333 were still very promising Sucha result indicated that the proposed classification modelcould be used as a potential technique for identifying NDpatients with different levels of severity

-e comparison of the classification performancebetween the proposed algorithm in this study and severalstate-of-the-art classification methods [1 6 7 12 30] onthe same dataset is tabulated in Table 8 It can be notedthat our proposed ANFIS-based method obtained themost accurate classification results in most cases Forexample for the purpose of identifying the ALS patientsrsquogaits from the normal gaits Wu et al [1] extracted featureslike swing interval turns count and averaged stride in-terval By using the least squares support vector machine(LS-SVM) as the classifier they obtained an overall ac-curacy of 8966 which is a bit lower than an accuracy of9310 obtained in the present study Zeng andWang [12]modelled the gait dynamics with the radial basis function(RBF) neural networks and by using the model learnedvia deterministic learning they reported a classificationaccuracy of 871 between PD patients and CO subjectswhile our method presented an accuracy of 9032 -esecompetitive results indicate that the proposed approachcan be effective for the classification of ND patients usinggait dynamics

However as a limitation of this study current gaitdataset available at Physionet is small at its size -ereforefor actual clinical purposes such as gait monitoring ofdisease progression and evaluation of the response oftherapy more ND patients at different severities and agelevels should be recruited into the database in the futurestudies When more ND patients and matched CO subjectswere enrolled into the study the training of the ANFISmodel could bemore effective and thus the proposed systemwould be more instructive for identifying the innate gaitdifferences between ND patients and CO subjects

4 Conclusions

In this study a new classification scheme has been proposedfor the purpose of classifying the gait of ND patients fromnormal gait -e presented ANFIS model combines neuralnetwork adaptive capabilities and the fuzzy logic qualitativeapproach which provides an ideal tool for modelling thenonstationary human gait dynamics Five gait cadence timeseries left stride interval right stride interval left stanceinterval right stance interval and double support intervalwere used as the input variables of the ANFIS model Whenevaluated with the LOOCV method the classification ac-curacies of the ANFIS model for discriminating ND vs COALS vs CO PD vs CO and HD vs CO groups were 90639310 9022 and 9444 respectively By taking intoconsideration the misclassification rates the proposed

Table 3 Confusion matrix for the classification performance of theproposed system on differentiating ALS patients from the COsubjects

Positiveclass(ALS)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP12 1 15 1 9375 9231 9310

Table 4 Confusion matrix for the classification performance of theproposed system on differentiating PD patients from the COsubjects

Positiveclass(PD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP13 2 15 1 9375 8667 9032

Table 5 Confusion matrix for the classification performance of theproposed system on differentiating HD patients from the COsubjects

Positiveclass(HD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP19 1 15 1 9375 9500 9444

Table 6 Confusion matrix for the classification performance of theproposed system on differentiating ND patients from the COsubjects

Positiveclass(ND)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP44 4 14 2 8750 9167 9063


ANFIS model showed the potentials that help in identifyingpatients with ND from CO subjects by analyzing the gaitdata

Data Availability

-e data used to support the findings of this study areavailable at PhysioNet website (httpwwwphysionetorgphysiobankdatabasegaitndd)

Ethical Approval

-e authors would like to inform that no ethical approvalwas required since a deidentified public database was used inthis study

Conflicts of Interest

-ere are no conflicts of interest regarding the publication ofthis paper

Acknowledgments

-is work was supported by Anhui Provincial NaturalScience Foundation (grant number 1608085MF136) theNational Natural Science Foundation of China (NSFC) forYouth (grant number 61701483) the Major Project ofNatural Science of Jiangsu Provincial Department of Edu-cation (17KJA330001) and the 333 Project of JiangsuProvince (BRA2017454)

References

[1] Y Wu and S Krishnan ldquoComputer-aided analysis of gaitrhythm fluctuations in amyotrophic lateral sclerosisrdquoMedicaland Biological Engineering amp Computing vol 47 pp 1165ndash1171 2009

[2] J M Hausdorff A Lertratanakul M E CudkowiczA L Peterson D Kaliton and A L Goldberger ldquoDynamicmarkers of altered gait rhythm in amyotrophic lateral scle-rosisrdquo Journal of Applied Physiology vol 88 no 6pp 2045ndash2053 2000

Table 7 Classification results between the CO group and the two ND groups

Performance metricsResults (negative vs positive)

Mild ND vs severe NDCO vs mild ND CO vs severe ND

Sensibility 1821 8571 2727100 2327 8519Specificity 1416 8750 1616100 1721 8095Accuracy 3237 8649 4343100 4048 8333

Table 8 Performance comparison of several state-of-the-art methods for discriminating ND gaits from normal gaits

Features Classifier Evaluation method Overall accuracy ()

ALS vsCO

Swing-interval turns count averaged strideinterval [1] LS-SVM LOO 8966

Entropy and coherence extracted from thewavelet approximation of the gait signal [6] LDA LOO 862

ANFIS models for left and right stride intervalleft and right stance interval and double

support interval (proposed)Distance rule LOO 9310

PD vsCO

Swing-interval turns count gait rhythmstandard deviation [7] LS-SVM LOO 9032

Constant RBF networks learned viadeterministic learning [12] Distance rule LOO 871

ANFIS models for left and right strideinterval left and right stance interval and

double support interval (proposed)Distance rule LOO 9032

HD vsCO


Statistical features such as minimum maximumaverage and standard deviation [30] SVM Random subsampling 9028



ND vsCO


Constant RBF networks learned viadeterministic learning [12] Distance rule ATAT 9375



LS-SVM least squares support vector machines LDA linear discriminant analysis ATAT all-training-all-testing LOO leave-one-out


[3] B Nefussy and V E Drory ldquoMoving toward a predictive andpersonalized clinical approach in amyotrophic lateral scle-rosis novel developments and future directions in diagnosisgenetics pathogenesis and therapiesrdquo EPMA Journal vol 1no 2 pp 329ndash341 2010

[4] B Goldfarb and S Simon ldquoGait patterns in patients withamyotrophic lateral sclerosisrdquo Archives of Physical Medicineand Rehabilitation vol 65 pp 61ndash65 1984

[5] J M Hausdorff M E Cudkowicz R Firtion J Y Wei andA L Goldberger ldquoGait variability and basal ganglia disordersStride-to-stride variations of gait cycle timing in parkinsonrsquosdisease and Huntingtonrsquos diseaserdquo Movement Disordersvol 13 no 3 pp 428ndash437 1998

[6] E Baratin L Sugavaneswaran K Umapathy C Ioana andS Krishnan ldquoWavelet-based characterization of gait signal forneurological abnormalitiesrdquo Gait Posture vol 41 no 2pp 634ndash639 2015

[7] Y Wu and S Krishnan ldquoStatistical analysis of gait rhythm inpatients with Parkinsonrsquos diseaserdquo IEEE Transactions onNeural Systems and Rehabilitation Engineering vol 18 no 2pp 150ndash158 2010

[8] Y Wu and L Shi ldquoAnalysis of altered gait cycle duration inamyotrophic lateral sclerosis based on nonparametric prob-ability density function estimationrdquo Medical Engineering ampPhysics vol 33 no 3 pp 347ndash355 2011

[9] R C Wagenaar and R E A van Emmerik ldquoDynamics ofmovement disordersrdquo Human Movement Science vol 15no 2 pp 161ndash175 1996

[10] N Stergiou and L M Decker ldquoHuman movement variabilitynonlinear dynamics and pathology is there a connectionrdquoHuman Movement Science vol 30 no 5 pp 869ndash888 2011

[11] F Liao J Wang and P He ldquoMulti-resolution entropyanalysis of gait symmetry in neurological degenerative dis-eases and amyotrophic lateral sclerosisrdquo Medical Engineeringamp Physics vol 30 no 3 pp 299ndash310 2008

[12] W Zeng and C Wang ldquoClassification of neurodegenerativediseases using gait dynamics via deterministic learningrdquo In-formation Sciences vol 317 pp 246ndash258 2015

[13] H Chaoui and P Sicard ldquoAdaptive fuzzy logic control ofpermanent magnet synchronous machines with nonlinearfrictionrdquo IEEE Transactions on Industrial Electronics vol 59no 2 pp 1123ndash1133 2012

[14] N Todorovic and P Klan ldquoState of the art in nonlineardynamical system identification using artificial neural net-worksrdquo in Proceedings of 8th Seminar on Neural NetworkApplications in Electrical Engineering pp 103ndash108 AalborgDenmark September 2006

[15] J S R Jang ldquoANFIS Adaptive-network-based fuzzy inferencesystemrdquo IEEE Transactions on Systems Man and Cyberneticsvol 23 no 3 pp 665ndash685 1993

[16] J S R Jang ldquoSelf-learning fuzzy controllers based on tem-poral backpropagationrdquo IEEE Transactions on Neural Net-works vol 3 no 5 pp 714ndash723 1992

[17] P Gorgel A Sertbas and O N Ucan ldquoA fuzzy inferencesystem combined with wavelet transform for breast massclassificationrdquo in Proceedings of International Conference onTelecommunications and Signal Processing pp 644ndash647Prague Czech Republic July 2012

[18] E D Ubeyli ldquoAutomatic diagnosis of diabetes using adaptiveneuro-fuzzy inference systemsrdquo Expert Systems vol 27 no 4pp 259ndash266 2010

[19] C Gomez R Hornero D Abasolo A Fernandez andJ Escudero ldquoAnalysis of MEG background activity in Alz-heimerrsquos disease using nonlinear methods and ANFISrdquo

Annals of Biomedical Engineering vol 37 no 3 pp 586ndash5942009

[20] Y Xia Q Gao and Q Ye ldquoClassification of gait rhythmsignals between patients with neuro-degenerative diseases andnormal subjects Experiments with statistical features anddifferent classification modelsrdquo Biomed Signal Procesvol 18pp 254ndash262 2015

[21] M M Hoehn and M D Yahr ldquoParkinsonism onset pro-gression and mortalityrdquo Neurology vol 50 no 2 pp 11ndash261998

[22] I Shoulson and S Fahn ldquoHuntington disease clinical careand evaluationrdquo Neurology vol 29 no 1 p 1 1979

[23] M Buragohain and C Mahanta ldquoA novel approach forANFIS modelling based on full factorial designrdquo Applied SoftComputing vol 8 no 1 pp 609ndash625 2008

[24] J Catalatildeo H Pousinho and V Mendes ldquoHybrid wavelet-PSO-ANFIS approach for short-term electricity prices fore-castingrdquo IEEE Transactions on Power Systems vol 26 no 1pp 137ndash144 2011

[25] M A Shoorehdeli M Teshnehlab A K Sedigh andM A Khanesar ldquoIdentification using ANFIS with intelligenthybrid stable learning algorithm approaches and stabilityanalysis of training methodsrdquo Applied Soft Computing vol 9no 2 pp 833ndash850 2009

[26] J Kennedy and R Eberhart ldquoParticle swarm optimizationrdquo inProceedings of International Conference on Neural Networkspp 1942ndash1948 Perth Australia November-December 1995

[27] L Xue J Yin Z Ji and L Jiang ldquoA particle swarm opti-mization for hidden Markov model trainingrdquo in Proceedingsof 8th International International Conference on Signal Pro-cessing IEEE pp 16ndash20 Guilin China November 2006

[28] W Yu and X Li ldquoFuzzy identification using fuzzy neuralnetworks with stable learning algorithmsrdquo IEEE Transactionson Fuzzy Systems vol 12 no 3 pp 411ndash420 2004

[29] B Vasumathi and S Moorthi ldquoImplementation of hybridANNndashPSO algorithm on FPGA for harmonic estimationrdquoEngineering Applications of Artificial Intelligence vol 25no 3 pp 476ndash483 2012

[30] M R Daliri ldquoAutomatic diagnosis of neuro-degenerativediseases using gait dynamicsrdquo Measurement vol 45 no 7pp 1729ndash1734 2012


Stem Cells International

Hindawiwwwhindawicom Volume 2018


MEDIATORSINFLAMMATION

of

EndocrinologyInternational Journal of



Disease Markers


BioMed Research International

OncologyJournal of



Oxidative Medicine and Cellular Longevity


PPAR Research

Hindawi Publishing Corporation httpwwwhindawicom Volume 2013Hindawiwwwhindawicom

The Scientific World Journal

Volume 2018

Immunology ResearchHindawiwwwhindawicom Volume 2018

Journal of

ObesityJournal of



Computational and Mathematical Methods in Medicine


Behavioural Neurology

OphthalmologyJournal of


Diabetes ResearchJournal of



Research and TreatmentAIDS


Gastroenterology Research and Practice


Parkinsonrsquos Disease

Evidence-Based Complementary andAlternative Medicine

Volume 2018Hindawiwwwhindawicom

Submit your manuscripts atwwwhindawicom

found that the ability to maintain a steady gait with lowstride-to-stride variability of gait cycle timing and its sub-phases was diminished for both PD and HD two typicaldisorders of basal ganglia Furthermore they found that gaitspeed was significantly lower in PD but not in HD andaverage gait cycle duration and the time spent in manysubphases of the gait cycle were similar in control (CO)subjects HD subjects and PD subjects In another similarstudy Hausdorff et al [2] studied the magnitude of thestride-to-stride fluctuations and perturbations of the gaitrhythm in subjects with ALS disease in comparison withhealthy CO subjects and PD and HD patients -eir studyrevealed a longer stride interval for ALS patients than all theother three compared groups -ey also found that patientswith ALS have less steady gait between successive strideintervals in comparison with control subjects using thedetrended fluctuation analysis technique

In order to facilitate automated identification of neuro-degenerative diseases like PD HD and ALS for the clinicalassistant purpose by using gait signals different machinelearning tools have also been applied in previous studiesBaratin et al [6] introduced a wavelet-based scheme for ef-fective characterization of gait associated with ALS patients-ey extracted features such as entropy that reflects theregularity of gait and coherence between left and right limbsfrom the wavelet approximation part of the raw gait signal-en the classification of ALS gaits and normal gaits wasimplemented with a linear discriminant analysis (LDA)classifier Wu et al [7] quantified the gait variability in pa-tients with PD by counting the signal turns -ey found thatthe signal turns in the stride interval time series presenteda significant difference between the healthy control subjectsand the PD patients In another similar study by the sameauthors [8] the nonparametric Parzen window method wasutilized to estimate the probability density functions (PDFs)of gait stride interval time series and then by computing themean of left foot stride interval time series and the modifiedKullbackndashLeibler divergence of the PDFs between left footand right foot as the two features they realized the classifi-cation of gait in subjects with ALS and normal subjects byusing the least squares support vector machines (LS-SVM)

-ese studies suggest that gait cadence exhibits complexand nonlinear behaviour in both patients with ND diseasesand control (CO) subjects because of the nonlinear dy-namics of the human system [9 10] However due to thepresence of neurological disorders the ND patientsrsquo gaitrhythm manifests significant differences than that of thehealthy CO subjects such as the magnitude of stride-to-stride gait variability gait speed [2] and the degree ofasymmetry in gait rhythm [11] Previous evidence has shownthat discrimination between normal and abnormal gaitcould be implemented by means of the statistical nonlinearcomputational method [1 6 8 12] ANFIS harnesses thepower of two paradigms fuzzy logic and artificial neuralnetworks which have shown significant results in modellingnonlinear functions [13 14] In the ANFIS the membershipfunction parameters are learned from a dataset that de-scribes the system behaviour under the constraints of a givenerror criterion [15 16] As far as we know ANFIS has been

successfully applied in biomedical engineering for differenttasks such as classification [17 18] and data analysis [19]

In this study a novel approach based on ANFIS ispresented for the classification of gait cadence between thepatients with ND disease and the healthy subjects -eANFIS model combined with the particle swarm optimi-zation (PSO) algorithm is used to learn the nonlinear gaitdynamics by using five gait cadence signals as the inputvariables and then the classification between patients withALS disease and normal subjects is implemented witha simple distance-based classifier by using the outputs of theANFIS model -is paper has been organized as follows thedescription of the gait dataset and an overview of ourproposed method are given in Section 2 -e experimentalresults and some discussions are presented in Section 3Finally some conclusions are given in Section 4

2 Methods

21 Description of Dataset -e gait dataset used in this studywas contributed by Hausdorff et al [2] and it has been usedby several different studies [1 6 8] to investigate the gaitcharacteristics of the patients with neurodegenerative dis-eases To measure the gait signals force-sensitive switcheswere placed in the subjectrsquos shoe -e gait signals weresampled at 300Hz via an onboard analog-to-digital converter(12 bit) -e experimental protocol required that each subjectshould walk at their normal pace along a 77m-long hallwayfor 5min According to the descriptions in [2] the first 20 s ofthe recorded data were excluded to eliminate the start-upeffects and a median filter was applied to remove data pointsthat had a standard deviation (SD) greater than 3 or were lessthan the median value As we had introduced in our previousstudy [20] the gait dataset used in this study provides thefollowing seven gait parameters left and right stride interval(time from initial contact of one foot to the immediatesubsequent contact) left and right stance interval (amount oftime when one foot is on the ground) left and right swinginterval (amount of time when one foot is in the air) anddouble support interval (time of bilateral foot contact)

-irteen ALS patients fifteen PD patients and twentyHD patients were recruited from the Neurology OutpatientClinic at Massachusetts General Hospital Boston MA USAAlso another sixteen healthy CO subjects were enrolled toconduct the comparative study Each subject provided in-formed consent as approved by the Institutional ReviewBoard of the Massachusetts General Hospital -ere was nosignificant difference between the weights and heights of theCO subjects and the ND patients -e ND patients were notusing a wheelchair or assistive device for mobility and werefree of other ailments that might affect lower extremityweakness -e presence or absence of symptoms that mightaffect the gait was determined by a qualified physician -eseverity of ALS is represented by the time since the onset ofthe disease -e Hoehn and Yahr (HampY) scale [21] and thetotal functional capacity (TFC) [22] are used to assess thedegree of neurologic impairment in PD and HD patientsrespectively Basic information of all the subjects partici-pated in this study has been listed in Table 1








(1)




Oi1 μAi

(x) i 1 2

Oi1 μBiminus 2

(y) i 3 4(2)



(x) is given by

μAi(x)

1

1 + xminus ci( 1113857ai1113858 11138592

1113966 1113967bi

(3)



Oi2 ωi μAi




Oi3 ωi

ωi

1113936ωi

ωi

ω1 + ω2for i 1 2 (5)





Oi5 fout 1113944

i


ω1 + ω2 (7)





A1x

prod

A2

B1

B2

prod

N

N

x

y

y

x y

ω1

ω2 ω2

ω1 ω1middotf1

ω2middotf2

fout


Σ










(8)



ItrmaxItr (9)



RMSE

1113936Nk1 z(k)minus zM(k)1113858 1113859

2

N

1113971

(10)




Sn TP

TP + FN

Sp TN

TN + FP

Ca TN + TP

TN + TP + FN + FP

(11)












0 100 200 300 400 500Epochs

06

065

07

075

08

085M

ean

squa

re er

ror



005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1








4 Conclusions



Positiveclass(ALS)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP12 1 15 1 9375 9231 9310


Positiveclass(PD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP13 2 15 1 9375 8667 9032


Positiveclass(HD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP19 1 15 1 9375 9500 9444


Positiveclass(ND)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP44 4 14 2 8750 9167 9063



Data Availability


Ethical Approval




Acknowledgments


References









ALS vsCO





PD vsCO





HD vsCO





ND vsCO









































of




Disease Markers



OncologyJournal of





PPAR Research



Volume 2018


Journal of

ObesityJournal of

























(1)




Oi1 μAi

(x) i 1 2

Oi1 μBiminus 2

(y) i 3 4(2)



(x) is given by

μAi(x)

1

1 + xminus ci( 1113857ai1113858 11138592

1113966 1113967bi

(3)



Oi2 ωi μAi




Oi3 ωi

ωi

1113936ωi

ωi

ω1 + ω2for i 1 2 (5)





Oi5 fout 1113944

i


ω1 + ω2 (7)





A1x

prod

A2

B1

B2

prod

N

N

x

y

y

x y

ω1

ω2 ω2

ω1 ω1middotf1

ω2middotf2

fout


Σ










(8)



ItrmaxItr (9)



RMSE

1113936Nk1 z(k)minus zM(k)1113858 1113859

2

N

1113971

(10)




Sn TP

TP + FN

Sp TN

TN + FP

Ca TN + TP

TN + TP + FN + FP

(11)












0 100 200 300 400 500Epochs

06

065

07

075

08

085M

ean

squa

re er

ror



005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1








4 Conclusions



Positiveclass(ALS)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP12 1 15 1 9375 9231 9310


Positiveclass(PD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP13 2 15 1 9375 8667 9032


Positiveclass(HD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP19 1 15 1 9375 9500 9444


Positiveclass(ND)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP44 4 14 2 8750 9167 9063



Data Availability


Ethical Approval




Acknowledgments


References









ALS vsCO





PD vsCO





HD vsCO





ND vsCO









































of




Disease Markers



OncologyJournal of





PPAR Research



Volume 2018


Journal of

ObesityJournal of


























(8)



ItrmaxItr (9)



RMSE

1113936Nk1 z(k)minus zM(k)1113858 1113859

2

N

1113971

(10)




Sn TP

TP + FN

Sp TN

TN + FP

Ca TN + TP

TN + TP + FN + FP

(11)












0 100 200 300 400 500Epochs

06

065

07

075

08

085M

ean

squa

re er

ror



005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1








4 Conclusions



Positiveclass(ALS)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP12 1 15 1 9375 9231 9310


Positiveclass(PD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP13 2 15 1 9375 8667 9032


Positiveclass(HD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP19 1 15 1 9375 9500 9444


Positiveclass(ND)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP44 4 14 2 8750 9167 9063



Data Availability


Ethical Approval




Acknowledgments


References









ALS vsCO





PD vsCO





HD vsCO





ND vsCO









































of




Disease Markers



OncologyJournal of





PPAR Research



Volume 2018


Journal of

ObesityJournal of
























0 100 200 300 400 500Epochs

06

065

07

075

08

085M

ean

squa

re er

ror



005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1


005

1








4 Conclusions



Positiveclass(ALS)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP12 1 15 1 9375 9231 9310


Positiveclass(PD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP13 2 15 1 9375 8667 9032


Positiveclass(HD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP19 1 15 1 9375 9500 9444


Positiveclass(ND)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP44 4 14 2 8750 9167 9063



Data Availability


Ethical Approval




Acknowledgments


References









ALS vsCO





PD vsCO





HD vsCO





ND vsCO









































of




Disease Markers



OncologyJournal of





PPAR Research



Volume 2018


Journal of

ObesityJournal of
























4 Conclusions



Positiveclass(ALS)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP12 1 15 1 9375 9231 9310


Positiveclass(PD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP13 2 15 1 9375 8667 9032


Positiveclass(HD)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP19 1 15 1 9375 9500 9444


Positiveclass(ND)

Negativeclass(CO)

Specificity(Sp)

Sensitivity(Sn)

Accuracy(Ca)

TP FN TN FP44 4 14 2 8750 9167 9063



Data Availability


Ethical Approval




Acknowledgments


References









ALS vsCO





PD vsCO





HD vsCO





ND vsCO









































of




Disease Markers



OncologyJournal of





PPAR Research



Volume 2018


Journal of

ObesityJournal of




















Data Availability


Ethical Approval




Acknowledgments


References









ALS vsCO





PD vsCO





HD vsCO





ND vsCO









































of




Disease Markers



OncologyJournal of





PPAR Research



Volume 2018


Journal of

ObesityJournal of





















































of




Disease Markers



OncologyJournal of





PPAR Research



Volume 2018


Journal of

ObesityJournal of



















ClassificationofGaitPatternsinPatientswithNeurodegenerativ...

Documents

Transcript of ClassificationofGaitPatternsinPatientswithNeurodegenerativ...