Research Article ECG Prediction Based on Classification via …downloads.hindawi.com/journals/tswj/2015/205749.pdf · 2019-07-31 · Research Article ECG Prediction Based on Classification
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Research ArticleECG Prediction Based on Classification via Neural Networks andLinguistic Fuzzy Logic Forecaster
Eva Volna Martin Kotyrba and Hashim Habiballa
University of Ostrava 30 Dubna 22 70103 Ostrava Czech Republic
Correspondence should be addressed to Martin Kotyrba martinkotyrbaosucz
Received 16 July 2014 Accepted 20 November 2014
Academic Editor Mohammed Chadli
Copyright copy 2015 Eva Volna et al This 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 is properly cited
The paper deals with ECG prediction based on neural networks classification of different types of time courses of ECG signals Themain objective is to recognise normal cycles and arrhythmias and perform further diagnosis We proposed two detection systemsthat have been created with usage of neural networks The experimental part makes it possible to load ECG signals preprocessthem and classify them into given classes Outputs from the classifiers carry a predictive character All experimental results fromboth of the proposed classifiers aremutually compared in the conclusionWe also experimented with the newmethod of time seriestransparent prediction based on fuzzy transform with linguistic IF-THEN rules Preliminary results show interesting results basedon the unique capability of this approach bringing natural language interpretation of particular prediction that is the propertiesof time series
1 Background
Biometrical data is typically represented as an image ora quantification of measured physiological or behaviouralcharacteristics As this data should refer to very complexhuman behaviour or describe very precisely physiologicalcharacteristic (typically iris scan fingerprint palm veinimage hand scan voice walk pattern etc) this data caneasily become very large and hard to process For this reasonmodern ways of data processing and classification are appliedfor biometrical data The leading method is the usage ofneural networks [1]
For more than four decades computers have been usedin the classification of the electrocardiogram (ECG) resultingin a huge variety of techniques [2] all designed to enhancethe classification accuracy to levels comparable to that of aldquogold standardrdquo of expert cardiology opinion Included inthese techniques are multivariate statistics decision treesfuzzy logic expert systems and hybrid approaches [3] Therecent interest in neural networks coupled with their highlevels of performance has resulted in many instances of theirapplication in this field [4]
The electrocardiogram is a technique of recording bio-electric currents generated by the heart Clinicians canevaluate the conditions of a patientrsquos heart from the ECG andperform further diagnosis ECG records are obtained by sam-pling the bioelectric currents sensed by several electrodesknown as leads A typical one-cycle ECG tracing is shown inFigure 3
11 Backpropagation Neural Networks A neural networkis a parallel distributed information processing structureconsisting of processing elements (which can possess a localmemory and can carry out localized information process-ing operations) interconnected together with unidirectionalsignal channels called connections Each processing elementhas a single output connection which branches into as manycollateral connections as desired (each carrying the samesignal the processing element output signal) The processingelement output signal can be of any mathematical typedesired All of the processing that goes on within eachprocessing element must be completely local that is it mustdepend only upon the current values of the input signals
Hindawi Publishing Corporatione Scientific World JournalVolume 2015 Article ID 205749 10 pageshttpdxdoiorg1011552015205749
2 The Scientific World Journal
x1
x2
f1(e)
f2(e)
f3(e)
f4(e)
f5(e)
f6(e) y
Figure 1 A backpropagation network architecture
arriving at the processing element via impinging connectionsand upon values stored in the processing elementrsquos localmemory [5]
The backpropagation neural network architecture is ahierarchical design consisting of fully interconnected layersor rows of processing units (with each unit itself comprisedof several individual processing elements) Backpropagationbelongs to the class of mapping neural network architecturesand therefore the information processing function that itcarries out is the approximation of a bounded mappingor function 119891 119860 sub 119877119899 rarr 119877119898 from a compactsubset A of n-dimensional Euclidean space to a boundedsubset 119891[A] of m-dimensional Euclidean space by meansof training on examples (119909
1 1199111) (1199092 1199112) (119909
119896 119911119896) It
will always be assumed that such examples of a mapping 119891are generated by selecting xk vectors randomly from A inaccordance with a fixed probability density function 119901(x)The operational use to which the network is to be put aftertraining is also assumed to involve random selections ofinput vectors x in accordancewith119901(x)The backpropagationarchitecture described in this paper is the basic classicalversion (Figure 1) The backpropagation learning algorithmis composed of two procedures (a) forward propagation ofsignals and (b) backpropagation weight training [5]
Feed-Forward Assume that each input factor in the inputlayer is denoted by 119909
119894 the 119910
119895and 119911
119896represent the output in
the hidden layer and the output layer respectively And the 119910119895
and 119911119896can be expressed as follows (1)
119910119895= 119891 (119883
119895) = 119891(119908
119900119895+119868
sum119894=1
119908119894119895119909119894)
119911119896= 119891 (119884
119896) = 119891(119908
119900119896+119869
sum119895=1
119908119894119896119910119895)
(1)
where the 119908119900119895
and 119908119900119896
are the bias weights for settingthreshold values 119891 is the activation function used in bothhidden and output layers and 119883
119895and 119884
119896are the temporarily
computing results before applying activation function 119891 Inthis study a sigmoid function is selected as the activation
function Therefore the actual outputs 119910119895and 119911
119896in hidden
and output layers respectively can be also written as
119910119895= 119891 (119883
119895) =
1
1 + 119890minus119883119895
119911119896= 119891 (119884
119896) =
1
1 + 119890minus119884119896
(2)
The activation function 119891 introduces the nonlinear effect tothe network and maps the result of computation to a domain(0 1) This sigmoid function is differentiable The derivativeof the sigmoid function in (2) can be easily derived as 1198911015840 =119891(1 + minus119891)
Backpropagation Weight Training The error function isdefined as
119864 =1
2
119870
sum119896=1
1198902119896=119870
sum119896=1
(119905119896minus 119911119896)2
(3)
where 119905119896is a predefined network output (or desired output or
target value) and 119890119896is the error in each output nodeThe goal
is to minimize 119864 so that the weight in each link is accordinglyadjusted and the final output can match the desired outputTo get the weight adjustment the gradient descent strategyis employed In the link between hidden and output layerscomputing the partial derivative of 119864 with respect to theweight 119908
where 120575119896= (119905119896minus 119911119896) 1198911015840 (119884
119896)
(4)
Theweight adjustment in the link between hidden and outputlayers is computed by Δ119908
119895119896= 120572 times 119910
119895times 120575119896 where 120572 is
the learning rate a positive constant between 0 and 1 Thenew weight herein can be updated by the following 119908
119895119896(119899 +
1) = 119908119895119896(119899) + Δ119908
119895119896(119899) where 119899 is the number of iterations
Similarly the error gradient in links between input andhidden layers can be obtained by taking the partial derivativewith respect to 119908
119894119895as
120597119864
120597119908119894119895
= minusΔ119895119909119895= 1198911015840 (119883
119895)119870
sum119896=1
120597119896119908119895119896 (5)
The new weight in the hidden-input links can be nowcorrected as Δ119908
119894119895= 120572 times 119909
119894times Δ119895and 119908
119894119895(119899 + 1) = 119908
119894119895(119899) + Δ
119895
Training the BP-networks with many samples is sometimesa time-consuming task The learning speed can be improvedby introducing the momentum term 120578 Usually 120578 falls in therange ⟨0 1⟩ For the iteration 119899 the weight change Δ119908 can beexpressed The backpropagation learning algorithm used inartificial neural networks is shown in many text books [3ndash6]
12 Fuzzy Logic Fuzzy logics form heterogeneous family offormalisms capable of successful modelling of uncertain andvague information processing [7] The usage of fuzzy logicfor analysis and prediction of time series can be perceived
The Scientific World Journal 3
as a complement method to neural network based methodsThe symbolic background of fuzzy logic brings an advantageof human readable symbolic representation of predictioninterpretation It does not necessarily mean that fuzzy logicbased time series analysis is more accurate and more efficientbut its power lies in transparent and interpretable results thatit gives [8ndash11]
Time series analysis and prediction are an importanttask that can be used in many areas of practice The task ofgetting the best prediction to given series may bring inter-esting engineering applications in wide number of areas likeeconomics geography or industry Solution to the problemof obtaining best results in prediction of time series can bebased on well-known and simple methods like Winters orLinear method In this paper we use a method based on twomethods originally developed by members of Institute forResearch andApplications of FuzzyModeling which is a partofUniversity ofOstravaThe aimof the paper is not to presentthe details of the methods already published but to presenta tool implementing them The first method is based on thenotion of F-transform (fuzzy transform) devised by the groupof Professor Perfilieva et al [12] The second approach usesthe linguistic rules utilizing fuzzy logic and deduction that isa well-known formalism with very good results in variety ofpractical applications like industrial ones
The idea of the fuzzy transform is to transform a givenfunction defined in one space into another usually simplerspace and then to transform it back The simpler spaceconsists of a finite vector of numbers The reverse transformthen leads to a function which approximates the original oneMore details can be found in [12]
The fuzzy transform is defined with respect to a fuzzypartition which consists of basic functions Let 119888
1lt sdot sdot sdot lt 119888
119899
be fixed nodes within [119886 119887] such that 1198881
= 119886 119888119899
= 119887 and119899 ge 2 We say that fuzzy sets 119860
1 119860
119899isin 119865([119886 119887]) are basic
functions forming a fuzzy partition of [119886 119887] if they fulfill thefollowing conditions for 119894 = 1 119899
(1) 119860119894(119888119894) = 1
(2) 119860119894(119909) = 0 for 119909 isin (119888
119894minus1 119888119894+1
) where for uniformity ofnotation we put 119888
0= 1198881= 119886 and 119888
119899+1= 119888119899= 119887
(3) 119860119894is continuous
(4) 119860119894strictly increases on [119888
119894minus1 119888119894] and strictly decreases
on [119888119894 119888119894+1
](5) for all 119909 isin [119886 119887]
119899
sum119894=1
119860119894(119909) = 1 (6)
Let a fuzzy partition of [119886 119887] be given by basic functions1198601 119860
119899 119899 ge 2 and let 119891 [119886 119887] rarr 119877 be a function that is
is a fuzzy transform of 119891 with respect to the given fuzzypartition
The numbers 1198651 119865
119899are called the components of the
fuzzy transform of 119891Let 119865119899[119891] be the fuzzy transform of 119891 with respect to
1198601 119860
119899isin 119865([119886 119887])
Then the function 119891119865119899
given on [119886 119887] by
119891119865119899
(119909) =119899
sum119894=1
119865119894119860119894(119909) (8)
is called the inverse fuzzy transform of 119891Fuzzy IF-THEN rules can be understood as a specific
conditional sentence of natural language of the form IF1198831is
1198601AND sdot sdot sdot AND 119883
119899is 119860119899THEN 119884 is 119861 where 119860
1 119860
119899
and 119861 are evaluative expressions (very small roughly bigetc) An example fuzzy IF-THEN rule is as follows
IF the number of cars sold in the current year is more or lesssmall and the half-year sales increment is medium THEN theupcoming half-year increment will be medium
The part of the rule before THEN is called the antecedentand the part after it is consequent Fuzzy IF-THEN rules areusually gathered in a linguistic description
1198771= IF 119883
1is 11986011
AND sdot sdot sdotAND 119883119899
is 1198601119899
THEN 119884 is 1198611
119877119898
= IF 1198831is 1198601198981
AND sdot sdot sdotAND 119883119899
is 119860119898119899
THEN 119884 is 119861119898
(9)
Time series prediction based on these two mainapproaches works as follows Let time series 119909
119905 119905 = 1 119879
be viewed as a discrete function 119909 on a time axis 119905 Then119865119899[119909] = [119883
1 119883
119899] is the fuzzy transform of the function
119909 with respect to a given fuzzy partition The inverse fuzzytransform then serves us as a model of the trend-cycle ofa given time series By subtracting the trend-cycle (inversefuzzy transform) values from the time series lags we get pureseasonal components This is how the fuzzy transform helpsus to model and decompose a given time series
Logical dependencies between components1198831 119883
119899of
the fuzzy transform may be described by the fuzzy rulesThese rules are generated automatically from the given dataand are used for forecasting the next components Fuzzytransform components as well as their first and second orderdifferences are used as antecedent variables For forecastingfuture fuzzy transform components based on the generatedfuzzy rules a special inference methodmdashperception-basedlogical deduction is used The seasonal components are fore-casted autoregressively Finally both forecasted componentstrend-cycle and seasonal are composed together to obtainthe forecast of time series lags These methods are integratedinto an implementation PC application called linguistic fuzzylogic forecaster (LFLF) which enables as to produce linguisticdescriptions that describe properties of data treated like atime series
4 The Scientific World Journal(m
V)
+20
0
minus90
0 100 200 300 400
(ms)
Figure 2 The cardiac action potentials
2 Basic Principles of ECG Evaluation
ECG scanning has its own rules which are in accordancewith the laws of physics The heart irritation spreads inall directions In the case that the depolarisation spreadstowards the electrode which is placed on the body surfacea positive deflection is recorded on an ECG monitor Anegative deflection is recorded at the opposite end of thebody The ECG waveform is written with a chart speed of25mmsdotsminus1 An algorithm describing the curve goes in thefollowing steps First we evaluate the shape and rhythm ofventricular complexes or atrial which can be either regularor irregular Then we evaluate the frequency of ventricularcomplexes and atrial fibrillations Contraction of eachmuscleof the human body (and thus the heart as well) is associatedwith electrical changes called depolarization which can bedetected by electrodes The heart contains two basic types ofcells myocardial cells which are responsible for generatingthe pressure necessary to pump blood throughout the bodyand conduction cells which are responsible for rapidlyspreading electrical signals to the myocardial cells in orderto coordinate pumping A graph of an action potential of amuscle of cardiac cells is shown in Figure 2
A normal electrocardiogram is illustrated in Figure 3The figure also includes definitions for various segmentsand intervals in the ECG The deflections in this signalare denoted in alphabetic order starting with the letterP which represents atrial depolarization The ventriculardepolarization causes the QRS complex and repolarizationis responsible for the T-wave Atrial repolarization occursduring the QRS complex and produces such a low signalamplitude that it cannot be seen apart from the normal ECG
3 Signal Processing Using NeuralNetworks and Fuzzy Logic
In practice a relatively reliable diagnostic program storedin ECG monitors has been used which is a guideline fordetermining the final diagnosis of heart disorders This pro-gramworks according to the principle of IF-THEN rulesThevalues of the electrical signal are discretized and uploaded
P
R
TU
Q
QRSinterval
interval interval
segment segment
S
P-R
P-R
S-T
S-T
Figure 3 A typical one-cycle ECG tracing (adapted fromhttpwwwnicomwhite-paperapplargeimagelang=csampimageurl=2Fcms2Fimages2Fdevzone2Ftut2F2007-07-09141618jpg)
0
500
1000
1500
2000
1 101 201 301
Healthy personsSick persons
minus500
Figure 4 Comparison of mean values of ECG waveforms forhealthysick persons
into expert systems in the form of thousand rules The aim ofthis paper is to use a different approach based on the principleof neural networks The proposed methodology of solutioncould be summarized into the following steps
(1) a conversion of analog signal from the ECG monitorto a computer
(2) using multilayer networks that are fully connected(3) obtaining ECG waveforms in collaboration with the
University Hospital in Poruba specifically at theDepartmentCardiac Surgery from sick patients and atthe Department Traumatology from healthy patients(ie ldquohealthyrdquo with regard to heart diseases)
31 Technical Equipment ECG measurements were per-formed using ADDA Junior with converter ADDA Junior
The Scientific World Journal 5
0
10
20
30
40
50
0-01 01-02 04-05
()
Test error
Healthy personsSick persons
02ndash04 gt05
Figure 5 Experimental results test error for healthysick persons
002040608
112
1 2 3 4 5 6 7 8 9 10
H1H2
H3H4
Figure 6 Patterns representing healthy persons
which was connected to a computer via bidirectional parallelcable (CETRONICS) Technical parameters of the AD con-verter (8-bit conversion) were the following
(i) 3 measuring ranges(ii) measuring of a frequency of AC voltage at any
channel(iii) autoranging for measuring the frequency of 100Hz
1 kHz and 10 kHz(iv) input resistance of 300 kΩ(v) measurement accuracy 1
Technical parameters of the DA converter (a pro-grammable voltage source plusmn10V) were the following
(i) maximum current consumption of 15mA (after opti-mizing 4A at the output)
(ii) power of the converter plusmn15 V (stabilized)
4 Experimental Results
41 Time Series Classification and Prediction via NeuralNetworks The training set consisted of modified ECG wave-forms We used a backpropagation neural network withtopology 101-10-1 The output unit represents a diagnose 01a healthysick person A smaller number of inputs would
002040608
112
1 2 3 4 5 6 7 8 9 10
S1S2
S3S4
Figure 7 Patterns representing sick persons
0
500
1000
1500
1 26 51 76 101
Sick persons
S2S3
S4
minus500
Figure 8 Some recognized patterns that occur in ECG time series
not be appropriate due to the nature of the ECG waveformWe use 34 ECG time series associated with sick personsand 36 ECG time series associated with healthy persons25 time series of each group were used as a training setand the rest as a test set Figure 4 shows a comparison ofmean values of ECG waveforms for healthysick persons Weused the backpropagation method [5 6] for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 The conducted experimental studiesalso showed that training patterns are mixed randomly ineach cycle of adaptation This ensures their greater diversitywhich acts as a measure of system stability Uniform systemin a crisis usually collapses entirely while system with suchdiversity of trained patterns remains functional despite ofcrisis of its individual parts The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
The test set consisted of 20 samples (11 health and 9sick persons) that were not included in the training set Thesummary results for this type of experiment are shown in agraph in Figure 5 For clarity the results of testing are givenin percentage The average test error was 0194 A healthypopulation was detected with an average error of 0263 andsick population with an average error of 0109
411 Pattern Recognition Classifier Leading to Prediction Forthe purpose of adaptation of the pattern recognition classifierit is necessary to remark that determination of trainingpatterns is one of the key tasks Improperly chosen patternscan lead to confusion of neural networks During our exper-imental work we made some study which included ECGpattern recognition When creating appropriate patterns ofthe training set we used characteristic curves shown as mean
6 The Scientific World Journal
0100
1 2 3 4 5 6 7 8 9 10
S2 trainS2 test
S2 test
minus500
minus400
minus300
minus200
minus100
050
100150200250
1 2 3 4 5 6 7 8 9 10
S3 trainS3 test
S3 test
minus150
minus100
minus50
0100200300400500600
1 2 3 4 5 6 7 8 9 10
Train H3Test H3
minus200
minus100
Figure 9 Training patterns their representation in used test sets
values from ECG waveforms for healthy and sick persons(Figure 4) We use two different groups of patterns PatternsH1ndashH4 (Figure 6) represent healthy persons and patterns S1ndashS4 (Figure 7) represent sick personsThe whole training set isshown in Table 1
Pattern recognition classifier is based on backpropagationneural network and is able to recognise wave structures ingiven time series [13 14] Artificial neural networks needtraining sets for their adaptation In our experimental workthe training set consisted of 8 patterns representing thebasic structure of the various waves in ECG graphs seeFigures 6 and 7 Input data is sequences always including119899 consecutive numbers which are transformed into interval⟨0 1⟩ by formula (10) Samples are adjusted for the needs ofbackpropagation networks with sigmoid activation functionin this way [5 6]
1199091015840119895=
119909119895minusmin (119909
119894 119909
119894+119899minus1)
max (119909119894 119909
119894+119899minus1) minusmin (119909
119894 119909
119894+119899minus1)
(119895 = 119894 119894 + 119899 minus 1)
(10)
where 1199091015840119895is normalized output value of the 119895th neuron (119895 =
119894 119894 + 119899minus1) and (119909119894 119909119894+119899minus1
) are 119899minus1 consecutive outputvalues that specify sequences (patterns) from the trainingset (eg training pars of input and corresponding outputvectors) Input vector contains 10 components Output vectorhas got 8 components and each output unit represents oneof 8 different types of ECG wave samples A neural networkarchitecture is 10-10-8 (eg 10 units in the input layer 10 unitsin the hidden layer and 8 units in the output layer)The net isfully connected Adaptation of the neural network starts withrandomly generated weight values
We used the backpropagation method for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 We have utilized our experience fromearlier times that is training patterns were mixed randomlyin each cycle of adaptation The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
In order to test the efficiency of the method we appliedthe same set of data that we used in the previous experimentalpart Outputs from the classifier produce sets of values thatare assigned to each recognized training pattern in the giventest time series It is important to appreciate what can beconsidered as an effective criterion related to consensusof similarity The proposed threshold resulting from ourexperimental study was determined at least 119901 = 70Figure 9 shows a comparison of patterns how were learned(S2 S3 H3 train) and how were recognized in test time series(S2 S3 H3 test) The neural network is able to discover someconnections which are almost imperceptible Illustration ofsome recognized patterns that occur in ECG time seriesis shown in Figure 8 Outputs from the classifier carry apredictive character The neural network determines if thetime series belongs to a healthy or sick person on the basis ofthe recognised ECG patterns which appear in the time serieshistory
The methodology of testing is shown in Figure 10 Thismeans that if the test pattern S1 S2 S3 or S4 appearedin ECG waveform with probability 119901S ge 119901 (119901 = 70)thus it was predicted to be ldquoa sick personrdquo Then we workonly with the remaining time series If the test pattern H1H2 H3 or H4 appeared in ECG waveform with probability119901H ge 119901 (119901 = 70) thus it was predicted to be ldquoahealthy personrdquo In all other cases the ECG time series wasunspecifiedWe examined a total of 20 data sets Each of them
contains 101 values that assign 92 possible patternsThewholenumber of examined patterns is 1840 The graph in Figure 11demonstrates a summary of results where ldquosick personsrdquorepresent patterns S1ndashS4 and ldquohealthy personsrdquo representpatterns H1ndashH4 The resulting prediction is based on themethodology see Figure 10
Figure 12 LFLF application
Figure 13 Winning predictor linguistic description (trend-cyclemodel)
42 Time Series Classification and Prediction via LinguisticFuzzy Logic Forecaster We tried also to utilize above pre-sentedmethod of time series analysis through linguistic fuzzylogic forecaster (LFLF) [15] see Figure 12
Basic usage of the application is to analyse given timeseries and find best predictor with respect to validation partof time series given We evaluate efficiency of predictorsby SMAPE (symmetric mean absolute percentage error) Itenables us to make analysis of trend-cycle of a time seriesand also seasonal part The main advantage lies in predictionbased on transparent linguistic descriptions that provide themodel of a time series behaviour Linguistic variables are ofthe following types
(i) value we directly mean the components of the fuzzytransform
(ii) difference first order differences of fuzzy transformcomponents that are given as follows differencesbetween components Δ119883
119894= 119883119894minus 119883119894minus1
8 The Scientific World Journal
Healthypattern
typical series (HS)
Sick patterntypical series
(SS)
Tested series (TS)
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
TS is ldquohealthyrdquo TS is ldquosickrdquo
Yes No
Find best predictorfor TS as validation
interval
Find best predictorfor TS as validation
interval
Concatenate series (SS + TS)Concatenate series (HS + TS)
+ TS) forSMAPE (HS SMAPE (SS + TS) for
If SMAPE (HS + TS) lt SMAPE (SS + TS)
Figure 14 Recognition by linguistic fuzzy logic predictors for typical learning series
(iii) second difference these are values of second orderdifferences of components of the fuzzy transform asfollows Δ2119883
119894= Δ119883
119894minus Δ119883119894minus1
LFLF application enables us to define minimal andmaximal number of these particular variables in a ruleof linguistic description as well as the total number ofantecedent variables
A rule consisting of these variables has the followingstructure and can be described as a signature (fuzzy rulesdescribing the trend-cycle model) Particularly 119878 denotes the
trend-cycle components 119889119878 their differences and 1198892119878 theirsecond order differences The argument (119905) (119905 minus 1) and soforth denotes the time lag of the component
For example taking signature 119878(119905)amp119889119878(119905) rarr 119889119878(119905 + 1)denotes the fact that119883
119894and Δ119883
119894are the antecedent variables
and Δ119883119894+1
is the consequent variable of the winning modeland hence we deal with rules of the form
IF 119883119894is 119860119894AND Δ119883
119894is 119860Δ119894
THEN Δ119883119894+1
is 119860Delta119894(11)
The Scientific World Journal 9
Table 2 Example of algorithm evaluation on 10 ldquohealthyrdquo and 10 ldquosickrdquo patients
Every single fuzzy rule can be taken as a sentence ofnatural language for example first rule from Figure 13
IF 119883119894is ml sm AND Δ119883
119894is qr sm THEN Δ119883
119894+1is minusme
may be read as followsIf the number of cars sold in the current year is more or
less small and the half-year sales increment is quite roughlysmall then the upcoming half-year increment will be negativemedium
421 Recognition of ldquoHealthyrdquo and ldquoSickrdquo Patterns by LFLFOur method to use linguistic fuzzy logic forecasting isbased on simple idea that best predictor learning from bothldquohealthyrdquo and ldquosickrdquo pattern samples respectively can be usedfor validation with tested pattern taken as validation part ofthe seriesThen we can evaluate SMAPE for both these casescompound series SMAPE (ldquohealthyrdquo + tested) and SMAPE(ldquosickrdquo + tested)
If SMAPE (ldquohealthyrdquo + tested) lt SMAPE (ldquosickrdquo + tested)then the tested pattern is supposed to be ldquohealthyrdquo otherwisethe tested pattern is supposed to be ldquosickrdquo
The idea is schematically shown in Figure 14For testing purposes we created two necessary typical
learning time series ldquohealthyrdquo (HS) and ldquosickrdquo (SS) accordingto the algorithm above They both consist of 1010 samplesmade from 10 typical series of ldquohealthyrdquo and ldquosickrdquo patientswith 101 measured ECG values Then we have created 10concatenated series according to the scheme in Figure 14with 10 randomly selected patients with ldquohealthyrdquo ECGmeasurement that is 20 files were produced (10x HS + TSand 10x SS + TS) The same concatenated series were alsomade from 10 ldquosickrdquo patients measurements This made usadditional 20 files with concatenated series (10x HS + TS and
0102030405060708090
100
Sick Healthy
FalseTrue
()
Figure 15 Experimental results LFLF
10x SS + TS) For 20 patients (Table 2) tested ECG we have2 concatenated series giving SMAPE (HS + TS) and SMAPE(SS + TS)
Our method based on LFLF proved very good results forright identification of sick patient records Nevertheless itproduces large amount of false positive identification of sickpattern for healthy patients (Figure 15) This result is con-sistent with our approach using neural networks Of courseour preliminary research has a limited extent and should beperceived only as narrative result which shows interestingproperties especially in complementation of neural networkresults
10 The Scientific World Journal
5 Conclusion
In this paper a short introduction into the field of ECGwaves recognition using backpropagation neural network hasbeen given Main objective was to recognise the normalcycles and arrhythmias and perform further diagnosis Weproposed two detection systems that have been created withusage of neural networks One of them is adapted accordingto the training set Here each pattern represents the wholeone ECG cycle Then an output unit represents a diagnose01 a healthysick person The second one approach usesneural network in which training set contains two differentgroups of patterns for healthysick persons According to theresults of experimental studies it can be stated that ECGwaves patterns were successfully extracted in given timeseries and recognised using suggestedmethod as can be seenfrom figures in Experimental Result section It might resultin better mapping of the time series behaviour for betterprediction
Both approaches were able to predict with high probabil-ity if the ECG time series represents sick or healthy persons Itis interesting that a sick diagnose was recognised with higheraccuracy in both experimental works
The third approach based on LFLF is currently only inthe stage of preliminary experiments but it conforms to theformer results based on neural networks This approach isnovel and could be good supplement to other soft-computingmethods for this task
Disclaimer
Anyopinions findings and conclusions or recommendationsexpressed in this material are those of the authors and do notnecessarily reflect the views of the sponsors
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
The research described here has been financially sup-ported by University of Ostrava Grants SGS16PrF2014 andSGS14PrF2014
References
[1] P Tangkraingkij C Lursinsap S Sanguansintukul and TDesudchit ldquoSelecting relevant EEG signal locations for personalidentification problem using ICA and neural networkrdquo inProceedings of the 8th IEEEACIS International Conference onComputer and Information Science (ICIS rsquo09) pp 616ndash621 June2009
[2] C D Nugent J A C Webb N D Black and G T H WrightldquoElectrocardiogram 2 classificationrdquo Automedica vol 17 pp281ndash306 1999
[3] S Russell and P Norvig Artificial IntelligencemdashA ModernApproach Prentice Hall 2nd edition 2003
[4] C D Nugent J A Lopez A E Smith and N D BlackldquoReverse engineering of neural network classifiers in medicalapplicationsrdquo in Proceedings of the 7th EFOMP Congress vol 17p 184 Physica Medica 2001
[5] L Fausett Fundamentals of Neural Network PrenticeHall 1994[6] D E Rumelhart G E Hinton and R J Williams ldquoLearning
[7] V Novak Fuzzy Sets and Their Applications Adam HilgerBristol UK 1989
[8] S Aouaouda M Chadli P Shi and H R Karimi ldquoDiscrete-time119867
minus119867infinsensor fault detection observer design for nonlin-
ear systems with parameter uncertaintyrdquo International Journalof Robust and Nonlinear Control 2013
[9] M Chadli A Abdo and S X Ding ldquo119867minus119867infin
fault detectionfilter design for discrete-time Takagi-Sugeno fuzzy systemrdquoAutomatica vol 49 no 7 pp 1996ndash2005 2013
[10] M Chadli and H R Karimi ldquoRobust observer design forunknown inputs takagi-sugeno modelsrdquo IEEE Transactions onFuzzy Systems vol 21 no 1 pp 158ndash164 2013
[11] S Aouaouda M Chadli and H Karimi ldquoRobust static output-feedback controller design against sensor failure for vehicledynamicsrdquo IET Control Theory amp Applications vol 8 no 9 pp728ndash737 2014
[12] I Perfilieva V Novak V Pavliska A Dvorak andM StepnickaldquoAnalysis and prediction of time series using fuzzy transformrdquoin Proceedings of the International Joint Conference on NeuralNetworks (IJCNN rsquo08) pp 3876ndash3880 Hong Kong June 2008
[13] E Volna M Kotyrba and R Jarusek ldquoMulti-classifier basedon Elliott waversquos recognitionrdquo Computers amp Mathematics withApplications vol 66 no 2 pp 213ndash225 2013
[14] E Volna M Kotyrba and R Jarusek ldquoPrediction by means ofElliott waves recognitionrdquo in Nostradamus Modern Methods ofPrediction Modeling and Analysis of Nonlinear Systems vol 192of Advances in Intelligent Systems and Computing pp 241ndash250Springer Berlin Germany 2013
[15] H Habiballa V Pavliska and A Dvorak ldquoSoftware systemfor time series prediction based on F-transform and linguisticrulesrdquo in Proceedings of the 8th International Conference onApplied Mathematics Aplimat pp 381ndash386 Bratislava Slovakia2009
arriving at the processing element via impinging connectionsand upon values stored in the processing elementrsquos localmemory [5]
The backpropagation neural network architecture is ahierarchical design consisting of fully interconnected layersor rows of processing units (with each unit itself comprisedof several individual processing elements) Backpropagationbelongs to the class of mapping neural network architecturesand therefore the information processing function that itcarries out is the approximation of a bounded mappingor function 119891 119860 sub 119877119899 rarr 119877119898 from a compactsubset A of n-dimensional Euclidean space to a boundedsubset 119891[A] of m-dimensional Euclidean space by meansof training on examples (119909
1 1199111) (1199092 1199112) (119909
119896 119911119896) It
will always be assumed that such examples of a mapping 119891are generated by selecting xk vectors randomly from A inaccordance with a fixed probability density function 119901(x)The operational use to which the network is to be put aftertraining is also assumed to involve random selections ofinput vectors x in accordancewith119901(x)The backpropagationarchitecture described in this paper is the basic classicalversion (Figure 1) The backpropagation learning algorithmis composed of two procedures (a) forward propagation ofsignals and (b) backpropagation weight training [5]
Feed-Forward Assume that each input factor in the inputlayer is denoted by 119909
119894 the 119910
119895and 119911
119896represent the output in
the hidden layer and the output layer respectively And the 119910119895
and 119911119896can be expressed as follows (1)
119910119895= 119891 (119883
119895) = 119891(119908
119900119895+119868
sum119894=1
119908119894119895119909119894)
119911119896= 119891 (119884
119896) = 119891(119908
119900119896+119869
sum119895=1
119908119894119896119910119895)
(1)
where the 119908119900119895
and 119908119900119896
are the bias weights for settingthreshold values 119891 is the activation function used in bothhidden and output layers and 119883
119895and 119884
119896are the temporarily
computing results before applying activation function 119891 Inthis study a sigmoid function is selected as the activation
function Therefore the actual outputs 119910119895and 119911
119896in hidden
and output layers respectively can be also written as
119910119895= 119891 (119883
119895) =
1
1 + 119890minus119883119895
119911119896= 119891 (119884
119896) =
1
1 + 119890minus119884119896
(2)
The activation function 119891 introduces the nonlinear effect tothe network and maps the result of computation to a domain(0 1) This sigmoid function is differentiable The derivativeof the sigmoid function in (2) can be easily derived as 1198911015840 =119891(1 + minus119891)
Backpropagation Weight Training The error function isdefined as
119864 =1
2
119870
sum119896=1
1198902119896=119870
sum119896=1
(119905119896minus 119911119896)2
(3)
where 119905119896is a predefined network output (or desired output or
target value) and 119890119896is the error in each output nodeThe goal
is to minimize 119864 so that the weight in each link is accordinglyadjusted and the final output can match the desired outputTo get the weight adjustment the gradient descent strategyis employed In the link between hidden and output layerscomputing the partial derivative of 119864 with respect to theweight 119908
where 120575119896= (119905119896minus 119911119896) 1198911015840 (119884
119896)
(4)
Theweight adjustment in the link between hidden and outputlayers is computed by Δ119908
119895119896= 120572 times 119910
119895times 120575119896 where 120572 is
the learning rate a positive constant between 0 and 1 Thenew weight herein can be updated by the following 119908
119895119896(119899 +
1) = 119908119895119896(119899) + Δ119908
119895119896(119899) where 119899 is the number of iterations
Similarly the error gradient in links between input andhidden layers can be obtained by taking the partial derivativewith respect to 119908
119894119895as
120597119864
120597119908119894119895
= minusΔ119895119909119895= 1198911015840 (119883
119895)119870
sum119896=1
120597119896119908119895119896 (5)
The new weight in the hidden-input links can be nowcorrected as Δ119908
119894119895= 120572 times 119909
119894times Δ119895and 119908
119894119895(119899 + 1) = 119908
119894119895(119899) + Δ
119895
Training the BP-networks with many samples is sometimesa time-consuming task The learning speed can be improvedby introducing the momentum term 120578 Usually 120578 falls in therange ⟨0 1⟩ For the iteration 119899 the weight change Δ119908 can beexpressed The backpropagation learning algorithm used inartificial neural networks is shown in many text books [3ndash6]
12 Fuzzy Logic Fuzzy logics form heterogeneous family offormalisms capable of successful modelling of uncertain andvague information processing [7] The usage of fuzzy logicfor analysis and prediction of time series can be perceived
The Scientific World Journal 3
as a complement method to neural network based methodsThe symbolic background of fuzzy logic brings an advantageof human readable symbolic representation of predictioninterpretation It does not necessarily mean that fuzzy logicbased time series analysis is more accurate and more efficientbut its power lies in transparent and interpretable results thatit gives [8ndash11]
Time series analysis and prediction are an importanttask that can be used in many areas of practice The task ofgetting the best prediction to given series may bring inter-esting engineering applications in wide number of areas likeeconomics geography or industry Solution to the problemof obtaining best results in prediction of time series can bebased on well-known and simple methods like Winters orLinear method In this paper we use a method based on twomethods originally developed by members of Institute forResearch andApplications of FuzzyModeling which is a partofUniversity ofOstravaThe aimof the paper is not to presentthe details of the methods already published but to presenta tool implementing them The first method is based on thenotion of F-transform (fuzzy transform) devised by the groupof Professor Perfilieva et al [12] The second approach usesthe linguistic rules utilizing fuzzy logic and deduction that isa well-known formalism with very good results in variety ofpractical applications like industrial ones
The idea of the fuzzy transform is to transform a givenfunction defined in one space into another usually simplerspace and then to transform it back The simpler spaceconsists of a finite vector of numbers The reverse transformthen leads to a function which approximates the original oneMore details can be found in [12]
The fuzzy transform is defined with respect to a fuzzypartition which consists of basic functions Let 119888
1lt sdot sdot sdot lt 119888
119899
be fixed nodes within [119886 119887] such that 1198881
= 119886 119888119899
= 119887 and119899 ge 2 We say that fuzzy sets 119860
1 119860
119899isin 119865([119886 119887]) are basic
functions forming a fuzzy partition of [119886 119887] if they fulfill thefollowing conditions for 119894 = 1 119899
(1) 119860119894(119888119894) = 1
(2) 119860119894(119909) = 0 for 119909 isin (119888
119894minus1 119888119894+1
) where for uniformity ofnotation we put 119888
0= 1198881= 119886 and 119888
119899+1= 119888119899= 119887
(3) 119860119894is continuous
(4) 119860119894strictly increases on [119888
119894minus1 119888119894] and strictly decreases
on [119888119894 119888119894+1
](5) for all 119909 isin [119886 119887]
119899
sum119894=1
119860119894(119909) = 1 (6)
Let a fuzzy partition of [119886 119887] be given by basic functions1198601 119860
119899 119899 ge 2 and let 119891 [119886 119887] rarr 119877 be a function that is
is a fuzzy transform of 119891 with respect to the given fuzzypartition
The numbers 1198651 119865
119899are called the components of the
fuzzy transform of 119891Let 119865119899[119891] be the fuzzy transform of 119891 with respect to
1198601 119860
119899isin 119865([119886 119887])
Then the function 119891119865119899
given on [119886 119887] by
119891119865119899
(119909) =119899
sum119894=1
119865119894119860119894(119909) (8)
is called the inverse fuzzy transform of 119891Fuzzy IF-THEN rules can be understood as a specific
conditional sentence of natural language of the form IF1198831is
1198601AND sdot sdot sdot AND 119883
119899is 119860119899THEN 119884 is 119861 where 119860
1 119860
119899
and 119861 are evaluative expressions (very small roughly bigetc) An example fuzzy IF-THEN rule is as follows
IF the number of cars sold in the current year is more or lesssmall and the half-year sales increment is medium THEN theupcoming half-year increment will be medium
The part of the rule before THEN is called the antecedentand the part after it is consequent Fuzzy IF-THEN rules areusually gathered in a linguistic description
1198771= IF 119883
1is 11986011
AND sdot sdot sdotAND 119883119899
is 1198601119899
THEN 119884 is 1198611
119877119898
= IF 1198831is 1198601198981
AND sdot sdot sdotAND 119883119899
is 119860119898119899
THEN 119884 is 119861119898
(9)
Time series prediction based on these two mainapproaches works as follows Let time series 119909
119905 119905 = 1 119879
be viewed as a discrete function 119909 on a time axis 119905 Then119865119899[119909] = [119883
1 119883
119899] is the fuzzy transform of the function
119909 with respect to a given fuzzy partition The inverse fuzzytransform then serves us as a model of the trend-cycle ofa given time series By subtracting the trend-cycle (inversefuzzy transform) values from the time series lags we get pureseasonal components This is how the fuzzy transform helpsus to model and decompose a given time series
Logical dependencies between components1198831 119883
119899of
the fuzzy transform may be described by the fuzzy rulesThese rules are generated automatically from the given dataand are used for forecasting the next components Fuzzytransform components as well as their first and second orderdifferences are used as antecedent variables For forecastingfuture fuzzy transform components based on the generatedfuzzy rules a special inference methodmdashperception-basedlogical deduction is used The seasonal components are fore-casted autoregressively Finally both forecasted componentstrend-cycle and seasonal are composed together to obtainthe forecast of time series lags These methods are integratedinto an implementation PC application called linguistic fuzzylogic forecaster (LFLF) which enables as to produce linguisticdescriptions that describe properties of data treated like atime series
4 The Scientific World Journal(m
V)
+20
0
minus90
0 100 200 300 400
(ms)
Figure 2 The cardiac action potentials
2 Basic Principles of ECG Evaluation
ECG scanning has its own rules which are in accordancewith the laws of physics The heart irritation spreads inall directions In the case that the depolarisation spreadstowards the electrode which is placed on the body surfacea positive deflection is recorded on an ECG monitor Anegative deflection is recorded at the opposite end of thebody The ECG waveform is written with a chart speed of25mmsdotsminus1 An algorithm describing the curve goes in thefollowing steps First we evaluate the shape and rhythm ofventricular complexes or atrial which can be either regularor irregular Then we evaluate the frequency of ventricularcomplexes and atrial fibrillations Contraction of eachmuscleof the human body (and thus the heart as well) is associatedwith electrical changes called depolarization which can bedetected by electrodes The heart contains two basic types ofcells myocardial cells which are responsible for generatingthe pressure necessary to pump blood throughout the bodyand conduction cells which are responsible for rapidlyspreading electrical signals to the myocardial cells in orderto coordinate pumping A graph of an action potential of amuscle of cardiac cells is shown in Figure 2
A normal electrocardiogram is illustrated in Figure 3The figure also includes definitions for various segmentsand intervals in the ECG The deflections in this signalare denoted in alphabetic order starting with the letterP which represents atrial depolarization The ventriculardepolarization causes the QRS complex and repolarizationis responsible for the T-wave Atrial repolarization occursduring the QRS complex and produces such a low signalamplitude that it cannot be seen apart from the normal ECG
3 Signal Processing Using NeuralNetworks and Fuzzy Logic
In practice a relatively reliable diagnostic program storedin ECG monitors has been used which is a guideline fordetermining the final diagnosis of heart disorders This pro-gramworks according to the principle of IF-THEN rulesThevalues of the electrical signal are discretized and uploaded
P
R
TU
Q
QRSinterval
interval interval
segment segment
S
P-R
P-R
S-T
S-T
Figure 3 A typical one-cycle ECG tracing (adapted fromhttpwwwnicomwhite-paperapplargeimagelang=csampimageurl=2Fcms2Fimages2Fdevzone2Ftut2F2007-07-09141618jpg)
0
500
1000
1500
2000
1 101 201 301
Healthy personsSick persons
minus500
Figure 4 Comparison of mean values of ECG waveforms forhealthysick persons
into expert systems in the form of thousand rules The aim ofthis paper is to use a different approach based on the principleof neural networks The proposed methodology of solutioncould be summarized into the following steps
(1) a conversion of analog signal from the ECG monitorto a computer
(2) using multilayer networks that are fully connected(3) obtaining ECG waveforms in collaboration with the
University Hospital in Poruba specifically at theDepartmentCardiac Surgery from sick patients and atthe Department Traumatology from healthy patients(ie ldquohealthyrdquo with regard to heart diseases)
31 Technical Equipment ECG measurements were per-formed using ADDA Junior with converter ADDA Junior
The Scientific World Journal 5
0
10
20
30
40
50
0-01 01-02 04-05
()
Test error
Healthy personsSick persons
02ndash04 gt05
Figure 5 Experimental results test error for healthysick persons
002040608
112
1 2 3 4 5 6 7 8 9 10
H1H2
H3H4
Figure 6 Patterns representing healthy persons
which was connected to a computer via bidirectional parallelcable (CETRONICS) Technical parameters of the AD con-verter (8-bit conversion) were the following
(i) 3 measuring ranges(ii) measuring of a frequency of AC voltage at any
channel(iii) autoranging for measuring the frequency of 100Hz
1 kHz and 10 kHz(iv) input resistance of 300 kΩ(v) measurement accuracy 1
Technical parameters of the DA converter (a pro-grammable voltage source plusmn10V) were the following
(i) maximum current consumption of 15mA (after opti-mizing 4A at the output)
(ii) power of the converter plusmn15 V (stabilized)
4 Experimental Results
41 Time Series Classification and Prediction via NeuralNetworks The training set consisted of modified ECG wave-forms We used a backpropagation neural network withtopology 101-10-1 The output unit represents a diagnose 01a healthysick person A smaller number of inputs would
002040608
112
1 2 3 4 5 6 7 8 9 10
S1S2
S3S4
Figure 7 Patterns representing sick persons
0
500
1000
1500
1 26 51 76 101
Sick persons
S2S3
S4
minus500
Figure 8 Some recognized patterns that occur in ECG time series
not be appropriate due to the nature of the ECG waveformWe use 34 ECG time series associated with sick personsand 36 ECG time series associated with healthy persons25 time series of each group were used as a training setand the rest as a test set Figure 4 shows a comparison ofmean values of ECG waveforms for healthysick persons Weused the backpropagation method [5 6] for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 The conducted experimental studiesalso showed that training patterns are mixed randomly ineach cycle of adaptation This ensures their greater diversitywhich acts as a measure of system stability Uniform systemin a crisis usually collapses entirely while system with suchdiversity of trained patterns remains functional despite ofcrisis of its individual parts The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
The test set consisted of 20 samples (11 health and 9sick persons) that were not included in the training set Thesummary results for this type of experiment are shown in agraph in Figure 5 For clarity the results of testing are givenin percentage The average test error was 0194 A healthypopulation was detected with an average error of 0263 andsick population with an average error of 0109
411 Pattern Recognition Classifier Leading to Prediction Forthe purpose of adaptation of the pattern recognition classifierit is necessary to remark that determination of trainingpatterns is one of the key tasks Improperly chosen patternscan lead to confusion of neural networks During our exper-imental work we made some study which included ECGpattern recognition When creating appropriate patterns ofthe training set we used characteristic curves shown as mean
6 The Scientific World Journal
0100
1 2 3 4 5 6 7 8 9 10
S2 trainS2 test
S2 test
minus500
minus400
minus300
minus200
minus100
050
100150200250
1 2 3 4 5 6 7 8 9 10
S3 trainS3 test
S3 test
minus150
minus100
minus50
0100200300400500600
1 2 3 4 5 6 7 8 9 10
Train H3Test H3
minus200
minus100
Figure 9 Training patterns their representation in used test sets
values from ECG waveforms for healthy and sick persons(Figure 4) We use two different groups of patterns PatternsH1ndashH4 (Figure 6) represent healthy persons and patterns S1ndashS4 (Figure 7) represent sick personsThe whole training set isshown in Table 1
Pattern recognition classifier is based on backpropagationneural network and is able to recognise wave structures ingiven time series [13 14] Artificial neural networks needtraining sets for their adaptation In our experimental workthe training set consisted of 8 patterns representing thebasic structure of the various waves in ECG graphs seeFigures 6 and 7 Input data is sequences always including119899 consecutive numbers which are transformed into interval⟨0 1⟩ by formula (10) Samples are adjusted for the needs ofbackpropagation networks with sigmoid activation functionin this way [5 6]
1199091015840119895=
119909119895minusmin (119909
119894 119909
119894+119899minus1)
max (119909119894 119909
119894+119899minus1) minusmin (119909
119894 119909
119894+119899minus1)
(119895 = 119894 119894 + 119899 minus 1)
(10)
where 1199091015840119895is normalized output value of the 119895th neuron (119895 =
119894 119894 + 119899minus1) and (119909119894 119909119894+119899minus1
) are 119899minus1 consecutive outputvalues that specify sequences (patterns) from the trainingset (eg training pars of input and corresponding outputvectors) Input vector contains 10 components Output vectorhas got 8 components and each output unit represents oneof 8 different types of ECG wave samples A neural networkarchitecture is 10-10-8 (eg 10 units in the input layer 10 unitsin the hidden layer and 8 units in the output layer)The net isfully connected Adaptation of the neural network starts withrandomly generated weight values
We used the backpropagation method for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 We have utilized our experience fromearlier times that is training patterns were mixed randomlyin each cycle of adaptation The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
In order to test the efficiency of the method we appliedthe same set of data that we used in the previous experimentalpart Outputs from the classifier produce sets of values thatare assigned to each recognized training pattern in the giventest time series It is important to appreciate what can beconsidered as an effective criterion related to consensusof similarity The proposed threshold resulting from ourexperimental study was determined at least 119901 = 70Figure 9 shows a comparison of patterns how were learned(S2 S3 H3 train) and how were recognized in test time series(S2 S3 H3 test) The neural network is able to discover someconnections which are almost imperceptible Illustration ofsome recognized patterns that occur in ECG time seriesis shown in Figure 8 Outputs from the classifier carry apredictive character The neural network determines if thetime series belongs to a healthy or sick person on the basis ofthe recognised ECG patterns which appear in the time serieshistory
The methodology of testing is shown in Figure 10 Thismeans that if the test pattern S1 S2 S3 or S4 appearedin ECG waveform with probability 119901S ge 119901 (119901 = 70)thus it was predicted to be ldquoa sick personrdquo Then we workonly with the remaining time series If the test pattern H1H2 H3 or H4 appeared in ECG waveform with probability119901H ge 119901 (119901 = 70) thus it was predicted to be ldquoahealthy personrdquo In all other cases the ECG time series wasunspecifiedWe examined a total of 20 data sets Each of them
contains 101 values that assign 92 possible patternsThewholenumber of examined patterns is 1840 The graph in Figure 11demonstrates a summary of results where ldquosick personsrdquorepresent patterns S1ndashS4 and ldquohealthy personsrdquo representpatterns H1ndashH4 The resulting prediction is based on themethodology see Figure 10
Figure 12 LFLF application
Figure 13 Winning predictor linguistic description (trend-cyclemodel)
42 Time Series Classification and Prediction via LinguisticFuzzy Logic Forecaster We tried also to utilize above pre-sentedmethod of time series analysis through linguistic fuzzylogic forecaster (LFLF) [15] see Figure 12
Basic usage of the application is to analyse given timeseries and find best predictor with respect to validation partof time series given We evaluate efficiency of predictorsby SMAPE (symmetric mean absolute percentage error) Itenables us to make analysis of trend-cycle of a time seriesand also seasonal part The main advantage lies in predictionbased on transparent linguistic descriptions that provide themodel of a time series behaviour Linguistic variables are ofthe following types
(i) value we directly mean the components of the fuzzytransform
(ii) difference first order differences of fuzzy transformcomponents that are given as follows differencesbetween components Δ119883
119894= 119883119894minus 119883119894minus1
8 The Scientific World Journal
Healthypattern
typical series (HS)
Sick patterntypical series
(SS)
Tested series (TS)
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
TS is ldquohealthyrdquo TS is ldquosickrdquo
Yes No
Find best predictorfor TS as validation
interval
Find best predictorfor TS as validation
interval
Concatenate series (SS + TS)Concatenate series (HS + TS)
+ TS) forSMAPE (HS SMAPE (SS + TS) for
If SMAPE (HS + TS) lt SMAPE (SS + TS)
Figure 14 Recognition by linguistic fuzzy logic predictors for typical learning series
(iii) second difference these are values of second orderdifferences of components of the fuzzy transform asfollows Δ2119883
119894= Δ119883
119894minus Δ119883119894minus1
LFLF application enables us to define minimal andmaximal number of these particular variables in a ruleof linguistic description as well as the total number ofantecedent variables
A rule consisting of these variables has the followingstructure and can be described as a signature (fuzzy rulesdescribing the trend-cycle model) Particularly 119878 denotes the
trend-cycle components 119889119878 their differences and 1198892119878 theirsecond order differences The argument (119905) (119905 minus 1) and soforth denotes the time lag of the component
For example taking signature 119878(119905)amp119889119878(119905) rarr 119889119878(119905 + 1)denotes the fact that119883
119894and Δ119883
119894are the antecedent variables
and Δ119883119894+1
is the consequent variable of the winning modeland hence we deal with rules of the form
IF 119883119894is 119860119894AND Δ119883
119894is 119860Δ119894
THEN Δ119883119894+1
is 119860Delta119894(11)
The Scientific World Journal 9
Table 2 Example of algorithm evaluation on 10 ldquohealthyrdquo and 10 ldquosickrdquo patients
Every single fuzzy rule can be taken as a sentence ofnatural language for example first rule from Figure 13
IF 119883119894is ml sm AND Δ119883
119894is qr sm THEN Δ119883
119894+1is minusme
may be read as followsIf the number of cars sold in the current year is more or
less small and the half-year sales increment is quite roughlysmall then the upcoming half-year increment will be negativemedium
421 Recognition of ldquoHealthyrdquo and ldquoSickrdquo Patterns by LFLFOur method to use linguistic fuzzy logic forecasting isbased on simple idea that best predictor learning from bothldquohealthyrdquo and ldquosickrdquo pattern samples respectively can be usedfor validation with tested pattern taken as validation part ofthe seriesThen we can evaluate SMAPE for both these casescompound series SMAPE (ldquohealthyrdquo + tested) and SMAPE(ldquosickrdquo + tested)
If SMAPE (ldquohealthyrdquo + tested) lt SMAPE (ldquosickrdquo + tested)then the tested pattern is supposed to be ldquohealthyrdquo otherwisethe tested pattern is supposed to be ldquosickrdquo
The idea is schematically shown in Figure 14For testing purposes we created two necessary typical
learning time series ldquohealthyrdquo (HS) and ldquosickrdquo (SS) accordingto the algorithm above They both consist of 1010 samplesmade from 10 typical series of ldquohealthyrdquo and ldquosickrdquo patientswith 101 measured ECG values Then we have created 10concatenated series according to the scheme in Figure 14with 10 randomly selected patients with ldquohealthyrdquo ECGmeasurement that is 20 files were produced (10x HS + TSand 10x SS + TS) The same concatenated series were alsomade from 10 ldquosickrdquo patients measurements This made usadditional 20 files with concatenated series (10x HS + TS and
0102030405060708090
100
Sick Healthy
FalseTrue
()
Figure 15 Experimental results LFLF
10x SS + TS) For 20 patients (Table 2) tested ECG we have2 concatenated series giving SMAPE (HS + TS) and SMAPE(SS + TS)
Our method based on LFLF proved very good results forright identification of sick patient records Nevertheless itproduces large amount of false positive identification of sickpattern for healthy patients (Figure 15) This result is con-sistent with our approach using neural networks Of courseour preliminary research has a limited extent and should beperceived only as narrative result which shows interestingproperties especially in complementation of neural networkresults
10 The Scientific World Journal
5 Conclusion
In this paper a short introduction into the field of ECGwaves recognition using backpropagation neural network hasbeen given Main objective was to recognise the normalcycles and arrhythmias and perform further diagnosis Weproposed two detection systems that have been created withusage of neural networks One of them is adapted accordingto the training set Here each pattern represents the wholeone ECG cycle Then an output unit represents a diagnose01 a healthysick person The second one approach usesneural network in which training set contains two differentgroups of patterns for healthysick persons According to theresults of experimental studies it can be stated that ECGwaves patterns were successfully extracted in given timeseries and recognised using suggestedmethod as can be seenfrom figures in Experimental Result section It might resultin better mapping of the time series behaviour for betterprediction
Both approaches were able to predict with high probabil-ity if the ECG time series represents sick or healthy persons Itis interesting that a sick diagnose was recognised with higheraccuracy in both experimental works
The third approach based on LFLF is currently only inthe stage of preliminary experiments but it conforms to theformer results based on neural networks This approach isnovel and could be good supplement to other soft-computingmethods for this task
Disclaimer
Anyopinions findings and conclusions or recommendationsexpressed in this material are those of the authors and do notnecessarily reflect the views of the sponsors
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
The research described here has been financially sup-ported by University of Ostrava Grants SGS16PrF2014 andSGS14PrF2014
References
[1] P Tangkraingkij C Lursinsap S Sanguansintukul and TDesudchit ldquoSelecting relevant EEG signal locations for personalidentification problem using ICA and neural networkrdquo inProceedings of the 8th IEEEACIS International Conference onComputer and Information Science (ICIS rsquo09) pp 616ndash621 June2009
[2] C D Nugent J A C Webb N D Black and G T H WrightldquoElectrocardiogram 2 classificationrdquo Automedica vol 17 pp281ndash306 1999
[3] S Russell and P Norvig Artificial IntelligencemdashA ModernApproach Prentice Hall 2nd edition 2003
[4] C D Nugent J A Lopez A E Smith and N D BlackldquoReverse engineering of neural network classifiers in medicalapplicationsrdquo in Proceedings of the 7th EFOMP Congress vol 17p 184 Physica Medica 2001
[5] L Fausett Fundamentals of Neural Network PrenticeHall 1994[6] D E Rumelhart G E Hinton and R J Williams ldquoLearning
[7] V Novak Fuzzy Sets and Their Applications Adam HilgerBristol UK 1989
[8] S Aouaouda M Chadli P Shi and H R Karimi ldquoDiscrete-time119867
minus119867infinsensor fault detection observer design for nonlin-
ear systems with parameter uncertaintyrdquo International Journalof Robust and Nonlinear Control 2013
[9] M Chadli A Abdo and S X Ding ldquo119867minus119867infin
fault detectionfilter design for discrete-time Takagi-Sugeno fuzzy systemrdquoAutomatica vol 49 no 7 pp 1996ndash2005 2013
[10] M Chadli and H R Karimi ldquoRobust observer design forunknown inputs takagi-sugeno modelsrdquo IEEE Transactions onFuzzy Systems vol 21 no 1 pp 158ndash164 2013
[11] S Aouaouda M Chadli and H Karimi ldquoRobust static output-feedback controller design against sensor failure for vehicledynamicsrdquo IET Control Theory amp Applications vol 8 no 9 pp728ndash737 2014
[12] I Perfilieva V Novak V Pavliska A Dvorak andM StepnickaldquoAnalysis and prediction of time series using fuzzy transformrdquoin Proceedings of the International Joint Conference on NeuralNetworks (IJCNN rsquo08) pp 3876ndash3880 Hong Kong June 2008
[13] E Volna M Kotyrba and R Jarusek ldquoMulti-classifier basedon Elliott waversquos recognitionrdquo Computers amp Mathematics withApplications vol 66 no 2 pp 213ndash225 2013
[14] E Volna M Kotyrba and R Jarusek ldquoPrediction by means ofElliott waves recognitionrdquo in Nostradamus Modern Methods ofPrediction Modeling and Analysis of Nonlinear Systems vol 192of Advances in Intelligent Systems and Computing pp 241ndash250Springer Berlin Germany 2013
[15] H Habiballa V Pavliska and A Dvorak ldquoSoftware systemfor time series prediction based on F-transform and linguisticrulesrdquo in Proceedings of the 8th International Conference onApplied Mathematics Aplimat pp 381ndash386 Bratislava Slovakia2009
as a complement method to neural network based methodsThe symbolic background of fuzzy logic brings an advantageof human readable symbolic representation of predictioninterpretation It does not necessarily mean that fuzzy logicbased time series analysis is more accurate and more efficientbut its power lies in transparent and interpretable results thatit gives [8ndash11]
Time series analysis and prediction are an importanttask that can be used in many areas of practice The task ofgetting the best prediction to given series may bring inter-esting engineering applications in wide number of areas likeeconomics geography or industry Solution to the problemof obtaining best results in prediction of time series can bebased on well-known and simple methods like Winters orLinear method In this paper we use a method based on twomethods originally developed by members of Institute forResearch andApplications of FuzzyModeling which is a partofUniversity ofOstravaThe aimof the paper is not to presentthe details of the methods already published but to presenta tool implementing them The first method is based on thenotion of F-transform (fuzzy transform) devised by the groupof Professor Perfilieva et al [12] The second approach usesthe linguistic rules utilizing fuzzy logic and deduction that isa well-known formalism with very good results in variety ofpractical applications like industrial ones
The idea of the fuzzy transform is to transform a givenfunction defined in one space into another usually simplerspace and then to transform it back The simpler spaceconsists of a finite vector of numbers The reverse transformthen leads to a function which approximates the original oneMore details can be found in [12]
The fuzzy transform is defined with respect to a fuzzypartition which consists of basic functions Let 119888
1lt sdot sdot sdot lt 119888
119899
be fixed nodes within [119886 119887] such that 1198881
= 119886 119888119899
= 119887 and119899 ge 2 We say that fuzzy sets 119860
1 119860
119899isin 119865([119886 119887]) are basic
functions forming a fuzzy partition of [119886 119887] if they fulfill thefollowing conditions for 119894 = 1 119899
(1) 119860119894(119888119894) = 1
(2) 119860119894(119909) = 0 for 119909 isin (119888
119894minus1 119888119894+1
) where for uniformity ofnotation we put 119888
0= 1198881= 119886 and 119888
119899+1= 119888119899= 119887
(3) 119860119894is continuous
(4) 119860119894strictly increases on [119888
119894minus1 119888119894] and strictly decreases
on [119888119894 119888119894+1
](5) for all 119909 isin [119886 119887]
119899
sum119894=1
119860119894(119909) = 1 (6)
Let a fuzzy partition of [119886 119887] be given by basic functions1198601 119860
119899 119899 ge 2 and let 119891 [119886 119887] rarr 119877 be a function that is
is a fuzzy transform of 119891 with respect to the given fuzzypartition
The numbers 1198651 119865
119899are called the components of the
fuzzy transform of 119891Let 119865119899[119891] be the fuzzy transform of 119891 with respect to
1198601 119860
119899isin 119865([119886 119887])
Then the function 119891119865119899
given on [119886 119887] by
119891119865119899
(119909) =119899
sum119894=1
119865119894119860119894(119909) (8)
is called the inverse fuzzy transform of 119891Fuzzy IF-THEN rules can be understood as a specific
conditional sentence of natural language of the form IF1198831is
1198601AND sdot sdot sdot AND 119883
119899is 119860119899THEN 119884 is 119861 where 119860
1 119860
119899
and 119861 are evaluative expressions (very small roughly bigetc) An example fuzzy IF-THEN rule is as follows
IF the number of cars sold in the current year is more or lesssmall and the half-year sales increment is medium THEN theupcoming half-year increment will be medium
The part of the rule before THEN is called the antecedentand the part after it is consequent Fuzzy IF-THEN rules areusually gathered in a linguistic description
1198771= IF 119883
1is 11986011
AND sdot sdot sdotAND 119883119899
is 1198601119899
THEN 119884 is 1198611
119877119898
= IF 1198831is 1198601198981
AND sdot sdot sdotAND 119883119899
is 119860119898119899
THEN 119884 is 119861119898
(9)
Time series prediction based on these two mainapproaches works as follows Let time series 119909
119905 119905 = 1 119879
be viewed as a discrete function 119909 on a time axis 119905 Then119865119899[119909] = [119883
1 119883
119899] is the fuzzy transform of the function
119909 with respect to a given fuzzy partition The inverse fuzzytransform then serves us as a model of the trend-cycle ofa given time series By subtracting the trend-cycle (inversefuzzy transform) values from the time series lags we get pureseasonal components This is how the fuzzy transform helpsus to model and decompose a given time series
Logical dependencies between components1198831 119883
119899of
the fuzzy transform may be described by the fuzzy rulesThese rules are generated automatically from the given dataand are used for forecasting the next components Fuzzytransform components as well as their first and second orderdifferences are used as antecedent variables For forecastingfuture fuzzy transform components based on the generatedfuzzy rules a special inference methodmdashperception-basedlogical deduction is used The seasonal components are fore-casted autoregressively Finally both forecasted componentstrend-cycle and seasonal are composed together to obtainthe forecast of time series lags These methods are integratedinto an implementation PC application called linguistic fuzzylogic forecaster (LFLF) which enables as to produce linguisticdescriptions that describe properties of data treated like atime series
4 The Scientific World Journal(m
V)
+20
0
minus90
0 100 200 300 400
(ms)
Figure 2 The cardiac action potentials
2 Basic Principles of ECG Evaluation
ECG scanning has its own rules which are in accordancewith the laws of physics The heart irritation spreads inall directions In the case that the depolarisation spreadstowards the electrode which is placed on the body surfacea positive deflection is recorded on an ECG monitor Anegative deflection is recorded at the opposite end of thebody The ECG waveform is written with a chart speed of25mmsdotsminus1 An algorithm describing the curve goes in thefollowing steps First we evaluate the shape and rhythm ofventricular complexes or atrial which can be either regularor irregular Then we evaluate the frequency of ventricularcomplexes and atrial fibrillations Contraction of eachmuscleof the human body (and thus the heart as well) is associatedwith electrical changes called depolarization which can bedetected by electrodes The heart contains two basic types ofcells myocardial cells which are responsible for generatingthe pressure necessary to pump blood throughout the bodyand conduction cells which are responsible for rapidlyspreading electrical signals to the myocardial cells in orderto coordinate pumping A graph of an action potential of amuscle of cardiac cells is shown in Figure 2
A normal electrocardiogram is illustrated in Figure 3The figure also includes definitions for various segmentsand intervals in the ECG The deflections in this signalare denoted in alphabetic order starting with the letterP which represents atrial depolarization The ventriculardepolarization causes the QRS complex and repolarizationis responsible for the T-wave Atrial repolarization occursduring the QRS complex and produces such a low signalamplitude that it cannot be seen apart from the normal ECG
3 Signal Processing Using NeuralNetworks and Fuzzy Logic
In practice a relatively reliable diagnostic program storedin ECG monitors has been used which is a guideline fordetermining the final diagnosis of heart disorders This pro-gramworks according to the principle of IF-THEN rulesThevalues of the electrical signal are discretized and uploaded
P
R
TU
Q
QRSinterval
interval interval
segment segment
S
P-R
P-R
S-T
S-T
Figure 3 A typical one-cycle ECG tracing (adapted fromhttpwwwnicomwhite-paperapplargeimagelang=csampimageurl=2Fcms2Fimages2Fdevzone2Ftut2F2007-07-09141618jpg)
0
500
1000
1500
2000
1 101 201 301
Healthy personsSick persons
minus500
Figure 4 Comparison of mean values of ECG waveforms forhealthysick persons
into expert systems in the form of thousand rules The aim ofthis paper is to use a different approach based on the principleof neural networks The proposed methodology of solutioncould be summarized into the following steps
(1) a conversion of analog signal from the ECG monitorto a computer
(2) using multilayer networks that are fully connected(3) obtaining ECG waveforms in collaboration with the
University Hospital in Poruba specifically at theDepartmentCardiac Surgery from sick patients and atthe Department Traumatology from healthy patients(ie ldquohealthyrdquo with regard to heart diseases)
31 Technical Equipment ECG measurements were per-formed using ADDA Junior with converter ADDA Junior
The Scientific World Journal 5
0
10
20
30
40
50
0-01 01-02 04-05
()
Test error
Healthy personsSick persons
02ndash04 gt05
Figure 5 Experimental results test error for healthysick persons
002040608
112
1 2 3 4 5 6 7 8 9 10
H1H2
H3H4
Figure 6 Patterns representing healthy persons
which was connected to a computer via bidirectional parallelcable (CETRONICS) Technical parameters of the AD con-verter (8-bit conversion) were the following
(i) 3 measuring ranges(ii) measuring of a frequency of AC voltage at any
channel(iii) autoranging for measuring the frequency of 100Hz
1 kHz and 10 kHz(iv) input resistance of 300 kΩ(v) measurement accuracy 1
Technical parameters of the DA converter (a pro-grammable voltage source plusmn10V) were the following
(i) maximum current consumption of 15mA (after opti-mizing 4A at the output)
(ii) power of the converter plusmn15 V (stabilized)
4 Experimental Results
41 Time Series Classification and Prediction via NeuralNetworks The training set consisted of modified ECG wave-forms We used a backpropagation neural network withtopology 101-10-1 The output unit represents a diagnose 01a healthysick person A smaller number of inputs would
002040608
112
1 2 3 4 5 6 7 8 9 10
S1S2
S3S4
Figure 7 Patterns representing sick persons
0
500
1000
1500
1 26 51 76 101
Sick persons
S2S3
S4
minus500
Figure 8 Some recognized patterns that occur in ECG time series
not be appropriate due to the nature of the ECG waveformWe use 34 ECG time series associated with sick personsand 36 ECG time series associated with healthy persons25 time series of each group were used as a training setand the rest as a test set Figure 4 shows a comparison ofmean values of ECG waveforms for healthysick persons Weused the backpropagation method [5 6] for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 The conducted experimental studiesalso showed that training patterns are mixed randomly ineach cycle of adaptation This ensures their greater diversitywhich acts as a measure of system stability Uniform systemin a crisis usually collapses entirely while system with suchdiversity of trained patterns remains functional despite ofcrisis of its individual parts The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
The test set consisted of 20 samples (11 health and 9sick persons) that were not included in the training set Thesummary results for this type of experiment are shown in agraph in Figure 5 For clarity the results of testing are givenin percentage The average test error was 0194 A healthypopulation was detected with an average error of 0263 andsick population with an average error of 0109
411 Pattern Recognition Classifier Leading to Prediction Forthe purpose of adaptation of the pattern recognition classifierit is necessary to remark that determination of trainingpatterns is one of the key tasks Improperly chosen patternscan lead to confusion of neural networks During our exper-imental work we made some study which included ECGpattern recognition When creating appropriate patterns ofthe training set we used characteristic curves shown as mean
6 The Scientific World Journal
0100
1 2 3 4 5 6 7 8 9 10
S2 trainS2 test
S2 test
minus500
minus400
minus300
minus200
minus100
050
100150200250
1 2 3 4 5 6 7 8 9 10
S3 trainS3 test
S3 test
minus150
minus100
minus50
0100200300400500600
1 2 3 4 5 6 7 8 9 10
Train H3Test H3
minus200
minus100
Figure 9 Training patterns their representation in used test sets
values from ECG waveforms for healthy and sick persons(Figure 4) We use two different groups of patterns PatternsH1ndashH4 (Figure 6) represent healthy persons and patterns S1ndashS4 (Figure 7) represent sick personsThe whole training set isshown in Table 1
Pattern recognition classifier is based on backpropagationneural network and is able to recognise wave structures ingiven time series [13 14] Artificial neural networks needtraining sets for their adaptation In our experimental workthe training set consisted of 8 patterns representing thebasic structure of the various waves in ECG graphs seeFigures 6 and 7 Input data is sequences always including119899 consecutive numbers which are transformed into interval⟨0 1⟩ by formula (10) Samples are adjusted for the needs ofbackpropagation networks with sigmoid activation functionin this way [5 6]
1199091015840119895=
119909119895minusmin (119909
119894 119909
119894+119899minus1)
max (119909119894 119909
119894+119899minus1) minusmin (119909
119894 119909
119894+119899minus1)
(119895 = 119894 119894 + 119899 minus 1)
(10)
where 1199091015840119895is normalized output value of the 119895th neuron (119895 =
119894 119894 + 119899minus1) and (119909119894 119909119894+119899minus1
) are 119899minus1 consecutive outputvalues that specify sequences (patterns) from the trainingset (eg training pars of input and corresponding outputvectors) Input vector contains 10 components Output vectorhas got 8 components and each output unit represents oneof 8 different types of ECG wave samples A neural networkarchitecture is 10-10-8 (eg 10 units in the input layer 10 unitsin the hidden layer and 8 units in the output layer)The net isfully connected Adaptation of the neural network starts withrandomly generated weight values
We used the backpropagation method for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 We have utilized our experience fromearlier times that is training patterns were mixed randomlyin each cycle of adaptation The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
In order to test the efficiency of the method we appliedthe same set of data that we used in the previous experimentalpart Outputs from the classifier produce sets of values thatare assigned to each recognized training pattern in the giventest time series It is important to appreciate what can beconsidered as an effective criterion related to consensusof similarity The proposed threshold resulting from ourexperimental study was determined at least 119901 = 70Figure 9 shows a comparison of patterns how were learned(S2 S3 H3 train) and how were recognized in test time series(S2 S3 H3 test) The neural network is able to discover someconnections which are almost imperceptible Illustration ofsome recognized patterns that occur in ECG time seriesis shown in Figure 8 Outputs from the classifier carry apredictive character The neural network determines if thetime series belongs to a healthy or sick person on the basis ofthe recognised ECG patterns which appear in the time serieshistory
The methodology of testing is shown in Figure 10 Thismeans that if the test pattern S1 S2 S3 or S4 appearedin ECG waveform with probability 119901S ge 119901 (119901 = 70)thus it was predicted to be ldquoa sick personrdquo Then we workonly with the remaining time series If the test pattern H1H2 H3 or H4 appeared in ECG waveform with probability119901H ge 119901 (119901 = 70) thus it was predicted to be ldquoahealthy personrdquo In all other cases the ECG time series wasunspecifiedWe examined a total of 20 data sets Each of them
contains 101 values that assign 92 possible patternsThewholenumber of examined patterns is 1840 The graph in Figure 11demonstrates a summary of results where ldquosick personsrdquorepresent patterns S1ndashS4 and ldquohealthy personsrdquo representpatterns H1ndashH4 The resulting prediction is based on themethodology see Figure 10
Figure 12 LFLF application
Figure 13 Winning predictor linguistic description (trend-cyclemodel)
42 Time Series Classification and Prediction via LinguisticFuzzy Logic Forecaster We tried also to utilize above pre-sentedmethod of time series analysis through linguistic fuzzylogic forecaster (LFLF) [15] see Figure 12
Basic usage of the application is to analyse given timeseries and find best predictor with respect to validation partof time series given We evaluate efficiency of predictorsby SMAPE (symmetric mean absolute percentage error) Itenables us to make analysis of trend-cycle of a time seriesand also seasonal part The main advantage lies in predictionbased on transparent linguistic descriptions that provide themodel of a time series behaviour Linguistic variables are ofthe following types
(i) value we directly mean the components of the fuzzytransform
(ii) difference first order differences of fuzzy transformcomponents that are given as follows differencesbetween components Δ119883
119894= 119883119894minus 119883119894minus1
8 The Scientific World Journal
Healthypattern
typical series (HS)
Sick patterntypical series
(SS)
Tested series (TS)
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
TS is ldquohealthyrdquo TS is ldquosickrdquo
Yes No
Find best predictorfor TS as validation
interval
Find best predictorfor TS as validation
interval
Concatenate series (SS + TS)Concatenate series (HS + TS)
+ TS) forSMAPE (HS SMAPE (SS + TS) for
If SMAPE (HS + TS) lt SMAPE (SS + TS)
Figure 14 Recognition by linguistic fuzzy logic predictors for typical learning series
(iii) second difference these are values of second orderdifferences of components of the fuzzy transform asfollows Δ2119883
119894= Δ119883
119894minus Δ119883119894minus1
LFLF application enables us to define minimal andmaximal number of these particular variables in a ruleof linguistic description as well as the total number ofantecedent variables
A rule consisting of these variables has the followingstructure and can be described as a signature (fuzzy rulesdescribing the trend-cycle model) Particularly 119878 denotes the
trend-cycle components 119889119878 their differences and 1198892119878 theirsecond order differences The argument (119905) (119905 minus 1) and soforth denotes the time lag of the component
For example taking signature 119878(119905)amp119889119878(119905) rarr 119889119878(119905 + 1)denotes the fact that119883
119894and Δ119883
119894are the antecedent variables
and Δ119883119894+1
is the consequent variable of the winning modeland hence we deal with rules of the form
IF 119883119894is 119860119894AND Δ119883
119894is 119860Δ119894
THEN Δ119883119894+1
is 119860Delta119894(11)
The Scientific World Journal 9
Table 2 Example of algorithm evaluation on 10 ldquohealthyrdquo and 10 ldquosickrdquo patients
Every single fuzzy rule can be taken as a sentence ofnatural language for example first rule from Figure 13
IF 119883119894is ml sm AND Δ119883
119894is qr sm THEN Δ119883
119894+1is minusme
may be read as followsIf the number of cars sold in the current year is more or
less small and the half-year sales increment is quite roughlysmall then the upcoming half-year increment will be negativemedium
421 Recognition of ldquoHealthyrdquo and ldquoSickrdquo Patterns by LFLFOur method to use linguistic fuzzy logic forecasting isbased on simple idea that best predictor learning from bothldquohealthyrdquo and ldquosickrdquo pattern samples respectively can be usedfor validation with tested pattern taken as validation part ofthe seriesThen we can evaluate SMAPE for both these casescompound series SMAPE (ldquohealthyrdquo + tested) and SMAPE(ldquosickrdquo + tested)
If SMAPE (ldquohealthyrdquo + tested) lt SMAPE (ldquosickrdquo + tested)then the tested pattern is supposed to be ldquohealthyrdquo otherwisethe tested pattern is supposed to be ldquosickrdquo
The idea is schematically shown in Figure 14For testing purposes we created two necessary typical
learning time series ldquohealthyrdquo (HS) and ldquosickrdquo (SS) accordingto the algorithm above They both consist of 1010 samplesmade from 10 typical series of ldquohealthyrdquo and ldquosickrdquo patientswith 101 measured ECG values Then we have created 10concatenated series according to the scheme in Figure 14with 10 randomly selected patients with ldquohealthyrdquo ECGmeasurement that is 20 files were produced (10x HS + TSand 10x SS + TS) The same concatenated series were alsomade from 10 ldquosickrdquo patients measurements This made usadditional 20 files with concatenated series (10x HS + TS and
0102030405060708090
100
Sick Healthy
FalseTrue
()
Figure 15 Experimental results LFLF
10x SS + TS) For 20 patients (Table 2) tested ECG we have2 concatenated series giving SMAPE (HS + TS) and SMAPE(SS + TS)
Our method based on LFLF proved very good results forright identification of sick patient records Nevertheless itproduces large amount of false positive identification of sickpattern for healthy patients (Figure 15) This result is con-sistent with our approach using neural networks Of courseour preliminary research has a limited extent and should beperceived only as narrative result which shows interestingproperties especially in complementation of neural networkresults
10 The Scientific World Journal
5 Conclusion
In this paper a short introduction into the field of ECGwaves recognition using backpropagation neural network hasbeen given Main objective was to recognise the normalcycles and arrhythmias and perform further diagnosis Weproposed two detection systems that have been created withusage of neural networks One of them is adapted accordingto the training set Here each pattern represents the wholeone ECG cycle Then an output unit represents a diagnose01 a healthysick person The second one approach usesneural network in which training set contains two differentgroups of patterns for healthysick persons According to theresults of experimental studies it can be stated that ECGwaves patterns were successfully extracted in given timeseries and recognised using suggestedmethod as can be seenfrom figures in Experimental Result section It might resultin better mapping of the time series behaviour for betterprediction
Both approaches were able to predict with high probabil-ity if the ECG time series represents sick or healthy persons Itis interesting that a sick diagnose was recognised with higheraccuracy in both experimental works
The third approach based on LFLF is currently only inthe stage of preliminary experiments but it conforms to theformer results based on neural networks This approach isnovel and could be good supplement to other soft-computingmethods for this task
Disclaimer
Anyopinions findings and conclusions or recommendationsexpressed in this material are those of the authors and do notnecessarily reflect the views of the sponsors
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
The research described here has been financially sup-ported by University of Ostrava Grants SGS16PrF2014 andSGS14PrF2014
References
[1] P Tangkraingkij C Lursinsap S Sanguansintukul and TDesudchit ldquoSelecting relevant EEG signal locations for personalidentification problem using ICA and neural networkrdquo inProceedings of the 8th IEEEACIS International Conference onComputer and Information Science (ICIS rsquo09) pp 616ndash621 June2009
[2] C D Nugent J A C Webb N D Black and G T H WrightldquoElectrocardiogram 2 classificationrdquo Automedica vol 17 pp281ndash306 1999
[3] S Russell and P Norvig Artificial IntelligencemdashA ModernApproach Prentice Hall 2nd edition 2003
[4] C D Nugent J A Lopez A E Smith and N D BlackldquoReverse engineering of neural network classifiers in medicalapplicationsrdquo in Proceedings of the 7th EFOMP Congress vol 17p 184 Physica Medica 2001
[5] L Fausett Fundamentals of Neural Network PrenticeHall 1994[6] D E Rumelhart G E Hinton and R J Williams ldquoLearning
[7] V Novak Fuzzy Sets and Their Applications Adam HilgerBristol UK 1989
[8] S Aouaouda M Chadli P Shi and H R Karimi ldquoDiscrete-time119867
minus119867infinsensor fault detection observer design for nonlin-
ear systems with parameter uncertaintyrdquo International Journalof Robust and Nonlinear Control 2013
[9] M Chadli A Abdo and S X Ding ldquo119867minus119867infin
fault detectionfilter design for discrete-time Takagi-Sugeno fuzzy systemrdquoAutomatica vol 49 no 7 pp 1996ndash2005 2013
[10] M Chadli and H R Karimi ldquoRobust observer design forunknown inputs takagi-sugeno modelsrdquo IEEE Transactions onFuzzy Systems vol 21 no 1 pp 158ndash164 2013
[11] S Aouaouda M Chadli and H Karimi ldquoRobust static output-feedback controller design against sensor failure for vehicledynamicsrdquo IET Control Theory amp Applications vol 8 no 9 pp728ndash737 2014
[12] I Perfilieva V Novak V Pavliska A Dvorak andM StepnickaldquoAnalysis and prediction of time series using fuzzy transformrdquoin Proceedings of the International Joint Conference on NeuralNetworks (IJCNN rsquo08) pp 3876ndash3880 Hong Kong June 2008
[13] E Volna M Kotyrba and R Jarusek ldquoMulti-classifier basedon Elliott waversquos recognitionrdquo Computers amp Mathematics withApplications vol 66 no 2 pp 213ndash225 2013
[14] E Volna M Kotyrba and R Jarusek ldquoPrediction by means ofElliott waves recognitionrdquo in Nostradamus Modern Methods ofPrediction Modeling and Analysis of Nonlinear Systems vol 192of Advances in Intelligent Systems and Computing pp 241ndash250Springer Berlin Germany 2013
[15] H Habiballa V Pavliska and A Dvorak ldquoSoftware systemfor time series prediction based on F-transform and linguisticrulesrdquo in Proceedings of the 8th International Conference onApplied Mathematics Aplimat pp 381ndash386 Bratislava Slovakia2009
ECG scanning has its own rules which are in accordancewith the laws of physics The heart irritation spreads inall directions In the case that the depolarisation spreadstowards the electrode which is placed on the body surfacea positive deflection is recorded on an ECG monitor Anegative deflection is recorded at the opposite end of thebody The ECG waveform is written with a chart speed of25mmsdotsminus1 An algorithm describing the curve goes in thefollowing steps First we evaluate the shape and rhythm ofventricular complexes or atrial which can be either regularor irregular Then we evaluate the frequency of ventricularcomplexes and atrial fibrillations Contraction of eachmuscleof the human body (and thus the heart as well) is associatedwith electrical changes called depolarization which can bedetected by electrodes The heart contains two basic types ofcells myocardial cells which are responsible for generatingthe pressure necessary to pump blood throughout the bodyand conduction cells which are responsible for rapidlyspreading electrical signals to the myocardial cells in orderto coordinate pumping A graph of an action potential of amuscle of cardiac cells is shown in Figure 2
A normal electrocardiogram is illustrated in Figure 3The figure also includes definitions for various segmentsand intervals in the ECG The deflections in this signalare denoted in alphabetic order starting with the letterP which represents atrial depolarization The ventriculardepolarization causes the QRS complex and repolarizationis responsible for the T-wave Atrial repolarization occursduring the QRS complex and produces such a low signalamplitude that it cannot be seen apart from the normal ECG
3 Signal Processing Using NeuralNetworks and Fuzzy Logic
In practice a relatively reliable diagnostic program storedin ECG monitors has been used which is a guideline fordetermining the final diagnosis of heart disorders This pro-gramworks according to the principle of IF-THEN rulesThevalues of the electrical signal are discretized and uploaded
P
R
TU
Q
QRSinterval
interval interval
segment segment
S
P-R
P-R
S-T
S-T
Figure 3 A typical one-cycle ECG tracing (adapted fromhttpwwwnicomwhite-paperapplargeimagelang=csampimageurl=2Fcms2Fimages2Fdevzone2Ftut2F2007-07-09141618jpg)
0
500
1000
1500
2000
1 101 201 301
Healthy personsSick persons
minus500
Figure 4 Comparison of mean values of ECG waveforms forhealthysick persons
into expert systems in the form of thousand rules The aim ofthis paper is to use a different approach based on the principleof neural networks The proposed methodology of solutioncould be summarized into the following steps
(1) a conversion of analog signal from the ECG monitorto a computer
(2) using multilayer networks that are fully connected(3) obtaining ECG waveforms in collaboration with the
University Hospital in Poruba specifically at theDepartmentCardiac Surgery from sick patients and atthe Department Traumatology from healthy patients(ie ldquohealthyrdquo with regard to heart diseases)
31 Technical Equipment ECG measurements were per-formed using ADDA Junior with converter ADDA Junior
The Scientific World Journal 5
0
10
20
30
40
50
0-01 01-02 04-05
()
Test error
Healthy personsSick persons
02ndash04 gt05
Figure 5 Experimental results test error for healthysick persons
002040608
112
1 2 3 4 5 6 7 8 9 10
H1H2
H3H4
Figure 6 Patterns representing healthy persons
which was connected to a computer via bidirectional parallelcable (CETRONICS) Technical parameters of the AD con-verter (8-bit conversion) were the following
(i) 3 measuring ranges(ii) measuring of a frequency of AC voltage at any
channel(iii) autoranging for measuring the frequency of 100Hz
1 kHz and 10 kHz(iv) input resistance of 300 kΩ(v) measurement accuracy 1
Technical parameters of the DA converter (a pro-grammable voltage source plusmn10V) were the following
(i) maximum current consumption of 15mA (after opti-mizing 4A at the output)
(ii) power of the converter plusmn15 V (stabilized)
4 Experimental Results
41 Time Series Classification and Prediction via NeuralNetworks The training set consisted of modified ECG wave-forms We used a backpropagation neural network withtopology 101-10-1 The output unit represents a diagnose 01a healthysick person A smaller number of inputs would
002040608
112
1 2 3 4 5 6 7 8 9 10
S1S2
S3S4
Figure 7 Patterns representing sick persons
0
500
1000
1500
1 26 51 76 101
Sick persons
S2S3
S4
minus500
Figure 8 Some recognized patterns that occur in ECG time series
not be appropriate due to the nature of the ECG waveformWe use 34 ECG time series associated with sick personsand 36 ECG time series associated with healthy persons25 time series of each group were used as a training setand the rest as a test set Figure 4 shows a comparison ofmean values of ECG waveforms for healthysick persons Weused the backpropagation method [5 6] for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 The conducted experimental studiesalso showed that training patterns are mixed randomly ineach cycle of adaptation This ensures their greater diversitywhich acts as a measure of system stability Uniform systemin a crisis usually collapses entirely while system with suchdiversity of trained patterns remains functional despite ofcrisis of its individual parts The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
The test set consisted of 20 samples (11 health and 9sick persons) that were not included in the training set Thesummary results for this type of experiment are shown in agraph in Figure 5 For clarity the results of testing are givenin percentage The average test error was 0194 A healthypopulation was detected with an average error of 0263 andsick population with an average error of 0109
411 Pattern Recognition Classifier Leading to Prediction Forthe purpose of adaptation of the pattern recognition classifierit is necessary to remark that determination of trainingpatterns is one of the key tasks Improperly chosen patternscan lead to confusion of neural networks During our exper-imental work we made some study which included ECGpattern recognition When creating appropriate patterns ofthe training set we used characteristic curves shown as mean
6 The Scientific World Journal
0100
1 2 3 4 5 6 7 8 9 10
S2 trainS2 test
S2 test
minus500
minus400
minus300
minus200
minus100
050
100150200250
1 2 3 4 5 6 7 8 9 10
S3 trainS3 test
S3 test
minus150
minus100
minus50
0100200300400500600
1 2 3 4 5 6 7 8 9 10
Train H3Test H3
minus200
minus100
Figure 9 Training patterns their representation in used test sets
values from ECG waveforms for healthy and sick persons(Figure 4) We use two different groups of patterns PatternsH1ndashH4 (Figure 6) represent healthy persons and patterns S1ndashS4 (Figure 7) represent sick personsThe whole training set isshown in Table 1
Pattern recognition classifier is based on backpropagationneural network and is able to recognise wave structures ingiven time series [13 14] Artificial neural networks needtraining sets for their adaptation In our experimental workthe training set consisted of 8 patterns representing thebasic structure of the various waves in ECG graphs seeFigures 6 and 7 Input data is sequences always including119899 consecutive numbers which are transformed into interval⟨0 1⟩ by formula (10) Samples are adjusted for the needs ofbackpropagation networks with sigmoid activation functionin this way [5 6]
1199091015840119895=
119909119895minusmin (119909
119894 119909
119894+119899minus1)
max (119909119894 119909
119894+119899minus1) minusmin (119909
119894 119909
119894+119899minus1)
(119895 = 119894 119894 + 119899 minus 1)
(10)
where 1199091015840119895is normalized output value of the 119895th neuron (119895 =
119894 119894 + 119899minus1) and (119909119894 119909119894+119899minus1
) are 119899minus1 consecutive outputvalues that specify sequences (patterns) from the trainingset (eg training pars of input and corresponding outputvectors) Input vector contains 10 components Output vectorhas got 8 components and each output unit represents oneof 8 different types of ECG wave samples A neural networkarchitecture is 10-10-8 (eg 10 units in the input layer 10 unitsin the hidden layer and 8 units in the output layer)The net isfully connected Adaptation of the neural network starts withrandomly generated weight values
We used the backpropagation method for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 We have utilized our experience fromearlier times that is training patterns were mixed randomlyin each cycle of adaptation The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
In order to test the efficiency of the method we appliedthe same set of data that we used in the previous experimentalpart Outputs from the classifier produce sets of values thatare assigned to each recognized training pattern in the giventest time series It is important to appreciate what can beconsidered as an effective criterion related to consensusof similarity The proposed threshold resulting from ourexperimental study was determined at least 119901 = 70Figure 9 shows a comparison of patterns how were learned(S2 S3 H3 train) and how were recognized in test time series(S2 S3 H3 test) The neural network is able to discover someconnections which are almost imperceptible Illustration ofsome recognized patterns that occur in ECG time seriesis shown in Figure 8 Outputs from the classifier carry apredictive character The neural network determines if thetime series belongs to a healthy or sick person on the basis ofthe recognised ECG patterns which appear in the time serieshistory
The methodology of testing is shown in Figure 10 Thismeans that if the test pattern S1 S2 S3 or S4 appearedin ECG waveform with probability 119901S ge 119901 (119901 = 70)thus it was predicted to be ldquoa sick personrdquo Then we workonly with the remaining time series If the test pattern H1H2 H3 or H4 appeared in ECG waveform with probability119901H ge 119901 (119901 = 70) thus it was predicted to be ldquoahealthy personrdquo In all other cases the ECG time series wasunspecifiedWe examined a total of 20 data sets Each of them
contains 101 values that assign 92 possible patternsThewholenumber of examined patterns is 1840 The graph in Figure 11demonstrates a summary of results where ldquosick personsrdquorepresent patterns S1ndashS4 and ldquohealthy personsrdquo representpatterns H1ndashH4 The resulting prediction is based on themethodology see Figure 10
Figure 12 LFLF application
Figure 13 Winning predictor linguistic description (trend-cyclemodel)
42 Time Series Classification and Prediction via LinguisticFuzzy Logic Forecaster We tried also to utilize above pre-sentedmethod of time series analysis through linguistic fuzzylogic forecaster (LFLF) [15] see Figure 12
Basic usage of the application is to analyse given timeseries and find best predictor with respect to validation partof time series given We evaluate efficiency of predictorsby SMAPE (symmetric mean absolute percentage error) Itenables us to make analysis of trend-cycle of a time seriesand also seasonal part The main advantage lies in predictionbased on transparent linguistic descriptions that provide themodel of a time series behaviour Linguistic variables are ofthe following types
(i) value we directly mean the components of the fuzzytransform
(ii) difference first order differences of fuzzy transformcomponents that are given as follows differencesbetween components Δ119883
119894= 119883119894minus 119883119894minus1
8 The Scientific World Journal
Healthypattern
typical series (HS)
Sick patterntypical series
(SS)
Tested series (TS)
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
TS is ldquohealthyrdquo TS is ldquosickrdquo
Yes No
Find best predictorfor TS as validation
interval
Find best predictorfor TS as validation
interval
Concatenate series (SS + TS)Concatenate series (HS + TS)
+ TS) forSMAPE (HS SMAPE (SS + TS) for
If SMAPE (HS + TS) lt SMAPE (SS + TS)
Figure 14 Recognition by linguistic fuzzy logic predictors for typical learning series
(iii) second difference these are values of second orderdifferences of components of the fuzzy transform asfollows Δ2119883
119894= Δ119883
119894minus Δ119883119894minus1
LFLF application enables us to define minimal andmaximal number of these particular variables in a ruleof linguistic description as well as the total number ofantecedent variables
A rule consisting of these variables has the followingstructure and can be described as a signature (fuzzy rulesdescribing the trend-cycle model) Particularly 119878 denotes the
trend-cycle components 119889119878 their differences and 1198892119878 theirsecond order differences The argument (119905) (119905 minus 1) and soforth denotes the time lag of the component
For example taking signature 119878(119905)amp119889119878(119905) rarr 119889119878(119905 + 1)denotes the fact that119883
119894and Δ119883
119894are the antecedent variables
and Δ119883119894+1
is the consequent variable of the winning modeland hence we deal with rules of the form
IF 119883119894is 119860119894AND Δ119883
119894is 119860Δ119894
THEN Δ119883119894+1
is 119860Delta119894(11)
The Scientific World Journal 9
Table 2 Example of algorithm evaluation on 10 ldquohealthyrdquo and 10 ldquosickrdquo patients
Every single fuzzy rule can be taken as a sentence ofnatural language for example first rule from Figure 13
IF 119883119894is ml sm AND Δ119883
119894is qr sm THEN Δ119883
119894+1is minusme
may be read as followsIf the number of cars sold in the current year is more or
less small and the half-year sales increment is quite roughlysmall then the upcoming half-year increment will be negativemedium
421 Recognition of ldquoHealthyrdquo and ldquoSickrdquo Patterns by LFLFOur method to use linguistic fuzzy logic forecasting isbased on simple idea that best predictor learning from bothldquohealthyrdquo and ldquosickrdquo pattern samples respectively can be usedfor validation with tested pattern taken as validation part ofthe seriesThen we can evaluate SMAPE for both these casescompound series SMAPE (ldquohealthyrdquo + tested) and SMAPE(ldquosickrdquo + tested)
If SMAPE (ldquohealthyrdquo + tested) lt SMAPE (ldquosickrdquo + tested)then the tested pattern is supposed to be ldquohealthyrdquo otherwisethe tested pattern is supposed to be ldquosickrdquo
The idea is schematically shown in Figure 14For testing purposes we created two necessary typical
learning time series ldquohealthyrdquo (HS) and ldquosickrdquo (SS) accordingto the algorithm above They both consist of 1010 samplesmade from 10 typical series of ldquohealthyrdquo and ldquosickrdquo patientswith 101 measured ECG values Then we have created 10concatenated series according to the scheme in Figure 14with 10 randomly selected patients with ldquohealthyrdquo ECGmeasurement that is 20 files were produced (10x HS + TSand 10x SS + TS) The same concatenated series were alsomade from 10 ldquosickrdquo patients measurements This made usadditional 20 files with concatenated series (10x HS + TS and
0102030405060708090
100
Sick Healthy
FalseTrue
()
Figure 15 Experimental results LFLF
10x SS + TS) For 20 patients (Table 2) tested ECG we have2 concatenated series giving SMAPE (HS + TS) and SMAPE(SS + TS)
Our method based on LFLF proved very good results forright identification of sick patient records Nevertheless itproduces large amount of false positive identification of sickpattern for healthy patients (Figure 15) This result is con-sistent with our approach using neural networks Of courseour preliminary research has a limited extent and should beperceived only as narrative result which shows interestingproperties especially in complementation of neural networkresults
10 The Scientific World Journal
5 Conclusion
In this paper a short introduction into the field of ECGwaves recognition using backpropagation neural network hasbeen given Main objective was to recognise the normalcycles and arrhythmias and perform further diagnosis Weproposed two detection systems that have been created withusage of neural networks One of them is adapted accordingto the training set Here each pattern represents the wholeone ECG cycle Then an output unit represents a diagnose01 a healthysick person The second one approach usesneural network in which training set contains two differentgroups of patterns for healthysick persons According to theresults of experimental studies it can be stated that ECGwaves patterns were successfully extracted in given timeseries and recognised using suggestedmethod as can be seenfrom figures in Experimental Result section It might resultin better mapping of the time series behaviour for betterprediction
Both approaches were able to predict with high probabil-ity if the ECG time series represents sick or healthy persons Itis interesting that a sick diagnose was recognised with higheraccuracy in both experimental works
The third approach based on LFLF is currently only inthe stage of preliminary experiments but it conforms to theformer results based on neural networks This approach isnovel and could be good supplement to other soft-computingmethods for this task
Disclaimer
Anyopinions findings and conclusions or recommendationsexpressed in this material are those of the authors and do notnecessarily reflect the views of the sponsors
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
The research described here has been financially sup-ported by University of Ostrava Grants SGS16PrF2014 andSGS14PrF2014
References
[1] P Tangkraingkij C Lursinsap S Sanguansintukul and TDesudchit ldquoSelecting relevant EEG signal locations for personalidentification problem using ICA and neural networkrdquo inProceedings of the 8th IEEEACIS International Conference onComputer and Information Science (ICIS rsquo09) pp 616ndash621 June2009
[2] C D Nugent J A C Webb N D Black and G T H WrightldquoElectrocardiogram 2 classificationrdquo Automedica vol 17 pp281ndash306 1999
[3] S Russell and P Norvig Artificial IntelligencemdashA ModernApproach Prentice Hall 2nd edition 2003
[4] C D Nugent J A Lopez A E Smith and N D BlackldquoReverse engineering of neural network classifiers in medicalapplicationsrdquo in Proceedings of the 7th EFOMP Congress vol 17p 184 Physica Medica 2001
[5] L Fausett Fundamentals of Neural Network PrenticeHall 1994[6] D E Rumelhart G E Hinton and R J Williams ldquoLearning
[7] V Novak Fuzzy Sets and Their Applications Adam HilgerBristol UK 1989
[8] S Aouaouda M Chadli P Shi and H R Karimi ldquoDiscrete-time119867
minus119867infinsensor fault detection observer design for nonlin-
ear systems with parameter uncertaintyrdquo International Journalof Robust and Nonlinear Control 2013
[9] M Chadli A Abdo and S X Ding ldquo119867minus119867infin
fault detectionfilter design for discrete-time Takagi-Sugeno fuzzy systemrdquoAutomatica vol 49 no 7 pp 1996ndash2005 2013
[10] M Chadli and H R Karimi ldquoRobust observer design forunknown inputs takagi-sugeno modelsrdquo IEEE Transactions onFuzzy Systems vol 21 no 1 pp 158ndash164 2013
[11] S Aouaouda M Chadli and H Karimi ldquoRobust static output-feedback controller design against sensor failure for vehicledynamicsrdquo IET Control Theory amp Applications vol 8 no 9 pp728ndash737 2014
[12] I Perfilieva V Novak V Pavliska A Dvorak andM StepnickaldquoAnalysis and prediction of time series using fuzzy transformrdquoin Proceedings of the International Joint Conference on NeuralNetworks (IJCNN rsquo08) pp 3876ndash3880 Hong Kong June 2008
[13] E Volna M Kotyrba and R Jarusek ldquoMulti-classifier basedon Elliott waversquos recognitionrdquo Computers amp Mathematics withApplications vol 66 no 2 pp 213ndash225 2013
[14] E Volna M Kotyrba and R Jarusek ldquoPrediction by means ofElliott waves recognitionrdquo in Nostradamus Modern Methods ofPrediction Modeling and Analysis of Nonlinear Systems vol 192of Advances in Intelligent Systems and Computing pp 241ndash250Springer Berlin Germany 2013
[15] H Habiballa V Pavliska and A Dvorak ldquoSoftware systemfor time series prediction based on F-transform and linguisticrulesrdquo in Proceedings of the 8th International Conference onApplied Mathematics Aplimat pp 381ndash386 Bratislava Slovakia2009
Figure 5 Experimental results test error for healthysick persons
002040608
112
1 2 3 4 5 6 7 8 9 10
H1H2
H3H4
Figure 6 Patterns representing healthy persons
which was connected to a computer via bidirectional parallelcable (CETRONICS) Technical parameters of the AD con-verter (8-bit conversion) were the following
(i) 3 measuring ranges(ii) measuring of a frequency of AC voltage at any
channel(iii) autoranging for measuring the frequency of 100Hz
1 kHz and 10 kHz(iv) input resistance of 300 kΩ(v) measurement accuracy 1
Technical parameters of the DA converter (a pro-grammable voltage source plusmn10V) were the following
(i) maximum current consumption of 15mA (after opti-mizing 4A at the output)
(ii) power of the converter plusmn15 V (stabilized)
4 Experimental Results
41 Time Series Classification and Prediction via NeuralNetworks The training set consisted of modified ECG wave-forms We used a backpropagation neural network withtopology 101-10-1 The output unit represents a diagnose 01a healthysick person A smaller number of inputs would
002040608
112
1 2 3 4 5 6 7 8 9 10
S1S2
S3S4
Figure 7 Patterns representing sick persons
0
500
1000
1500
1 26 51 76 101
Sick persons
S2S3
S4
minus500
Figure 8 Some recognized patterns that occur in ECG time series
not be appropriate due to the nature of the ECG waveformWe use 34 ECG time series associated with sick personsand 36 ECG time series associated with healthy persons25 time series of each group were used as a training setand the rest as a test set Figure 4 shows a comparison ofmean values of ECG waveforms for healthysick persons Weused the backpropagation method [5 6] for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 The conducted experimental studiesalso showed that training patterns are mixed randomly ineach cycle of adaptation This ensures their greater diversitywhich acts as a measure of system stability Uniform systemin a crisis usually collapses entirely while system with suchdiversity of trained patterns remains functional despite ofcrisis of its individual parts The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
The test set consisted of 20 samples (11 health and 9sick persons) that were not included in the training set Thesummary results for this type of experiment are shown in agraph in Figure 5 For clarity the results of testing are givenin percentage The average test error was 0194 A healthypopulation was detected with an average error of 0263 andsick population with an average error of 0109
411 Pattern Recognition Classifier Leading to Prediction Forthe purpose of adaptation of the pattern recognition classifierit is necessary to remark that determination of trainingpatterns is one of the key tasks Improperly chosen patternscan lead to confusion of neural networks During our exper-imental work we made some study which included ECGpattern recognition When creating appropriate patterns ofthe training set we used characteristic curves shown as mean
6 The Scientific World Journal
0100
1 2 3 4 5 6 7 8 9 10
S2 trainS2 test
S2 test
minus500
minus400
minus300
minus200
minus100
050
100150200250
1 2 3 4 5 6 7 8 9 10
S3 trainS3 test
S3 test
minus150
minus100
minus50
0100200300400500600
1 2 3 4 5 6 7 8 9 10
Train H3Test H3
minus200
minus100
Figure 9 Training patterns their representation in used test sets
values from ECG waveforms for healthy and sick persons(Figure 4) We use two different groups of patterns PatternsH1ndashH4 (Figure 6) represent healthy persons and patterns S1ndashS4 (Figure 7) represent sick personsThe whole training set isshown in Table 1
Pattern recognition classifier is based on backpropagationneural network and is able to recognise wave structures ingiven time series [13 14] Artificial neural networks needtraining sets for their adaptation In our experimental workthe training set consisted of 8 patterns representing thebasic structure of the various waves in ECG graphs seeFigures 6 and 7 Input data is sequences always including119899 consecutive numbers which are transformed into interval⟨0 1⟩ by formula (10) Samples are adjusted for the needs ofbackpropagation networks with sigmoid activation functionin this way [5 6]
1199091015840119895=
119909119895minusmin (119909
119894 119909
119894+119899minus1)
max (119909119894 119909
119894+119899minus1) minusmin (119909
119894 119909
119894+119899minus1)
(119895 = 119894 119894 + 119899 minus 1)
(10)
where 1199091015840119895is normalized output value of the 119895th neuron (119895 =
119894 119894 + 119899minus1) and (119909119894 119909119894+119899minus1
) are 119899minus1 consecutive outputvalues that specify sequences (patterns) from the trainingset (eg training pars of input and corresponding outputvectors) Input vector contains 10 components Output vectorhas got 8 components and each output unit represents oneof 8 different types of ECG wave samples A neural networkarchitecture is 10-10-8 (eg 10 units in the input layer 10 unitsin the hidden layer and 8 units in the output layer)The net isfully connected Adaptation of the neural network starts withrandomly generated weight values
We used the backpropagation method for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 We have utilized our experience fromearlier times that is training patterns were mixed randomlyin each cycle of adaptation The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
In order to test the efficiency of the method we appliedthe same set of data that we used in the previous experimentalpart Outputs from the classifier produce sets of values thatare assigned to each recognized training pattern in the giventest time series It is important to appreciate what can beconsidered as an effective criterion related to consensusof similarity The proposed threshold resulting from ourexperimental study was determined at least 119901 = 70Figure 9 shows a comparison of patterns how were learned(S2 S3 H3 train) and how were recognized in test time series(S2 S3 H3 test) The neural network is able to discover someconnections which are almost imperceptible Illustration ofsome recognized patterns that occur in ECG time seriesis shown in Figure 8 Outputs from the classifier carry apredictive character The neural network determines if thetime series belongs to a healthy or sick person on the basis ofthe recognised ECG patterns which appear in the time serieshistory
The methodology of testing is shown in Figure 10 Thismeans that if the test pattern S1 S2 S3 or S4 appearedin ECG waveform with probability 119901S ge 119901 (119901 = 70)thus it was predicted to be ldquoa sick personrdquo Then we workonly with the remaining time series If the test pattern H1H2 H3 or H4 appeared in ECG waveform with probability119901H ge 119901 (119901 = 70) thus it was predicted to be ldquoahealthy personrdquo In all other cases the ECG time series wasunspecifiedWe examined a total of 20 data sets Each of them
contains 101 values that assign 92 possible patternsThewholenumber of examined patterns is 1840 The graph in Figure 11demonstrates a summary of results where ldquosick personsrdquorepresent patterns S1ndashS4 and ldquohealthy personsrdquo representpatterns H1ndashH4 The resulting prediction is based on themethodology see Figure 10
Figure 12 LFLF application
Figure 13 Winning predictor linguistic description (trend-cyclemodel)
42 Time Series Classification and Prediction via LinguisticFuzzy Logic Forecaster We tried also to utilize above pre-sentedmethod of time series analysis through linguistic fuzzylogic forecaster (LFLF) [15] see Figure 12
Basic usage of the application is to analyse given timeseries and find best predictor with respect to validation partof time series given We evaluate efficiency of predictorsby SMAPE (symmetric mean absolute percentage error) Itenables us to make analysis of trend-cycle of a time seriesand also seasonal part The main advantage lies in predictionbased on transparent linguistic descriptions that provide themodel of a time series behaviour Linguistic variables are ofthe following types
(i) value we directly mean the components of the fuzzytransform
(ii) difference first order differences of fuzzy transformcomponents that are given as follows differencesbetween components Δ119883
119894= 119883119894minus 119883119894minus1
8 The Scientific World Journal
Healthypattern
typical series (HS)
Sick patterntypical series
(SS)
Tested series (TS)
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
TS is ldquohealthyrdquo TS is ldquosickrdquo
Yes No
Find best predictorfor TS as validation
interval
Find best predictorfor TS as validation
interval
Concatenate series (SS + TS)Concatenate series (HS + TS)
+ TS) forSMAPE (HS SMAPE (SS + TS) for
If SMAPE (HS + TS) lt SMAPE (SS + TS)
Figure 14 Recognition by linguistic fuzzy logic predictors for typical learning series
(iii) second difference these are values of second orderdifferences of components of the fuzzy transform asfollows Δ2119883
119894= Δ119883
119894minus Δ119883119894minus1
LFLF application enables us to define minimal andmaximal number of these particular variables in a ruleof linguistic description as well as the total number ofantecedent variables
A rule consisting of these variables has the followingstructure and can be described as a signature (fuzzy rulesdescribing the trend-cycle model) Particularly 119878 denotes the
trend-cycle components 119889119878 their differences and 1198892119878 theirsecond order differences The argument (119905) (119905 minus 1) and soforth denotes the time lag of the component
For example taking signature 119878(119905)amp119889119878(119905) rarr 119889119878(119905 + 1)denotes the fact that119883
119894and Δ119883
119894are the antecedent variables
and Δ119883119894+1
is the consequent variable of the winning modeland hence we deal with rules of the form
IF 119883119894is 119860119894AND Δ119883
119894is 119860Δ119894
THEN Δ119883119894+1
is 119860Delta119894(11)
The Scientific World Journal 9
Table 2 Example of algorithm evaluation on 10 ldquohealthyrdquo and 10 ldquosickrdquo patients
Every single fuzzy rule can be taken as a sentence ofnatural language for example first rule from Figure 13
IF 119883119894is ml sm AND Δ119883
119894is qr sm THEN Δ119883
119894+1is minusme
may be read as followsIf the number of cars sold in the current year is more or
less small and the half-year sales increment is quite roughlysmall then the upcoming half-year increment will be negativemedium
421 Recognition of ldquoHealthyrdquo and ldquoSickrdquo Patterns by LFLFOur method to use linguistic fuzzy logic forecasting isbased on simple idea that best predictor learning from bothldquohealthyrdquo and ldquosickrdquo pattern samples respectively can be usedfor validation with tested pattern taken as validation part ofthe seriesThen we can evaluate SMAPE for both these casescompound series SMAPE (ldquohealthyrdquo + tested) and SMAPE(ldquosickrdquo + tested)
If SMAPE (ldquohealthyrdquo + tested) lt SMAPE (ldquosickrdquo + tested)then the tested pattern is supposed to be ldquohealthyrdquo otherwisethe tested pattern is supposed to be ldquosickrdquo
The idea is schematically shown in Figure 14For testing purposes we created two necessary typical
learning time series ldquohealthyrdquo (HS) and ldquosickrdquo (SS) accordingto the algorithm above They both consist of 1010 samplesmade from 10 typical series of ldquohealthyrdquo and ldquosickrdquo patientswith 101 measured ECG values Then we have created 10concatenated series according to the scheme in Figure 14with 10 randomly selected patients with ldquohealthyrdquo ECGmeasurement that is 20 files were produced (10x HS + TSand 10x SS + TS) The same concatenated series were alsomade from 10 ldquosickrdquo patients measurements This made usadditional 20 files with concatenated series (10x HS + TS and
0102030405060708090
100
Sick Healthy
FalseTrue
()
Figure 15 Experimental results LFLF
10x SS + TS) For 20 patients (Table 2) tested ECG we have2 concatenated series giving SMAPE (HS + TS) and SMAPE(SS + TS)
Our method based on LFLF proved very good results forright identification of sick patient records Nevertheless itproduces large amount of false positive identification of sickpattern for healthy patients (Figure 15) This result is con-sistent with our approach using neural networks Of courseour preliminary research has a limited extent and should beperceived only as narrative result which shows interestingproperties especially in complementation of neural networkresults
10 The Scientific World Journal
5 Conclusion
In this paper a short introduction into the field of ECGwaves recognition using backpropagation neural network hasbeen given Main objective was to recognise the normalcycles and arrhythmias and perform further diagnosis Weproposed two detection systems that have been created withusage of neural networks One of them is adapted accordingto the training set Here each pattern represents the wholeone ECG cycle Then an output unit represents a diagnose01 a healthysick person The second one approach usesneural network in which training set contains two differentgroups of patterns for healthysick persons According to theresults of experimental studies it can be stated that ECGwaves patterns were successfully extracted in given timeseries and recognised using suggestedmethod as can be seenfrom figures in Experimental Result section It might resultin better mapping of the time series behaviour for betterprediction
Both approaches were able to predict with high probabil-ity if the ECG time series represents sick or healthy persons Itis interesting that a sick diagnose was recognised with higheraccuracy in both experimental works
The third approach based on LFLF is currently only inthe stage of preliminary experiments but it conforms to theformer results based on neural networks This approach isnovel and could be good supplement to other soft-computingmethods for this task
Disclaimer
Anyopinions findings and conclusions or recommendationsexpressed in this material are those of the authors and do notnecessarily reflect the views of the sponsors
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
The research described here has been financially sup-ported by University of Ostrava Grants SGS16PrF2014 andSGS14PrF2014
References
[1] P Tangkraingkij C Lursinsap S Sanguansintukul and TDesudchit ldquoSelecting relevant EEG signal locations for personalidentification problem using ICA and neural networkrdquo inProceedings of the 8th IEEEACIS International Conference onComputer and Information Science (ICIS rsquo09) pp 616ndash621 June2009
[2] C D Nugent J A C Webb N D Black and G T H WrightldquoElectrocardiogram 2 classificationrdquo Automedica vol 17 pp281ndash306 1999
[3] S Russell and P Norvig Artificial IntelligencemdashA ModernApproach Prentice Hall 2nd edition 2003
[4] C D Nugent J A Lopez A E Smith and N D BlackldquoReverse engineering of neural network classifiers in medicalapplicationsrdquo in Proceedings of the 7th EFOMP Congress vol 17p 184 Physica Medica 2001
[5] L Fausett Fundamentals of Neural Network PrenticeHall 1994[6] D E Rumelhart G E Hinton and R J Williams ldquoLearning
[7] V Novak Fuzzy Sets and Their Applications Adam HilgerBristol UK 1989
[8] S Aouaouda M Chadli P Shi and H R Karimi ldquoDiscrete-time119867
minus119867infinsensor fault detection observer design for nonlin-
ear systems with parameter uncertaintyrdquo International Journalof Robust and Nonlinear Control 2013
[9] M Chadli A Abdo and S X Ding ldquo119867minus119867infin
fault detectionfilter design for discrete-time Takagi-Sugeno fuzzy systemrdquoAutomatica vol 49 no 7 pp 1996ndash2005 2013
[10] M Chadli and H R Karimi ldquoRobust observer design forunknown inputs takagi-sugeno modelsrdquo IEEE Transactions onFuzzy Systems vol 21 no 1 pp 158ndash164 2013
[11] S Aouaouda M Chadli and H Karimi ldquoRobust static output-feedback controller design against sensor failure for vehicledynamicsrdquo IET Control Theory amp Applications vol 8 no 9 pp728ndash737 2014
[12] I Perfilieva V Novak V Pavliska A Dvorak andM StepnickaldquoAnalysis and prediction of time series using fuzzy transformrdquoin Proceedings of the International Joint Conference on NeuralNetworks (IJCNN rsquo08) pp 3876ndash3880 Hong Kong June 2008
[13] E Volna M Kotyrba and R Jarusek ldquoMulti-classifier basedon Elliott waversquos recognitionrdquo Computers amp Mathematics withApplications vol 66 no 2 pp 213ndash225 2013
[14] E Volna M Kotyrba and R Jarusek ldquoPrediction by means ofElliott waves recognitionrdquo in Nostradamus Modern Methods ofPrediction Modeling and Analysis of Nonlinear Systems vol 192of Advances in Intelligent Systems and Computing pp 241ndash250Springer Berlin Germany 2013
[15] H Habiballa V Pavliska and A Dvorak ldquoSoftware systemfor time series prediction based on F-transform and linguisticrulesrdquo in Proceedings of the 8th International Conference onApplied Mathematics Aplimat pp 381ndash386 Bratislava Slovakia2009
Figure 9 Training patterns their representation in used test sets
values from ECG waveforms for healthy and sick persons(Figure 4) We use two different groups of patterns PatternsH1ndashH4 (Figure 6) represent healthy persons and patterns S1ndashS4 (Figure 7) represent sick personsThe whole training set isshown in Table 1
Pattern recognition classifier is based on backpropagationneural network and is able to recognise wave structures ingiven time series [13 14] Artificial neural networks needtraining sets for their adaptation In our experimental workthe training set consisted of 8 patterns representing thebasic structure of the various waves in ECG graphs seeFigures 6 and 7 Input data is sequences always including119899 consecutive numbers which are transformed into interval⟨0 1⟩ by formula (10) Samples are adjusted for the needs ofbackpropagation networks with sigmoid activation functionin this way [5 6]
1199091015840119895=
119909119895minusmin (119909
119894 119909
119894+119899minus1)
max (119909119894 119909
119894+119899minus1) minusmin (119909
119894 119909
119894+119899minus1)
(119895 = 119894 119894 + 119899 minus 1)
(10)
where 1199091015840119895is normalized output value of the 119895th neuron (119895 =
119894 119894 + 119899minus1) and (119909119894 119909119894+119899minus1
) are 119899minus1 consecutive outputvalues that specify sequences (patterns) from the trainingset (eg training pars of input and corresponding outputvectors) Input vector contains 10 components Output vectorhas got 8 components and each output unit represents oneof 8 different types of ECG wave samples A neural networkarchitecture is 10-10-8 (eg 10 units in the input layer 10 unitsin the hidden layer and 8 units in the output layer)The net isfully connected Adaptation of the neural network starts withrandomly generated weight values
We used the backpropagation method for the adaptationwith the following parameters the learning rate value is 01and momentum is 0 We have utilized our experience fromearlier times that is training patterns were mixed randomlyin each cycle of adaptation The condition of end of theadaptation algorithm specified the limit value of the overallnetwork error 119864 lt 01
In order to test the efficiency of the method we appliedthe same set of data that we used in the previous experimentalpart Outputs from the classifier produce sets of values thatare assigned to each recognized training pattern in the giventest time series It is important to appreciate what can beconsidered as an effective criterion related to consensusof similarity The proposed threshold resulting from ourexperimental study was determined at least 119901 = 70Figure 9 shows a comparison of patterns how were learned(S2 S3 H3 train) and how were recognized in test time series(S2 S3 H3 test) The neural network is able to discover someconnections which are almost imperceptible Illustration ofsome recognized patterns that occur in ECG time seriesis shown in Figure 8 Outputs from the classifier carry apredictive character The neural network determines if thetime series belongs to a healthy or sick person on the basis ofthe recognised ECG patterns which appear in the time serieshistory
The methodology of testing is shown in Figure 10 Thismeans that if the test pattern S1 S2 S3 or S4 appearedin ECG waveform with probability 119901S ge 119901 (119901 = 70)thus it was predicted to be ldquoa sick personrdquo Then we workonly with the remaining time series If the test pattern H1H2 H3 or H4 appeared in ECG waveform with probability119901H ge 119901 (119901 = 70) thus it was predicted to be ldquoahealthy personrdquo In all other cases the ECG time series wasunspecifiedWe examined a total of 20 data sets Each of them
contains 101 values that assign 92 possible patternsThewholenumber of examined patterns is 1840 The graph in Figure 11demonstrates a summary of results where ldquosick personsrdquorepresent patterns S1ndashS4 and ldquohealthy personsrdquo representpatterns H1ndashH4 The resulting prediction is based on themethodology see Figure 10
Figure 12 LFLF application
Figure 13 Winning predictor linguistic description (trend-cyclemodel)
42 Time Series Classification and Prediction via LinguisticFuzzy Logic Forecaster We tried also to utilize above pre-sentedmethod of time series analysis through linguistic fuzzylogic forecaster (LFLF) [15] see Figure 12
Basic usage of the application is to analyse given timeseries and find best predictor with respect to validation partof time series given We evaluate efficiency of predictorsby SMAPE (symmetric mean absolute percentage error) Itenables us to make analysis of trend-cycle of a time seriesand also seasonal part The main advantage lies in predictionbased on transparent linguistic descriptions that provide themodel of a time series behaviour Linguistic variables are ofthe following types
(i) value we directly mean the components of the fuzzytransform
(ii) difference first order differences of fuzzy transformcomponents that are given as follows differencesbetween components Δ119883
119894= 119883119894minus 119883119894minus1
8 The Scientific World Journal
Healthypattern
typical series (HS)
Sick patterntypical series
(SS)
Tested series (TS)
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
TS is ldquohealthyrdquo TS is ldquosickrdquo
Yes No
Find best predictorfor TS as validation
interval
Find best predictorfor TS as validation
interval
Concatenate series (SS + TS)Concatenate series (HS + TS)
+ TS) forSMAPE (HS SMAPE (SS + TS) for
If SMAPE (HS + TS) lt SMAPE (SS + TS)
Figure 14 Recognition by linguistic fuzzy logic predictors for typical learning series
(iii) second difference these are values of second orderdifferences of components of the fuzzy transform asfollows Δ2119883
119894= Δ119883
119894minus Δ119883119894minus1
LFLF application enables us to define minimal andmaximal number of these particular variables in a ruleof linguistic description as well as the total number ofantecedent variables
A rule consisting of these variables has the followingstructure and can be described as a signature (fuzzy rulesdescribing the trend-cycle model) Particularly 119878 denotes the
trend-cycle components 119889119878 their differences and 1198892119878 theirsecond order differences The argument (119905) (119905 minus 1) and soforth denotes the time lag of the component
For example taking signature 119878(119905)amp119889119878(119905) rarr 119889119878(119905 + 1)denotes the fact that119883
119894and Δ119883
119894are the antecedent variables
and Δ119883119894+1
is the consequent variable of the winning modeland hence we deal with rules of the form
IF 119883119894is 119860119894AND Δ119883
119894is 119860Δ119894
THEN Δ119883119894+1
is 119860Delta119894(11)
The Scientific World Journal 9
Table 2 Example of algorithm evaluation on 10 ldquohealthyrdquo and 10 ldquosickrdquo patients
Every single fuzzy rule can be taken as a sentence ofnatural language for example first rule from Figure 13
IF 119883119894is ml sm AND Δ119883
119894is qr sm THEN Δ119883
119894+1is minusme
may be read as followsIf the number of cars sold in the current year is more or
less small and the half-year sales increment is quite roughlysmall then the upcoming half-year increment will be negativemedium
421 Recognition of ldquoHealthyrdquo and ldquoSickrdquo Patterns by LFLFOur method to use linguistic fuzzy logic forecasting isbased on simple idea that best predictor learning from bothldquohealthyrdquo and ldquosickrdquo pattern samples respectively can be usedfor validation with tested pattern taken as validation part ofthe seriesThen we can evaluate SMAPE for both these casescompound series SMAPE (ldquohealthyrdquo + tested) and SMAPE(ldquosickrdquo + tested)
If SMAPE (ldquohealthyrdquo + tested) lt SMAPE (ldquosickrdquo + tested)then the tested pattern is supposed to be ldquohealthyrdquo otherwisethe tested pattern is supposed to be ldquosickrdquo
The idea is schematically shown in Figure 14For testing purposes we created two necessary typical
learning time series ldquohealthyrdquo (HS) and ldquosickrdquo (SS) accordingto the algorithm above They both consist of 1010 samplesmade from 10 typical series of ldquohealthyrdquo and ldquosickrdquo patientswith 101 measured ECG values Then we have created 10concatenated series according to the scheme in Figure 14with 10 randomly selected patients with ldquohealthyrdquo ECGmeasurement that is 20 files were produced (10x HS + TSand 10x SS + TS) The same concatenated series were alsomade from 10 ldquosickrdquo patients measurements This made usadditional 20 files with concatenated series (10x HS + TS and
0102030405060708090
100
Sick Healthy
FalseTrue
()
Figure 15 Experimental results LFLF
10x SS + TS) For 20 patients (Table 2) tested ECG we have2 concatenated series giving SMAPE (HS + TS) and SMAPE(SS + TS)
Our method based on LFLF proved very good results forright identification of sick patient records Nevertheless itproduces large amount of false positive identification of sickpattern for healthy patients (Figure 15) This result is con-sistent with our approach using neural networks Of courseour preliminary research has a limited extent and should beperceived only as narrative result which shows interestingproperties especially in complementation of neural networkresults
10 The Scientific World Journal
5 Conclusion
In this paper a short introduction into the field of ECGwaves recognition using backpropagation neural network hasbeen given Main objective was to recognise the normalcycles and arrhythmias and perform further diagnosis Weproposed two detection systems that have been created withusage of neural networks One of them is adapted accordingto the training set Here each pattern represents the wholeone ECG cycle Then an output unit represents a diagnose01 a healthysick person The second one approach usesneural network in which training set contains two differentgroups of patterns for healthysick persons According to theresults of experimental studies it can be stated that ECGwaves patterns were successfully extracted in given timeseries and recognised using suggestedmethod as can be seenfrom figures in Experimental Result section It might resultin better mapping of the time series behaviour for betterprediction
Both approaches were able to predict with high probabil-ity if the ECG time series represents sick or healthy persons Itis interesting that a sick diagnose was recognised with higheraccuracy in both experimental works
The third approach based on LFLF is currently only inthe stage of preliminary experiments but it conforms to theformer results based on neural networks This approach isnovel and could be good supplement to other soft-computingmethods for this task
Disclaimer
Anyopinions findings and conclusions or recommendationsexpressed in this material are those of the authors and do notnecessarily reflect the views of the sponsors
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
The research described here has been financially sup-ported by University of Ostrava Grants SGS16PrF2014 andSGS14PrF2014
References
[1] P Tangkraingkij C Lursinsap S Sanguansintukul and TDesudchit ldquoSelecting relevant EEG signal locations for personalidentification problem using ICA and neural networkrdquo inProceedings of the 8th IEEEACIS International Conference onComputer and Information Science (ICIS rsquo09) pp 616ndash621 June2009
[2] C D Nugent J A C Webb N D Black and G T H WrightldquoElectrocardiogram 2 classificationrdquo Automedica vol 17 pp281ndash306 1999
[3] S Russell and P Norvig Artificial IntelligencemdashA ModernApproach Prentice Hall 2nd edition 2003
[4] C D Nugent J A Lopez A E Smith and N D BlackldquoReverse engineering of neural network classifiers in medicalapplicationsrdquo in Proceedings of the 7th EFOMP Congress vol 17p 184 Physica Medica 2001
[5] L Fausett Fundamentals of Neural Network PrenticeHall 1994[6] D E Rumelhart G E Hinton and R J Williams ldquoLearning
[7] V Novak Fuzzy Sets and Their Applications Adam HilgerBristol UK 1989
[8] S Aouaouda M Chadli P Shi and H R Karimi ldquoDiscrete-time119867
minus119867infinsensor fault detection observer design for nonlin-
ear systems with parameter uncertaintyrdquo International Journalof Robust and Nonlinear Control 2013
[9] M Chadli A Abdo and S X Ding ldquo119867minus119867infin
fault detectionfilter design for discrete-time Takagi-Sugeno fuzzy systemrdquoAutomatica vol 49 no 7 pp 1996ndash2005 2013
[10] M Chadli and H R Karimi ldquoRobust observer design forunknown inputs takagi-sugeno modelsrdquo IEEE Transactions onFuzzy Systems vol 21 no 1 pp 158ndash164 2013
[11] S Aouaouda M Chadli and H Karimi ldquoRobust static output-feedback controller design against sensor failure for vehicledynamicsrdquo IET Control Theory amp Applications vol 8 no 9 pp728ndash737 2014
[12] I Perfilieva V Novak V Pavliska A Dvorak andM StepnickaldquoAnalysis and prediction of time series using fuzzy transformrdquoin Proceedings of the International Joint Conference on NeuralNetworks (IJCNN rsquo08) pp 3876ndash3880 Hong Kong June 2008
[13] E Volna M Kotyrba and R Jarusek ldquoMulti-classifier basedon Elliott waversquos recognitionrdquo Computers amp Mathematics withApplications vol 66 no 2 pp 213ndash225 2013
[14] E Volna M Kotyrba and R Jarusek ldquoPrediction by means ofElliott waves recognitionrdquo in Nostradamus Modern Methods ofPrediction Modeling and Analysis of Nonlinear Systems vol 192of Advances in Intelligent Systems and Computing pp 241ndash250Springer Berlin Germany 2013
[15] H Habiballa V Pavliska and A Dvorak ldquoSoftware systemfor time series prediction based on F-transform and linguisticrulesrdquo in Proceedings of the 8th International Conference onApplied Mathematics Aplimat pp 381ndash386 Bratislava Slovakia2009
contains 101 values that assign 92 possible patternsThewholenumber of examined patterns is 1840 The graph in Figure 11demonstrates a summary of results where ldquosick personsrdquorepresent patterns S1ndashS4 and ldquohealthy personsrdquo representpatterns H1ndashH4 The resulting prediction is based on themethodology see Figure 10
Figure 12 LFLF application
Figure 13 Winning predictor linguistic description (trend-cyclemodel)
42 Time Series Classification and Prediction via LinguisticFuzzy Logic Forecaster We tried also to utilize above pre-sentedmethod of time series analysis through linguistic fuzzylogic forecaster (LFLF) [15] see Figure 12
Basic usage of the application is to analyse given timeseries and find best predictor with respect to validation partof time series given We evaluate efficiency of predictorsby SMAPE (symmetric mean absolute percentage error) Itenables us to make analysis of trend-cycle of a time seriesand also seasonal part The main advantage lies in predictionbased on transparent linguistic descriptions that provide themodel of a time series behaviour Linguistic variables are ofthe following types
(i) value we directly mean the components of the fuzzytransform
(ii) difference first order differences of fuzzy transformcomponents that are given as follows differencesbetween components Δ119883
119894= 119883119894minus 119883119894minus1
8 The Scientific World Journal
Healthypattern
typical series (HS)
Sick patterntypical series
(SS)
Tested series (TS)
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
and set TS as validationinterval
Evaluate bestpredictor
TS validationinterval
TS is ldquohealthyrdquo TS is ldquosickrdquo
Yes No
Find best predictorfor TS as validation
interval
Find best predictorfor TS as validation
interval
Concatenate series (SS + TS)Concatenate series (HS + TS)
+ TS) forSMAPE (HS SMAPE (SS + TS) for
If SMAPE (HS + TS) lt SMAPE (SS + TS)
Figure 14 Recognition by linguistic fuzzy logic predictors for typical learning series
(iii) second difference these are values of second orderdifferences of components of the fuzzy transform asfollows Δ2119883
119894= Δ119883
119894minus Δ119883119894minus1
LFLF application enables us to define minimal andmaximal number of these particular variables in a ruleof linguistic description as well as the total number ofantecedent variables
A rule consisting of these variables has the followingstructure and can be described as a signature (fuzzy rulesdescribing the trend-cycle model) Particularly 119878 denotes the
trend-cycle components 119889119878 their differences and 1198892119878 theirsecond order differences The argument (119905) (119905 minus 1) and soforth denotes the time lag of the component
For example taking signature 119878(119905)amp119889119878(119905) rarr 119889119878(119905 + 1)denotes the fact that119883
119894and Δ119883
119894are the antecedent variables
and Δ119883119894+1
is the consequent variable of the winning modeland hence we deal with rules of the form
IF 119883119894is 119860119894AND Δ119883
119894is 119860Δ119894
THEN Δ119883119894+1
is 119860Delta119894(11)
The Scientific World Journal 9
Table 2 Example of algorithm evaluation on 10 ldquohealthyrdquo and 10 ldquosickrdquo patients
Every single fuzzy rule can be taken as a sentence ofnatural language for example first rule from Figure 13
IF 119883119894is ml sm AND Δ119883
119894is qr sm THEN Δ119883
119894+1is minusme
may be read as followsIf the number of cars sold in the current year is more or
less small and the half-year sales increment is quite roughlysmall then the upcoming half-year increment will be negativemedium
421 Recognition of ldquoHealthyrdquo and ldquoSickrdquo Patterns by LFLFOur method to use linguistic fuzzy logic forecasting isbased on simple idea that best predictor learning from bothldquohealthyrdquo and ldquosickrdquo pattern samples respectively can be usedfor validation with tested pattern taken as validation part ofthe seriesThen we can evaluate SMAPE for both these casescompound series SMAPE (ldquohealthyrdquo + tested) and SMAPE(ldquosickrdquo + tested)
If SMAPE (ldquohealthyrdquo + tested) lt SMAPE (ldquosickrdquo + tested)then the tested pattern is supposed to be ldquohealthyrdquo otherwisethe tested pattern is supposed to be ldquosickrdquo
The idea is schematically shown in Figure 14For testing purposes we created two necessary typical
learning time series ldquohealthyrdquo (HS) and ldquosickrdquo (SS) accordingto the algorithm above They both consist of 1010 samplesmade from 10 typical series of ldquohealthyrdquo and ldquosickrdquo patientswith 101 measured ECG values Then we have created 10concatenated series according to the scheme in Figure 14with 10 randomly selected patients with ldquohealthyrdquo ECGmeasurement that is 20 files were produced (10x HS + TSand 10x SS + TS) The same concatenated series were alsomade from 10 ldquosickrdquo patients measurements This made usadditional 20 files with concatenated series (10x HS + TS and
0102030405060708090
100
Sick Healthy
FalseTrue
()
Figure 15 Experimental results LFLF
10x SS + TS) For 20 patients (Table 2) tested ECG we have2 concatenated series giving SMAPE (HS + TS) and SMAPE(SS + TS)
Our method based on LFLF proved very good results forright identification of sick patient records Nevertheless itproduces large amount of false positive identification of sickpattern for healthy patients (Figure 15) This result is con-sistent with our approach using neural networks Of courseour preliminary research has a limited extent and should beperceived only as narrative result which shows interestingproperties especially in complementation of neural networkresults
10 The Scientific World Journal
5 Conclusion
In this paper a short introduction into the field of ECGwaves recognition using backpropagation neural network hasbeen given Main objective was to recognise the normalcycles and arrhythmias and perform further diagnosis Weproposed two detection systems that have been created withusage of neural networks One of them is adapted accordingto the training set Here each pattern represents the wholeone ECG cycle Then an output unit represents a diagnose01 a healthysick person The second one approach usesneural network in which training set contains two differentgroups of patterns for healthysick persons According to theresults of experimental studies it can be stated that ECGwaves patterns were successfully extracted in given timeseries and recognised using suggestedmethod as can be seenfrom figures in Experimental Result section It might resultin better mapping of the time series behaviour for betterprediction
Both approaches were able to predict with high probabil-ity if the ECG time series represents sick or healthy persons Itis interesting that a sick diagnose was recognised with higheraccuracy in both experimental works
The third approach based on LFLF is currently only inthe stage of preliminary experiments but it conforms to theformer results based on neural networks This approach isnovel and could be good supplement to other soft-computingmethods for this task
Disclaimer
Anyopinions findings and conclusions or recommendationsexpressed in this material are those of the authors and do notnecessarily reflect the views of the sponsors
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
The research described here has been financially sup-ported by University of Ostrava Grants SGS16PrF2014 andSGS14PrF2014
References
[1] P Tangkraingkij C Lursinsap S Sanguansintukul and TDesudchit ldquoSelecting relevant EEG signal locations for personalidentification problem using ICA and neural networkrdquo inProceedings of the 8th IEEEACIS International Conference onComputer and Information Science (ICIS rsquo09) pp 616ndash621 June2009
[2] C D Nugent J A C Webb N D Black and G T H WrightldquoElectrocardiogram 2 classificationrdquo Automedica vol 17 pp281ndash306 1999
[3] S Russell and P Norvig Artificial IntelligencemdashA ModernApproach Prentice Hall 2nd edition 2003
[4] C D Nugent J A Lopez A E Smith and N D BlackldquoReverse engineering of neural network classifiers in medicalapplicationsrdquo in Proceedings of the 7th EFOMP Congress vol 17p 184 Physica Medica 2001
[5] L Fausett Fundamentals of Neural Network PrenticeHall 1994[6] D E Rumelhart G E Hinton and R J Williams ldquoLearning
[7] V Novak Fuzzy Sets and Their Applications Adam HilgerBristol UK 1989
[8] S Aouaouda M Chadli P Shi and H R Karimi ldquoDiscrete-time119867
minus119867infinsensor fault detection observer design for nonlin-
ear systems with parameter uncertaintyrdquo International Journalof Robust and Nonlinear Control 2013
[9] M Chadli A Abdo and S X Ding ldquo119867minus119867infin
fault detectionfilter design for discrete-time Takagi-Sugeno fuzzy systemrdquoAutomatica vol 49 no 7 pp 1996ndash2005 2013
[10] M Chadli and H R Karimi ldquoRobust observer design forunknown inputs takagi-sugeno modelsrdquo IEEE Transactions onFuzzy Systems vol 21 no 1 pp 158ndash164 2013
[11] S Aouaouda M Chadli and H Karimi ldquoRobust static output-feedback controller design against sensor failure for vehicledynamicsrdquo IET Control Theory amp Applications vol 8 no 9 pp728ndash737 2014
[12] I Perfilieva V Novak V Pavliska A Dvorak andM StepnickaldquoAnalysis and prediction of time series using fuzzy transformrdquoin Proceedings of the International Joint Conference on NeuralNetworks (IJCNN rsquo08) pp 3876ndash3880 Hong Kong June 2008
[13] E Volna M Kotyrba and R Jarusek ldquoMulti-classifier basedon Elliott waversquos recognitionrdquo Computers amp Mathematics withApplications vol 66 no 2 pp 213ndash225 2013
[14] E Volna M Kotyrba and R Jarusek ldquoPrediction by means ofElliott waves recognitionrdquo in Nostradamus Modern Methods ofPrediction Modeling and Analysis of Nonlinear Systems vol 192of Advances in Intelligent Systems and Computing pp 241ndash250Springer Berlin Germany 2013
[15] H Habiballa V Pavliska and A Dvorak ldquoSoftware systemfor time series prediction based on F-transform and linguisticrulesrdquo in Proceedings of the 8th International Conference onApplied Mathematics Aplimat pp 381ndash386 Bratislava Slovakia2009
Concatenate series (SS + TS)Concatenate series (HS + TS)
+ TS) forSMAPE (HS SMAPE (SS + TS) for
If SMAPE (HS + TS) lt SMAPE (SS + TS)
Figure 14 Recognition by linguistic fuzzy logic predictors for typical learning series
(iii) second difference these are values of second orderdifferences of components of the fuzzy transform asfollows Δ2119883
119894= Δ119883
119894minus Δ119883119894minus1
LFLF application enables us to define minimal andmaximal number of these particular variables in a ruleof linguistic description as well as the total number ofantecedent variables
A rule consisting of these variables has the followingstructure and can be described as a signature (fuzzy rulesdescribing the trend-cycle model) Particularly 119878 denotes the
trend-cycle components 119889119878 their differences and 1198892119878 theirsecond order differences The argument (119905) (119905 minus 1) and soforth denotes the time lag of the component
For example taking signature 119878(119905)amp119889119878(119905) rarr 119889119878(119905 + 1)denotes the fact that119883
119894and Δ119883
119894are the antecedent variables
and Δ119883119894+1
is the consequent variable of the winning modeland hence we deal with rules of the form
IF 119883119894is 119860119894AND Δ119883
119894is 119860Δ119894
THEN Δ119883119894+1
is 119860Delta119894(11)
The Scientific World Journal 9
Table 2 Example of algorithm evaluation on 10 ldquohealthyrdquo and 10 ldquosickrdquo patients
Every single fuzzy rule can be taken as a sentence ofnatural language for example first rule from Figure 13
IF 119883119894is ml sm AND Δ119883
119894is qr sm THEN Δ119883
119894+1is minusme
may be read as followsIf the number of cars sold in the current year is more or
less small and the half-year sales increment is quite roughlysmall then the upcoming half-year increment will be negativemedium
421 Recognition of ldquoHealthyrdquo and ldquoSickrdquo Patterns by LFLFOur method to use linguistic fuzzy logic forecasting isbased on simple idea that best predictor learning from bothldquohealthyrdquo and ldquosickrdquo pattern samples respectively can be usedfor validation with tested pattern taken as validation part ofthe seriesThen we can evaluate SMAPE for both these casescompound series SMAPE (ldquohealthyrdquo + tested) and SMAPE(ldquosickrdquo + tested)
If SMAPE (ldquohealthyrdquo + tested) lt SMAPE (ldquosickrdquo + tested)then the tested pattern is supposed to be ldquohealthyrdquo otherwisethe tested pattern is supposed to be ldquosickrdquo
The idea is schematically shown in Figure 14For testing purposes we created two necessary typical
learning time series ldquohealthyrdquo (HS) and ldquosickrdquo (SS) accordingto the algorithm above They both consist of 1010 samplesmade from 10 typical series of ldquohealthyrdquo and ldquosickrdquo patientswith 101 measured ECG values Then we have created 10concatenated series according to the scheme in Figure 14with 10 randomly selected patients with ldquohealthyrdquo ECGmeasurement that is 20 files were produced (10x HS + TSand 10x SS + TS) The same concatenated series were alsomade from 10 ldquosickrdquo patients measurements This made usadditional 20 files with concatenated series (10x HS + TS and
0102030405060708090
100
Sick Healthy
FalseTrue
()
Figure 15 Experimental results LFLF
10x SS + TS) For 20 patients (Table 2) tested ECG we have2 concatenated series giving SMAPE (HS + TS) and SMAPE(SS + TS)
Our method based on LFLF proved very good results forright identification of sick patient records Nevertheless itproduces large amount of false positive identification of sickpattern for healthy patients (Figure 15) This result is con-sistent with our approach using neural networks Of courseour preliminary research has a limited extent and should beperceived only as narrative result which shows interestingproperties especially in complementation of neural networkresults
10 The Scientific World Journal
5 Conclusion
In this paper a short introduction into the field of ECGwaves recognition using backpropagation neural network hasbeen given Main objective was to recognise the normalcycles and arrhythmias and perform further diagnosis Weproposed two detection systems that have been created withusage of neural networks One of them is adapted accordingto the training set Here each pattern represents the wholeone ECG cycle Then an output unit represents a diagnose01 a healthysick person The second one approach usesneural network in which training set contains two differentgroups of patterns for healthysick persons According to theresults of experimental studies it can be stated that ECGwaves patterns were successfully extracted in given timeseries and recognised using suggestedmethod as can be seenfrom figures in Experimental Result section It might resultin better mapping of the time series behaviour for betterprediction
Both approaches were able to predict with high probabil-ity if the ECG time series represents sick or healthy persons Itis interesting that a sick diagnose was recognised with higheraccuracy in both experimental works
The third approach based on LFLF is currently only inthe stage of preliminary experiments but it conforms to theformer results based on neural networks This approach isnovel and could be good supplement to other soft-computingmethods for this task
Disclaimer
Anyopinions findings and conclusions or recommendationsexpressed in this material are those of the authors and do notnecessarily reflect the views of the sponsors
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
The research described here has been financially sup-ported by University of Ostrava Grants SGS16PrF2014 andSGS14PrF2014
References
[1] P Tangkraingkij C Lursinsap S Sanguansintukul and TDesudchit ldquoSelecting relevant EEG signal locations for personalidentification problem using ICA and neural networkrdquo inProceedings of the 8th IEEEACIS International Conference onComputer and Information Science (ICIS rsquo09) pp 616ndash621 June2009
[2] C D Nugent J A C Webb N D Black and G T H WrightldquoElectrocardiogram 2 classificationrdquo Automedica vol 17 pp281ndash306 1999
[3] S Russell and P Norvig Artificial IntelligencemdashA ModernApproach Prentice Hall 2nd edition 2003
[4] C D Nugent J A Lopez A E Smith and N D BlackldquoReverse engineering of neural network classifiers in medicalapplicationsrdquo in Proceedings of the 7th EFOMP Congress vol 17p 184 Physica Medica 2001
[5] L Fausett Fundamentals of Neural Network PrenticeHall 1994[6] D E Rumelhart G E Hinton and R J Williams ldquoLearning
[7] V Novak Fuzzy Sets and Their Applications Adam HilgerBristol UK 1989
[8] S Aouaouda M Chadli P Shi and H R Karimi ldquoDiscrete-time119867
minus119867infinsensor fault detection observer design for nonlin-
ear systems with parameter uncertaintyrdquo International Journalof Robust and Nonlinear Control 2013
[9] M Chadli A Abdo and S X Ding ldquo119867minus119867infin
fault detectionfilter design for discrete-time Takagi-Sugeno fuzzy systemrdquoAutomatica vol 49 no 7 pp 1996ndash2005 2013
[10] M Chadli and H R Karimi ldquoRobust observer design forunknown inputs takagi-sugeno modelsrdquo IEEE Transactions onFuzzy Systems vol 21 no 1 pp 158ndash164 2013
[11] S Aouaouda M Chadli and H Karimi ldquoRobust static output-feedback controller design against sensor failure for vehicledynamicsrdquo IET Control Theory amp Applications vol 8 no 9 pp728ndash737 2014
[12] I Perfilieva V Novak V Pavliska A Dvorak andM StepnickaldquoAnalysis and prediction of time series using fuzzy transformrdquoin Proceedings of the International Joint Conference on NeuralNetworks (IJCNN rsquo08) pp 3876ndash3880 Hong Kong June 2008
[13] E Volna M Kotyrba and R Jarusek ldquoMulti-classifier basedon Elliott waversquos recognitionrdquo Computers amp Mathematics withApplications vol 66 no 2 pp 213ndash225 2013
[14] E Volna M Kotyrba and R Jarusek ldquoPrediction by means ofElliott waves recognitionrdquo in Nostradamus Modern Methods ofPrediction Modeling and Analysis of Nonlinear Systems vol 192of Advances in Intelligent Systems and Computing pp 241ndash250Springer Berlin Germany 2013
[15] H Habiballa V Pavliska and A Dvorak ldquoSoftware systemfor time series prediction based on F-transform and linguisticrulesrdquo in Proceedings of the 8th International Conference onApplied Mathematics Aplimat pp 381ndash386 Bratislava Slovakia2009
Every single fuzzy rule can be taken as a sentence ofnatural language for example first rule from Figure 13
IF 119883119894is ml sm AND Δ119883
119894is qr sm THEN Δ119883
119894+1is minusme
may be read as followsIf the number of cars sold in the current year is more or
less small and the half-year sales increment is quite roughlysmall then the upcoming half-year increment will be negativemedium
421 Recognition of ldquoHealthyrdquo and ldquoSickrdquo Patterns by LFLFOur method to use linguistic fuzzy logic forecasting isbased on simple idea that best predictor learning from bothldquohealthyrdquo and ldquosickrdquo pattern samples respectively can be usedfor validation with tested pattern taken as validation part ofthe seriesThen we can evaluate SMAPE for both these casescompound series SMAPE (ldquohealthyrdquo + tested) and SMAPE(ldquosickrdquo + tested)
If SMAPE (ldquohealthyrdquo + tested) lt SMAPE (ldquosickrdquo + tested)then the tested pattern is supposed to be ldquohealthyrdquo otherwisethe tested pattern is supposed to be ldquosickrdquo
The idea is schematically shown in Figure 14For testing purposes we created two necessary typical
learning time series ldquohealthyrdquo (HS) and ldquosickrdquo (SS) accordingto the algorithm above They both consist of 1010 samplesmade from 10 typical series of ldquohealthyrdquo and ldquosickrdquo patientswith 101 measured ECG values Then we have created 10concatenated series according to the scheme in Figure 14with 10 randomly selected patients with ldquohealthyrdquo ECGmeasurement that is 20 files were produced (10x HS + TSand 10x SS + TS) The same concatenated series were alsomade from 10 ldquosickrdquo patients measurements This made usadditional 20 files with concatenated series (10x HS + TS and
0102030405060708090
100
Sick Healthy
FalseTrue
()
Figure 15 Experimental results LFLF
10x SS + TS) For 20 patients (Table 2) tested ECG we have2 concatenated series giving SMAPE (HS + TS) and SMAPE(SS + TS)
Our method based on LFLF proved very good results forright identification of sick patient records Nevertheless itproduces large amount of false positive identification of sickpattern for healthy patients (Figure 15) This result is con-sistent with our approach using neural networks Of courseour preliminary research has a limited extent and should beperceived only as narrative result which shows interestingproperties especially in complementation of neural networkresults
10 The Scientific World Journal
5 Conclusion
In this paper a short introduction into the field of ECGwaves recognition using backpropagation neural network hasbeen given Main objective was to recognise the normalcycles and arrhythmias and perform further diagnosis Weproposed two detection systems that have been created withusage of neural networks One of them is adapted accordingto the training set Here each pattern represents the wholeone ECG cycle Then an output unit represents a diagnose01 a healthysick person The second one approach usesneural network in which training set contains two differentgroups of patterns for healthysick persons According to theresults of experimental studies it can be stated that ECGwaves patterns were successfully extracted in given timeseries and recognised using suggestedmethod as can be seenfrom figures in Experimental Result section It might resultin better mapping of the time series behaviour for betterprediction
Both approaches were able to predict with high probabil-ity if the ECG time series represents sick or healthy persons Itis interesting that a sick diagnose was recognised with higheraccuracy in both experimental works
The third approach based on LFLF is currently only inthe stage of preliminary experiments but it conforms to theformer results based on neural networks This approach isnovel and could be good supplement to other soft-computingmethods for this task
Disclaimer
Anyopinions findings and conclusions or recommendationsexpressed in this material are those of the authors and do notnecessarily reflect the views of the sponsors
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
The research described here has been financially sup-ported by University of Ostrava Grants SGS16PrF2014 andSGS14PrF2014
References
[1] P Tangkraingkij C Lursinsap S Sanguansintukul and TDesudchit ldquoSelecting relevant EEG signal locations for personalidentification problem using ICA and neural networkrdquo inProceedings of the 8th IEEEACIS International Conference onComputer and Information Science (ICIS rsquo09) pp 616ndash621 June2009
[2] C D Nugent J A C Webb N D Black and G T H WrightldquoElectrocardiogram 2 classificationrdquo Automedica vol 17 pp281ndash306 1999
[3] S Russell and P Norvig Artificial IntelligencemdashA ModernApproach Prentice Hall 2nd edition 2003
[4] C D Nugent J A Lopez A E Smith and N D BlackldquoReverse engineering of neural network classifiers in medicalapplicationsrdquo in Proceedings of the 7th EFOMP Congress vol 17p 184 Physica Medica 2001
[5] L Fausett Fundamentals of Neural Network PrenticeHall 1994[6] D E Rumelhart G E Hinton and R J Williams ldquoLearning
[7] V Novak Fuzzy Sets and Their Applications Adam HilgerBristol UK 1989
[8] S Aouaouda M Chadli P Shi and H R Karimi ldquoDiscrete-time119867
minus119867infinsensor fault detection observer design for nonlin-
ear systems with parameter uncertaintyrdquo International Journalof Robust and Nonlinear Control 2013
[9] M Chadli A Abdo and S X Ding ldquo119867minus119867infin
fault detectionfilter design for discrete-time Takagi-Sugeno fuzzy systemrdquoAutomatica vol 49 no 7 pp 1996ndash2005 2013
[10] M Chadli and H R Karimi ldquoRobust observer design forunknown inputs takagi-sugeno modelsrdquo IEEE Transactions onFuzzy Systems vol 21 no 1 pp 158ndash164 2013
[11] S Aouaouda M Chadli and H Karimi ldquoRobust static output-feedback controller design against sensor failure for vehicledynamicsrdquo IET Control Theory amp Applications vol 8 no 9 pp728ndash737 2014
[12] I Perfilieva V Novak V Pavliska A Dvorak andM StepnickaldquoAnalysis and prediction of time series using fuzzy transformrdquoin Proceedings of the International Joint Conference on NeuralNetworks (IJCNN rsquo08) pp 3876ndash3880 Hong Kong June 2008
[13] E Volna M Kotyrba and R Jarusek ldquoMulti-classifier basedon Elliott waversquos recognitionrdquo Computers amp Mathematics withApplications vol 66 no 2 pp 213ndash225 2013
[14] E Volna M Kotyrba and R Jarusek ldquoPrediction by means ofElliott waves recognitionrdquo in Nostradamus Modern Methods ofPrediction Modeling and Analysis of Nonlinear Systems vol 192of Advances in Intelligent Systems and Computing pp 241ndash250Springer Berlin Germany 2013
[15] H Habiballa V Pavliska and A Dvorak ldquoSoftware systemfor time series prediction based on F-transform and linguisticrulesrdquo in Proceedings of the 8th International Conference onApplied Mathematics Aplimat pp 381ndash386 Bratislava Slovakia2009
In this paper a short introduction into the field of ECGwaves recognition using backpropagation neural network hasbeen given Main objective was to recognise the normalcycles and arrhythmias and perform further diagnosis Weproposed two detection systems that have been created withusage of neural networks One of them is adapted accordingto the training set Here each pattern represents the wholeone ECG cycle Then an output unit represents a diagnose01 a healthysick person The second one approach usesneural network in which training set contains two differentgroups of patterns for healthysick persons According to theresults of experimental studies it can be stated that ECGwaves patterns were successfully extracted in given timeseries and recognised using suggestedmethod as can be seenfrom figures in Experimental Result section It might resultin better mapping of the time series behaviour for betterprediction
Both approaches were able to predict with high probabil-ity if the ECG time series represents sick or healthy persons Itis interesting that a sick diagnose was recognised with higheraccuracy in both experimental works
The third approach based on LFLF is currently only inthe stage of preliminary experiments but it conforms to theformer results based on neural networks This approach isnovel and could be good supplement to other soft-computingmethods for this task
Disclaimer
Anyopinions findings and conclusions or recommendationsexpressed in this material are those of the authors and do notnecessarily reflect the views of the sponsors
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
The research described here has been financially sup-ported by University of Ostrava Grants SGS16PrF2014 andSGS14PrF2014
References
[1] P Tangkraingkij C Lursinsap S Sanguansintukul and TDesudchit ldquoSelecting relevant EEG signal locations for personalidentification problem using ICA and neural networkrdquo inProceedings of the 8th IEEEACIS International Conference onComputer and Information Science (ICIS rsquo09) pp 616ndash621 June2009
[2] C D Nugent J A C Webb N D Black and G T H WrightldquoElectrocardiogram 2 classificationrdquo Automedica vol 17 pp281ndash306 1999
[3] S Russell and P Norvig Artificial IntelligencemdashA ModernApproach Prentice Hall 2nd edition 2003
[4] C D Nugent J A Lopez A E Smith and N D BlackldquoReverse engineering of neural network classifiers in medicalapplicationsrdquo in Proceedings of the 7th EFOMP Congress vol 17p 184 Physica Medica 2001
[5] L Fausett Fundamentals of Neural Network PrenticeHall 1994[6] D E Rumelhart G E Hinton and R J Williams ldquoLearning
[7] V Novak Fuzzy Sets and Their Applications Adam HilgerBristol UK 1989
[8] S Aouaouda M Chadli P Shi and H R Karimi ldquoDiscrete-time119867
minus119867infinsensor fault detection observer design for nonlin-
ear systems with parameter uncertaintyrdquo International Journalof Robust and Nonlinear Control 2013
[9] M Chadli A Abdo and S X Ding ldquo119867minus119867infin
fault detectionfilter design for discrete-time Takagi-Sugeno fuzzy systemrdquoAutomatica vol 49 no 7 pp 1996ndash2005 2013
[10] M Chadli and H R Karimi ldquoRobust observer design forunknown inputs takagi-sugeno modelsrdquo IEEE Transactions onFuzzy Systems vol 21 no 1 pp 158ndash164 2013
[11] S Aouaouda M Chadli and H Karimi ldquoRobust static output-feedback controller design against sensor failure for vehicledynamicsrdquo IET Control Theory amp Applications vol 8 no 9 pp728ndash737 2014
[12] I Perfilieva V Novak V Pavliska A Dvorak andM StepnickaldquoAnalysis and prediction of time series using fuzzy transformrdquoin Proceedings of the International Joint Conference on NeuralNetworks (IJCNN rsquo08) pp 3876ndash3880 Hong Kong June 2008
[13] E Volna M Kotyrba and R Jarusek ldquoMulti-classifier basedon Elliott waversquos recognitionrdquo Computers amp Mathematics withApplications vol 66 no 2 pp 213ndash225 2013
[14] E Volna M Kotyrba and R Jarusek ldquoPrediction by means ofElliott waves recognitionrdquo in Nostradamus Modern Methods ofPrediction Modeling and Analysis of Nonlinear Systems vol 192of Advances in Intelligent Systems and Computing pp 241ndash250Springer Berlin Germany 2013
[15] H Habiballa V Pavliska and A Dvorak ldquoSoftware systemfor time series prediction based on F-transform and linguisticrulesrdquo in Proceedings of the 8th International Conference onApplied Mathematics Aplimat pp 381ndash386 Bratislava Slovakia2009