An Automatic Epilepsy Detection Method Based on Improved...

Research ArticleAn Automatic Epilepsy Detection Method Based on ImprovedInductive Transfer Learning

Yufeng Yao 12 and Zhiming Cui 3

1The Institute of Intelligent Information Processing and Application Soochow University Suzhou 215006 China2Department of Computer Science and Engineering Changshu Institute of Technology Changshu 215500 China3Suzhou University of Science and Technology Suzhou 215009 China

Correspondence should be addressed to Yufeng Yao yyfcslgeducn and Zhiming Cui zmcuiustseducn

Received 20 June 2020 Accepted 15 July 2020 Published 3 August 2020

Guest Editor Yi-Zhang Jiang

Copyright copy 2020 Yufeng Yao and Zhiming Cui This is an open access article distributed under the Creative Commons AttributionLicense which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

Epilepsy is a chronic disease caused by sudden abnormal discharge of brain neurons causing transient brain dysfunction Theseizures of epilepsy have the characteristics of being sudden and repetitive which has seriously endangered patientsrsquo healthcognition etc In the current condition EEG plays a vital role in the diagnosis judgment and qualitative location of epilepsyamong the clinical diagnosis of various epileptic seizures and is an indispensable means of detection The study of the EEGsignals of patients with epilepsy can provide a strong basis and useful information for in-depth understanding of itspathogenesis Although intelligent classification technologies based on machine learning have been widely used to theclassification of epilepsy EEG signals and show the effectiveness In fact it is difficult to ensure that there is always enough EEGdata available for training the model in real life which will affect the performance of the algorithms In view of this to reducethe impact of insufficient data on the detection performance of the algorithms a novel discriminate least squares regression-(DLSR-) based inductive transfer learning method was introduced which is on the basis of DLSR and the inductive transferlearning And it is applied to promote the adaptability and accuracy of the epilepsy EEG signal recognition The proposedmethod inherits the advantages of DLSR it can be more suitable for classification scenarios by expanding the interval betweendifferent classes Meanwhile it can simultaneously use the data of the target domain and the knowledge of the source domainwhich is helpful for getting better performance The results show that the improved method has more advantages in EEG signalrecognition comparing to several other representative methods

1 Introduction

At present epilepsy has become a common disease in neurol-ogy Its pathogenesis has not yet been fully elucidated and itis usually defined as a chronic neurological disease caused bysudden abnormal discharge of brain neurons The epilepticseizures are sudden and repetitive Its onset is accompaniedby clinical manifestations such as loss of consciousness faint-ing and twitching of extremities It also has cognitive andmental disorders that seriously endanger patientsrsquo healthcognition etc [1 2] According to statistics more than onepercent of the worldrsquos population suffer from the disease[3] and there are approximately 9 million people with epi-

lepsy in China Therefore the depth research and preventionof epilepsy play an indispensable role to alleviate the sufferingof patients improve the quality of life and promote healthydevelopment As an important method for studying epilepsyEEG uses electrodes to record the electrical activity of nervecells in the brain which contains a large amount of physio-logical and pathological information and is of great impor-tance in the clinical examination location and therapy ofepilepsy Therefore for people with a tendency to epilepsyautomatic detection of epilepsy can analyze and screen theEEG signals of people at high risk for epilepsy so as to realizeearly detection perform timely intervention and reduce theimpact of epilepsy on people and the incidence of epilepsy

HindawiComputational and Mathematical Methods in MedicineVolume 2020 Article ID 5046315 13 pageshttpsdoiorg10115520205046315

In view of this it is of great value to study the epilepsyautomatic detection algorithm based on EEG signals anddevelop an efficient and accurate epilepsy automatic detec-tion system

In fact the study of automatic epileptic detection basedon EEG signals has attracted extensive attention fromscholars and experts at home and abroad since the 1970sTo predict the onset or preonset of epilepsy in the processof seizure detection machine learning and pattern classifica-tion algorithms are generally applied to classify the EEG sig-nals after extracting the characteristics of the time domainfrequency domain time frequency domain or nonlineardomain of the EEGWith the development of computer tech-nology and digital signal processing technology more andmore methods are widely used in the study of seizure detec-tion methods and have achieved certain research results suchas the Bayesian classifier [4] artificial neural network [5ndash9]support vector machine (SVM) [10ndash13] and fuzzy reasoning[14 15] For example Obeyli extracted the Lyapunov expo-nential features of EEG signals and used probabilistic neu-ral networks to classify EEG signals so as to achieve highclassification results [9] Chan et al extracted the time-frequency features of five subbands in the wavelet trans-form domain of epilepsy EEG signals and then used sup-port vector machines and cluster regression models torecognize the onset of seizures Aarabi et al [14] extractedthe features such as sample entropy dominant frequencyaverage amplitude and amplitude variation coefficient ofintracranial EEG da4ta of patients with epilepsy and usedthe established fuzzy inference rules to fuse the EEG featureinformation for seizure detection Although many of theabove intelligent classification methods have shown theeffectiveness of epilepsy EEG signal classification they stillface a challenge that is it is very hard to get enough EEGdata for epilepsy to train the model in real life Thereforeit has important practical value to explore how to usethe knowledge acquired from related fields to enhancethe classification performance of EEG data in the currentscenario [16]

To solve the above challenges a novel inductive transferlearning method based on discriminant least squares regres-sion (TDLSR) was proposed Meanwhile it was applied tospecific medical application scenarios namely epilepsyEEG signal classification so as to relieve the effect of severedata shortage on the performance of the algorithms Transferlearning is an effective way to transfer knowledge fromrelated fields and is helpful of obtaining more informationin the absence of sufficient data or information It focuseson how to use the useful information from similar but differ-ent source domains to improve the classification result of theclassifier in the target domain When studying epilepsy EEGsignal classification the inductive transfer learning methodnaturally becomes the first choice because of insufficientlabeled epilepsy EEG samples in the target domain some-times What is more since the discriminant LSR is still basedon least squares regression (LSR) [17] which can explain theimportance of each feature in the prediction model based onthe original data space we introduced the inductive transferlearning method based on DLSR to use for epilepsy EEG sig-

nal classification In summary the innovations of this workare summarized as follows

Point 1 The improved method is on the basis of theinductive transfer learning but it has some difference fromthe traditional inductive transfer learning The latter directlytransfers the samples or features used in the source domain tothe target domain for transfer learning while the former usesa knowledge lever mechanism that transfers some knowledgefrom the source domain to the target domain Then the secu-rity of the data in the source domain can be well protectedand the data in the target domain and the knowledge in thesource domain can be used simultaneously so that the classi-fication effect is better

Point 2 The improved method expands DLSR that can bewell applied to classification scenes into a novel method withcertain transfer learning ability so that it can be used in morecomplex scenes

Point 3 The improved method inherits the characteris-tics of DLSR that is it can be better applied to classificationscenarios by expanding the interval between different catego-ries And it can transfer knowledge from the source domainthus ensuring the rationality of its training model

Finally to better illustrate the basic idea of this study thestructure of the paper is as follows

Part 1 Introduction to the research background statusand significance of the thesis

Part 2 The related work including the related technologyof epilepsy EEG signal detection and the epilepsy EEG signalclassification based on transfer learning is summarized inadvance so that the following sections becomemore readable

Part 3 The notations of the inductive transfer learningalgorithm based on DLSR was introduced in detail

Part 4 The reliability and validity of TDLSR algorithm indetection of epilepsy EEG signals based on a series of exper-imental were verified

2 Related Work

Automatic epilepsy detection is based on signal processingtechnology and pattern recognition It analyzes EEG data toidentify the location and duration of seizures Usually theEEG signals collected during seizures was called seizureEEG and the EEG signals collected in the nonseizure arecalled nonseizure EEG The problem of automatic detectionof epilepsy EEG is to effectively judge the above two typesof EEG signals and identify seizures The related detectiontechnologies are introduced as follows

21 The Related Technology of Epilepsy EEG Signal DetectionBecause the EEG signal of epilepsy is easily interfered bymany factors it is very random and it is a nonstationarysignal and its rule is generally difficult to grasp There-fore researchers often use quantitative analysis to extractcharacteristic information of epilepsy EEG signals Theexisting methods of automatic epilepsy detection includethe following

(1) Time-domain analysis time-domain analysis is oneof the earliest methods used in signal analysis It

2 Computational and Mathematical Methods in Medicine

analyzes the time-domain waveforms of EEG signalsto study the difference between EEG waveform dur-ing seizures and EEG waveform during nonseizuresand directly extracts the waveform characteristics ofthe signals to distinguish the two types of EEGsignals The time-domain analysis method has thecharacteristics of intuitiveness and clear physicalmeaning and it can reflect the important informa-tion in transient waveform such as spike wave andharp wave The representative methods of time-domain analysis have a template matching methodtime-domain waveform or energy characteristicsetc In 2001 Litt et al [18] performed feature extrac-tion on EEG signals based on time-domain analysismethods In addition researchers performed time-domain analysis of EEG signals to extract waveformsor energy features different from background activi-ties and use amplitude rhythm period and otherparameters as classification criteria to identifyepilepsy EEG signals Gotman et al performed aldquohalf-waverdquo decomposition of EEG signals and thenextracted EEG features including average amplitudeduration and coefficient of variation relative to thebackground and set thresholds based on expert expe-rience The characteristic parameters are comparedwith the threshold to judge whether it is an epilepsysignal [19ndash21] Time-domain analysis only performsstatistical analysis from the time domain and it iseasy to miss other important changes in abnormalsignals such as slow waves

(2) Frequency analysis unlike time-domain analysiswhich mainly analyzes the waveform characteristicsof epilepsy EEG frequency domain analysis analyzesthe frequency characteristics of EEG It recognizesdifferent rhythms according to the frequency of brainwaves Each brain wave of different rhythms corre-sponds to epilepsy EEG signals in different timeperiods or different parts of the brain [22] Frequencydomain analysis is based on the Fourier transformand is mainly used for power spectrum analysis ofEEG signals It performs the Fourier transform onthe EEG signal to obtain its frequency componentsand spectrum distribution and extracts the corre-sponding EEG features in the frequency domain forepilepsy detection and recognition Representativemethods include power spectrum estimation autore-gressive (AR) model spectrum estimation and higherorder spectrum [23] Among them the power spec-trum estimation transforms the EEG signal whoseamplitude changes with time into the EEG spectrumchart with power varying with frequency and ana-lyzes the distribution and change of each frequencyband of the EEG signal intuitively and quantitatively[24 25] Although frequency domain analysis canprovide a lot of effective information allowingresearchers to detect epilepsy based on the frequencydomain characteristics of EEG the overall spectrumof the signal obtained by the Fourier transform nei-

ther can reflect the local characteristics of the signalnor can reflect the signal frequency componentchanges with time Therefore the detection resultsobtained by frequency domain analysis are not verysatisfactory and are greatly restricted in practicalapplications

(3) Time-frequency analysis the epilepsy EEG is a typi-cal nonstationary signal which contains not onlythe waveform parameter characteristics in the timedomain but also the energy distribution characteris-tics in the frequency domain However neither theabove two methods can fully extract the transientcharacteristics and information of the EEG signalsand can get the ideal results With the developmentof digital signal theory and methods the time-frequency analysis method combining time domainand frequency domain is widely used in the analysisof nonstationary EEG signals It can obtain timeand frequency domain information at the same timeand capture transient information of EEG In recentyears more and more studies have adopted time-frequency analysis methods to analyze EEG signalsamong which various wavelet change methods arerepresented The wavelet transform uses the transla-tion and expansion of the window function to imple-ment a wide time window for the low-frequencycomponents of the signal and a narrow time windowfor the high-frequency components to complete themultiscale analysis of the signal This analysismethod conforms to the laws of nature and has agood ability to characterize the local characteristicsof the signal [26] It can capture the transient charac-teristics of the EEG signal and accurately locate it inthe time and frequency domains In addition towavelet transform commonly used time-frequencyanalysis methods also include empirical modedecomposition [27ndash30] the Wigner-Ville distribu-tion [31 32] and the Stockwell transform [33 34]However most of these time-frequency analysismethods can only be used for multiresolution analy-sis of the original signal and then need to be com-bined with other algorithms to achieve the featureextraction and selection of EEG Figure 1 shows thecomparison of time-domain analysis frequencydomain analysis and wavelet transform analysis

Observe the above figures it is easy to draw aconclusion that the time-frequency analysis can pro-vide more useful information compared to time-domain analysis and frequency domain analysis InFigure 1(c) the wavelet transform improves the timeresolution at the high frequency of the signal bychanging the time window and improves the fre-quency resolution at the low frequency which has abetter classification effect

(4) Nonlinear dynamic analysis nowadays with theprogress of nonlinear dynamics theory researchersare devoted to studying the nonlinear of EEG signals

3Computational and Mathematical Methods in Medicine

to solve the problem of automatic detection ofepilepsy Using nonlinear dynamics theory methodsfor EEG signal analysis various nonlinear featuresof EEG signals can be extracted to distinguish epi-lepsy EEG signals from normal EEG signals Thisprovides some new research ideas for automaticepilepsy detection technology Kannathal et al useddifferent entropies to measure the chaotic character-istics of EEG signals and used them as EEG featuresto distinguish EEG signals in different periods [35]including the Shannon entropy Renyi entropyKolmogorov-Sinai entropy and approximate entropyThe results show that the complexity of the EEG signalin patients with epilepsy during the intermittentperiod is higher than that during the seizure periodthat is the complexity of the EEG signal during theseizure is reduced and the bet value is less than thenormal EEG signal Although nonlinear analysis canreflect the dynamic mechanism of seizures well mostof the nonlinear features are computationally intensiveand generally time-consuming which is not suitablefor real-time epilepsy automatic detection systems

22 The Epilepsy EEG Signal Classification Based on TransferLearning Traditional classification methods use a largeamount of data with label information to train a decisionfunction and then use this function to classify and identifytest samples with unknown label information Howeverthese classification methods all have a presupposition train-ing data and test data need to obey the same distributioncharacteristics as shown in Figure 2 For the differences in

the distribution of training samples and test samples asdescribed above the performance of the traditional methodssignificantly decrease as shown in Figure 3 In response tothis challenge transfer learning is a promising researchdirection Transfer learning focuses on knowledge transferproblems that are similar to different domains or havedifferent data distributions It enhances the performanceof the classifier used for target area recognition by learn-ing useful knowledge from the source domain Accordingto whether the target domain used contains samples withlabeled information transfer learning techniques are dividedinto three categories inductive transfer learning methoddirect transductive transfer learning method and unsuper-vised transfer learning method [36] In the paper we willfocus on an inductive transfer learning method with goodperformance that is inductive transfer learning method

Time

Amplitude

(a)

Time

Frequency

(b)

Time

Frequency

(c)

Figure 1 Comparison of time-domain analysis frequency domain analysis and time-frequency analysis (a) Time-domain analysis(b) Frequency analysis (Fourier transform) (c) Wavelet transform

Same distribution

Trainingdata

Testingdata

Figure 2 Scene suitable for the traditional method


based on discriminant least squares regression (TDLSR)And its application and actual effect in EEG signal detectionof epilepsy will be studied The framework structure of epi-lepsy EEG signal detection based on transfer learning theoryis given as shown in Figure 4

In short transfer learning is to transfer the knowledge(useful knowledge) from the source domain with a largeamount of labeled data for learning in the target domain withno or little labeled data thereby improving the trainingquality of the target domain This can reduce the workloadof collecting labeled data in the target domain

3 The Inductive Transfer Learning AlgorithmBased on DLSR

To better describe the algorithm proposed in this paperTable 1 gives a detailed description of the symbols in thealgorithm

Since DLSR is a nontransforming algorithm the symbolsin Table 1 refer to the parameter variables of the originaltraining sample

31 The Least Squares Regression As a widely used methodbased on statistical theory LSR has become a typicalmethod LSR uses the Frobenius norm to constrain thematrix of representation coefficients In the paper toexpand the classification ability of the LSR algorithm wepreconstructed the binary label matrix Y correspondingto the training samples X so that it can better cope withmore complex classification scenarios The jth column ofY indicates that only the data belonging to the jth classcorresponds to an element equal to 1 and all other ele-ments are 0 Then the objective function of LSR can beredefined as follows

arg minZ

XZ minus Yk k2F + λ Zk k2F eth1THORN

Since the LSR has an analytical solution which can beeasily obtained formula (1) can be rewritten as

J Zeth THORN = arg minZ


Trainingdata

Testingdata

Different distribution

Figure 3 New challenge for the traditional methods

EEG signal

Signal data

Data in the samedistribution

Data in the differentdistribution

Feature extraction

Transfer learning

Sourcedomain

Useful knowledge

Identification results

Targetdomain

Figure 4 Framework of adaptive recognition for epileptic EEGsignals based on transfer learning

Table 1 Symbol description

Symbol Description

XTraining matrix of the target domain X = x1 x2⋯xNfrac12

and xi isin Rd

N Total number of training samples of the target domain

d Dimension of sample

C The number of classes

Y

The binary label matrix corresponding to the trainingsamples in the target domain The jth column of Yindicates that only the data belonging to the jth classcorresponds to an element equal to 1 and all other

elements are 0

B Indicator matrix constructed based on Y

W Label offset matrix of the target domain

Z Mapping matrix of target domain

ZS Mapping matrix of source domain

e Vector representing all 1 e isin RN

Θ Hadamard product operator for matrix

p The mapping offset vector of target domain

λ η The regularization parameter


Then the solution process of the analytical solution ofLSR is as follows let

partJ Zeth THORNpartZ

= 0

⟹2XTXZ minus 2XTY + 2λZ = 0

⟹ XTX + λId

Z = XTY

eth3THORN

Finally the analytical solution of LSR can beobtained as

Z = XTX + λId minus1

XTY eth4THORN

32 The Discriminate Least Squares Regression As weknow LSR can be directly used for classification tasksHowever since the interval between any two differentclasses in the constructed binary class label matrix isffiffiffi2

p the DLSR proposed in literature [37] introduces

the relaxation technique into the LSR so as to expandthe interval between the two data from different classesTo improve the compactness of the classification taskDLSR will comprehensively consider the class factorsand build an indicator matrix B on the basis of thebinary label matrix Y of the sample Each element ofmatrix B is defined as follows

Bij=+1 if yi = 1

minus1 otherwise

(

eth5THORN

In essence each element of matrix B represents theoffset direction of the corresponding class label Thenε relaxation on each element of Y is performed andthe amount of ε relaxation through the matrix W isrecorded Then the objective function of DLSR can beexpressed as

minstWge0

XZ + epT minus Y + BΘWeth THORN 2F + λ Zk k2F eth6THORN

The objective function of DLSR is a convex optimiza-tion problem However it cannot directly optimize thesolution Literature [37] adopts an alternative optimizationstrategy and ensures that a closed solution is obtained ateach step The specific derivation process is as follows

(1) Fix W and update Z and p Let L = Y + BΘWformula (6) can be rewritten as

J Z peth THORN = arg minZp

XZ + epT minus L 2

F + λ Zk k2F eth7THORN

According to the optimization theory we take apartial derivative of p namely

partJ Z peth THORNpartp

= 0

⟹ZTXTe + peTe minus LTe = 0

⟹p =LTe minus ZTXTe

N

eth8THORN

Furthermore we find the partial derivative of Z andwe can get


= 0

⟹2XTXZ minus 2XT epT minus L

+ 2λZ = 0

⟹XTXZ minus XT eeTL minus eTXZ

Nminus L

+ λZ = 0

⟹XT IN minuseeT

N

XZ minus XT IN minus

eeT

N

L + λZ = 0

⟹Z = XTHX + λId minus1

XTHL

eth9THORN

(2) Fix Z and p and update W Let G = XZ + epT minus Y then W can be solved as follows

arg minW

stWge0

G minus BΘWk k2F eth10THORN

According to literature [37] the Frobenius norm squareof a matrix can be solved element by element so Eq (10)can be equivalent to solve N times C subproblems For elementWij in row i and column j there can be

minWij

Gij minus BijWij

2

stWijge0

eth11THORN

where Gij and Bij represent the jth element of the ith row ofthe matrix G and matrix B respectively and satisfy Bij

2 = 1Then we can get

Gij minus BijWij

2 = BijGij minusWij

2 eth12THORN

And because each element of W satisfies Wij ge 0formula (12) can be written as

Wij =max BijGij 0

eth13THORN


Therefore the final solution formula of W is

W =max BΘG 0eth THORN eth14THORN

According to the above derivation Algorithm 1 gives adetailed description of the DLSR algorithm

33 The Inductive Transfer Learning Algorithm Based onDLSR Most inductive transfer learning algorithms areimplemented by directly learning from the data in the sourcedomain through some classes However in the paper we

used a knowledge-based inductive transfer learning frame-work instead of raw data to study inductive transfer learningmethods based on source domain knowledge Inspired bythis an inductive transfer learning algorithm based on DLSRwas introduced Its objective function is

minZp

XZ + epT minus Y + BΘWeth THORN 2F

stWge0

+ λ Zk k2F + η Z minus Zsk k2

eth15THORN

The description of DLSRInputThe training samples X = frac12x1 x2⋯xN isin RNtimesd and their corresponding class labels yi isin f1 2⋯Cgethi = 1 2⋯NTHORNwhere xi isin Rdethi = 1 2⋯NTHORN The maximum number of iterations is T OutputThe mapping matrix Z and the translation vector pTrainingStep 1 Construct the label matrix Y and the indicator matrix B respectivelyStep 2 Get the value of λ by grid searching in the set of 10-4 10-3 10-2 10-1 100 10 102Step 3 Initialize Z prime = 0 p = 0 andW = 0 Set t = 1Step 4 Update Q by Q = ethXTHX + λIdTHORNminus1XTHRepeatStep 5 Update the label matrix L by L = Y + BΘWStep 6 Update p and Z by Eqs (8) and (10)Step 7 Update the regression error matrix G by G = XZ + epT minus Y Step 8 Update the label shift matrix W by Eq (14)Step 9 Let Z prime = Z and t = t + 1Until kZ prime minus Zk2F le 10minus4 or t gt T Step 10 Output Z and p

Algorithm 1 The description of the DLSR algorithm

The description of TDLSRInputThe training samples in target domain X = frac12x1 x2⋯xN isin RNtimesd and their corresponding class labels yi isin f1 2⋯Cgethi = 1 2⋯NTHORNwhere xi isin Rdethi = 1 2⋯NTHORN The maximum number of iterations is T AS is learned in advanced by using DLSR from source domainOutputThe mapping matrix Z and the translation vector pTrainingStep 1 Construct the label matrix Y and the indicator matrix B respectivelyStep 2 Get the value of λ and η by grid searching in the set of 10-4 10-3 10-2 10-1 100 10 102 respectivelyStep 3 Initialize Z prime = 0 p = 0 andW = 0 Set t = 1Step 4 Compute Q by Q = ethXTHX + λIdTHORNminus1XTH and V by V = ηethXTHX + ethλ + ηTHORNIdTHORNminus1ZS respectivelyRepeatStep 5 Update the label matrix L by L = Y + BΘWStep 6 Update p and Z by Eqs (17) (18)Step 7 Update the regression error matrix G by G = XZ + epT minus Y Step 8 Update the label shift matrix W by Eq (20)Step 9 Let Z prime = Z and t = t + 1Until kZ prime minus Zk2F le 10minus4 or t gt T Step 10 Output Z and p

Algorithm 2 The description of the TDLSR algorithm


Y

Initializationparameters

Construct the label matrix Yand the indicator matrix B

Update the label matrix L

Update the mapping matrixof the target domain Z andthe mapping offset vector of

target domain p

Update the regression errormatrix G

Update the label shift matrix W

Output Z and p

N

If 997919ZprimeminusZ9979192le 10minus4 ort gt T

F

Figure 5 The process of the TDLSR algorithm

It can be found from the formula (15) that the first twoitems directly inherit the DLSR for learning in the targetdomain The third item is used to transfer the knowledgeZs of the source domain to the target domain When η = 0DSLR is DSLR

In short TDLSR summarizes DLSR from the perspectiveof transfer learning but it has more transfer learning capabil-ities than DLSR and has better applicability Similar to DLSRthe objective function of TDLSR can also be solved usingan alternate optimization strategy The specific derivationprocess is as follows



XZ + epT minus L 2

F + λ Zk k2F+ η Z minus ZSk k2F

eth16THORN



= 0



N

eth17THORN

Furthermore we find the partial derivative of Z wecan get


= 0


+ 2λZ + 2η Z minus ZSeth THORN = 0


Nminus L

+ λZ

+ η Z minus ZSeth THORN = 0

⟹XT IN minuseeT

N


eeT

N

L

+ λ + ηeth THORNZ minus ηZS = 0

⟹Z = XTHX + λ + ηeth THORNId minus1

XTHL + ηZS

eth18THORN

(2) Fix Z and p and update W Let G = XZ + epT minus Y then W can be solved as

arg minW

stWge0


Similar to the way of solving W in DLSR the finalsolution formula of W is


According to the above derivation process Algorithm 2gives a detailed description of the TDLSR algorithm

To understand the TDLSR algorithm more clearlyFigure 5 shows the specific process of the TDLSR

4 Experimental Research

41 The Environment and Parameter Settings To evaluatethe performance of the TDLSR a lot of experiments werecarried out based on EEG signal recognition Experimentswill be conducted from two aspects (1) the comparisonexperiments with classic LSR SRC RLR DLSR and SVMand (2) the comparison experiments with related methodswith transform ability such as AuSVM Tr-Adaboost


TSVM and LMPROJ The hardware environment used inthe experiment is Intel (R) Core(TM) I7-9700 30GHztimes28G RAM and the software environment is Windows 1064 bit MATLAB R2012b Table 2 shows the parametersettings of the above algorithms

42 Experimental Dataset In the experiments the EEGdataset used for epilepsy is the Bonn dataset [38 39] whichwas collected by Andrzejak et al at an epilepsy center at theUniversity of Bonn The EEG dataset contains five datasetsdenoted by A to E This dataset compares the EEG of thepatient during the onset and nononset period with the EEGof the normal person Dataset A and dataset B are EEG sig-nals collected by healthy testers with their eyes open andclosed Dataset C and dataset D are the EEG signals collectedby epilepsy patients outside and inside the lesion during theseizure period and dataset E is the EEG signals collected bythe patients in dataset C and dataset D during the seizureEach of the 5 datasets includes 100 single-channel EEGs (thatis 100 samples) the sampling frequency is 17361Hz eachsegment of the signal collects 4097 frequency points andeach EEG segment lasts 236 s

In our experiments the use of the Bonn dataset is signif-icantly different from many previous works We constructed8 subdatasets from the original 5 datasets to simulatedifferent scenarios in the experiments The source and target

domains of the 8 subdatasets of experiments are composed ofthe partial data extracted from the 5 sets We randomly select75 of the data from a certain dataset as the source domainand the remaining 25 as the target domain The sampledata of source and target domains of subdatasets 1-4 arederived from the same distribution but the samples takenare different The sample data of the source domain and thetarget domain of subdatasets 5-8 have different distributionsFinally for the data in the target domain 20 is randomlyselected for testing and the remaining 80 is used for

Table 2 The parameter settings in the experiments

Methods Parameter settings

LSR The regularization parameter λ isin 10minus4 10minus3 10minus2 10minus1 100 10 102

SRC The regularization parameter α isin 10‐3 5 times 10‐3 10‐2 2 times 10‐2 5 times 10‐2 10‐1

RLR The regularization parameter λ isin 10minus4 10minus3 10minus2 10minus1 100

the heart parameter δ isin 2‐5 2‐4 2‐3⋯23 24 25

DLSR The regularization parameter λ isin 10minus4 10minus3 10minus2 10minus1 100 10 102

SVM The penalty factor c isin 2‐5 2‐4 2‐3⋯23 24 25

the kernel parameter δ isin 2‐5 2‐4 2‐3⋯23 24 25

TSVM The Lagrange multiplier upper bound c isin 2‐5 2‐4 2‐3⋯23 24 25


Au-SVM The penalty factor c isin 2‐5 2‐4 2‐3⋯23 24 25


Tr-Adaboost The penalty factor c isin 2‐5 2‐4 2‐3⋯23 24 25


LMPROJ The regularization parameters τ isin 025 01⋯200f g σ isin 025 01⋯200f g the kernel parameter δ isin 025 01⋯200f gTDLSR The regularization parameter λ isin 10minus4 10minus3 10minus2 10minus1 100 10 102

η isin 10minus4 10minus3 10minus2 10minus1 100 10 102

Table 3 The construction of experimental datasets

The subdataset Source domain Target domain

D1 AE-each 75 AE-each 25

D2 BDE-each 75 BDE-each 25

D3 ABCD-each 75 ABCD-each 25

D4 BCDE-each 75 BCDE-each 25

D5 BE-each 75 BC-each 25

D6 ACE-each 75 BCE-each 25

D7 ADE-each 75 BDE-each 25

D8 ACDE-each 75 BCDE-each 25

Table 4 Experimental results based on nontransfer learningalgorithms

Datasets LSR SRC RLR DLSR SVM

D1 087 084 089 090 082

D2 079 072 081 080 074

D3 071 062 072 072 068

D4 071 064 070 072 066

D5 082 078 083 084 078

D6 073 060 072 074 061

D7 070 060 070 072 059

D8 063 053 065 064 050

Table 5 Experimental results based on transfer learningalgorithms

Datasets Au-SVM TSVM Tr-Adaboost LMPROJ TDLSR

D1 090 092 089 092 096

D2 082 085 082 084 090

D3 072 075 073 076 083

D4 072 075 071 078 082

D5 085 086 084 087 091

D6 075 078 075 080 083

D7 071 078 074 078 083

D8 068 070 066 070 073


training In addition the training and test datasets do notcontain the same samples and are independent of each otherTable 3 lists the detailed information of the 8 subdatasets

43 Experimental Results and Analysis To verify the effec-tiveness of TDLSR in epilepsy EEG data recognition thecomparative experiments were conducted among the classicalgorithms The experimental results are shown in Tables 4and 5 respectively

As shown in the above experimental results the conclu-sions are summarized as follows

(1) In the case of the same distribution of the sourcedomain dataset and the target domain dataset boththe nontransfer learning algorithms and the trans-fer learning algorithms achieve good classificationresults For the datasets with certain differences indistribution the classification effects of algorithmswithout transforming abilities are quite differentand the experimental results of the algorithms inTable 5 are generally better than results of thealgorithms in Table 4

(2) As a whole the performance of TDLSR introducedin this paper is obviously superior to all other algo-rithms It means that by using the knowledgetransferred from the source domain to the targetdomain the TDLSR algorithm obtains better per-formance and becomes effective for epilepsy EEGrecognition

(3) Comparing the performance of the algorithms inTable 4 it can be found that DLSR has the best per-formance while SRC and SVM have poor perfor-mance This is because DLSR expands the ability todistinguish between classes by using the class infor-mation in the label space for the classification taskAs the results shown in Table 5 the TDLSR algo-rithm performs best This is because it not onlyinherits the advantages of DLSR by increasing theinterval between different classes but also transfersmore useful information from the source domain tothe target domain It has stronger transfer learningability than several other algorithms And becausethe TDLSR method requires fewer parameters toadjust it is easier to use and the stability and faulttolerance are stronger than the transfer learning algo-rithms such as LMPROJ

In addition to further observe and compare the overallclassification performance of all algorithms Figures 6 and 7also give the average classification accuracy of each algorithmon all datasets

Observing Figures 6 and 7 it can be seen that the classi-fication accuracy of the algorithms on the same distributeddataset is higher than the result on the different distributeddataset Secondly compared with the traditional nontransferlearning algorithms the algorithm with transfer learningability has more advantages in classification performanceand the TDLSR has a significant improvement in classifica-tion performance Finally for all datasets with different clas-ses the performance of all algorithms decreases continuously

D1 D2 D3 D4 D5 D6 D7 D8

05

06

07

08

09

Datasets

Acc

urac

y (

)

LSRSRCRLR

DLSRSVM

Figure 6 The average classification accuracy of all nontransfer learning algorithms


as the number of categories increases This is because as thenumber of classes increases the information in the labelspace becomes more complicated and learning the informa-tion in the label space becomes more difficult

Meanwhile to further verify the reliability and stability ofthe algorithms we randomly added 15 white Gaussiannoise to the data in the source domain to prove that thealgorithm in this paper can adapt to more complex scenariosThe experimental results are shown in Tables 6 and 7respectively

Through the above experimental results it can be foundthat the classification performance of all nontransfer learningalgorithms under noisy conditions decreases more The rea-son is that they cannot obtain useful knowledge from noisydata (source domain) for classification However all the algo-rithms in Table 7 can transfer some useful knowledge fromthe source domain for classification in the target domain sotheir performance is better than the results in Table 6

In summary the TDLSR algorithm introduced in thepaper is superior to the other algorithms in the detection ofepilepsy EEG signals And it is easy to learn and train hashigh stability and shows certain advantages compared withother intelligent algorithms

5 Conclusion

To solve the problem of serious shortage of training data inthe current scene and improve the accuracy of classificationa novel DLSR-based inductive transfer learning algorithm(TDLSR) was introduced for the detection of epilepsy EEGsignals It can take advantage of both inductive transfer learn-

ing and DLSR On the one hand it can not only protect thesecurity of the source domain data but also use the data ofthe target domain and the knowledge of the source domain

Table 6 Experimental results based on nontransfer learningalgorithms with 15 noise in the source domain


D1 083 080 084 085 077

D2 074 067 076 076 070

D3 066 059 068 069 065

D4 066 061 067 068 063

D5 078 074 080 080 074

D6 065 056 069 069 058

D7 064 057 067 071 056

D8 060 050 062 061 047

Table 7 Experimental results based on transfer learning algorithmswith 15 noise in the source domain


D1 087 086 089 089 093

D2 079 079 082 081 087

D3 069 070 072 073 080

D4 069 068 072 075 079

D5 082 081 083 084 088

D6 072 072 075 077 080

D7 068 071 075 075 080

D8 065 063 067 067 070

D1 D2 D3 D4 D5 D6 D7 D8

065

07

075

08

085

09

095

1

Datasets

Acc

urac

y (

)

AuminusSVMTrminusAdaboostTSVM

LMPROJTDLSR

Figure 7 The average classification accuracy of all transfer learning algorithms


to get better performance On the other hand it inherits theDLSRrsquos characteristics of being more suitable for real classifi-cation scenarios by expanding the interval between differentcategories Therefore compared with DLSR the new algo-rithm not only enhances the ability of transfer learning butalso ensures that the model is more reasonable The resultsreflect that the improved algorithm has more advantages inepilepsy EEG signal classification compared with the tradi-tional algorithms However it is found that the results iseasily affected by the parameters In a word the quality ofthe parameter selection will directly affect the final detectionaccuracy Therefore to obtain higher detection accuracy it isworth to further study the characteristics of the EEG signal toguide the setting range of the parameters in the transferlearning algorithm

Data Availability

The labeled dataset used to support the findings of this studyare available from the corresponding author upon request

Conflicts of Interest

The authors declare no conflict of interest regarding thepublication

References

[1] K S Wilcox P J West and C S Metcalf ldquoThe currentapproach of the Epilepsy Therapy Screening Program contractsite for identifying improved therapies for the treatment ofpharmacoresistant seizures in epilepsyrdquo Neuropharmacologyvol 166 p 107811 2020

[2] L Harris and H Angus-Leppan ldquoEpilepsy diagnosis classifi-cation and managementrdquo Medicine 2020

[3] T Gandhi B K Panigrahi M Bhatia and S Anand ldquoExpertmodel for detection of epileptic activity in EEG signaturerdquoExpert System with Applications vol 37 no 4 pp 3513ndash3520 2010

[4] R Sahu S R Dash L A Cacha R R Poznanski and S ParidaldquoEpileptic seizure detection a comparative study between deepand traditional machine learning techniquesrdquo Journal of Inte-grative Neuroscience vol 19 no 1 pp 1ndash9 2020

[5] S Panahi T Shirzadian M Jalili and S Jafari ldquoA new chaoticnetwork model for epilepsyrdquo Applied Mathematics and Com-putation vol 346 no 1 pp 395ndash407 2019

[6] S Panahi Z Aram S Jafari J Ma and J C Sprott ldquoModelingof epilepsy based on chaotic artificial neural networkrdquo ChaosSolitons amp Fractals vol 105 pp 150ndash156 2017

[7] S T George M S P Subathra N J Sairamya L Susmithaand M J Premkumar ldquoClassification of epileptic EEG signalsusing PSO based artificial neural network and tunable-Qwavelet transformrdquo Biocybernetics and Biomedical Engineer-ing vol 40 no 2 pp 709ndash728 2020

[8] S P Kumar N Sriraam P G Benakop and B C JinagaldquoEntropies based detection of epileptic seizures with artificialneural network classifiersrdquo Expert Systems with Applicationsvol 37 no 4 pp 3284ndash3291 2010

[9] E D Uneyli ldquoLyapunov exponentsprobabilistic neural net-works for analysis of EEG signalsrdquo Expert Systems with Appli-cations vol 37 no 2 pp 985ndash992 2010

[10] M Sharma S Patel and U R Acharya ldquoAutomated detectionof abnormal EEG signals using localized wavelet filter banksrdquoPattern Recognition Letters vol 133 pp 188ndash194 2020

[11] S Raghu N Sriraam Y Temel S V Rao A S Hegde andP L Kubben ldquoPerformance evaluation of DWT based sigmoidentropy in time and frequency domains for automated detec-tion of epileptic seizures using SVM classifierrdquo Computers inBiology and Medicine vol 110 pp 127ndash143 2019

[12] A Temko E Thomas W Marnane G Lightbody andG Boylan ldquoEEG-based neonatal seizure detection with sup-port vector machinesrdquo Clinical Neurophysiology vol 112no 3 pp 464ndash473 2011

[13] N Mahmoodian A Boese M Friebe and J HaddadnialdquoEpileptic seizure detection using cross-bispectrum of electro-encephalogram signalrdquo Seizure European Journal of Epilepsyvol 66 pp 4ndash11 2019

[14] S Ramakrishnan and A S Muthanantha Murugavel ldquoEpilep-tic seizure detection using fuzzy-rules-based sub-band specificfeatures and layered multi-class SVMrdquo Pattern Analysis andApplications vol 22 no 3 pp 1161ndash1176 2019

[15] H S Alaei M A Khalilzadeh and A Gorji ldquoOptimalselection of SOP and SPH using fuzzy inference system foron-line epileptic seizure prediction based on EEG phase syn-chronizationrdquo Australasian Physical amp Engineering Sciencesin Medicine vol 42 no 4 pp 1049ndash1068 2019

[16] J M Lichtenberg and O Simsek ldquoSimple regression modelsrdquoin Imperfect Decision Makes Admitting Real-Word Rational-ity pp 13ndash25 Barcelona 2017

[17] N Imran R Tongeri and M Bennamoun ldquoLinear regressionfor face classificationrdquo IEEE Transactions on Patter Analysisand Machine Intelligence vol 32 no 11 pp 2106ndash2112 2011

[18] B Litt R Esteller J Echauz et al ldquoEpileptic seizures maybegin hours in advance of clinical onsetrdquo Neuron vol 30no 1 pp 51ndash64 2001

[19] A Sharmila and P Geethanjali ldquoEffect of filtering with timedomain features for the detection of epileptic seizure fromEEG signalsrdquo Journal of medical engineering amp technologyvol 42 no 3 pp 217ndash227 2018

[20] J Gotman ldquoAutomatic seizure detection improvements andevaluationrdquo Electroencephalography and Clinical Neurophysi-ology vol 76 no 4 pp 317ndash324 1990

[21] A Sharmila and P Mahalakshmi ldquoWavelet-based featureextraction for classification of epileptic seizure EEG signalrdquoJournal of Medical Engineering ampamp Technology vol 41no 8 pp 670ndash680 2017

[22] I Goker O Osman S Ozekes M B Baslo M Ertas andY Ulgen ldquoClassification of juvenile myoclonic epilepsy dataacquired through scanning electromyography with machinelearning algorithmsrdquo Journal of Medical Systems vol 36no 5 pp 2705ndash2711 2012

[23] O Faust U R Acharya L C Min and B H Sputh ldquoAuto-matic identification of epileptic and background EEG signalsusing frequency domain parametersrdquo International journalof neural systems vol 20 no 2 pp 159ndash176 2012

[24] R Hopfengartner F Kerling V Bauer and H Stefan ldquoAnefficient robust and fast method for the offline detection ofepileptic seizures in long-term scalp EEG recordingsrdquo Clin-ical Neurophysiology vol 118 no 11 pp 2332ndash2343 2007

[25] A Maheshwari ldquoRodent EEG expanding the spectrum ofanalysisrdquo Epilepsy Currents vol 20 no 3 pp 149ndash1532020


[26] S Ashok and G Purusothaman ldquoA review on the patterndetection methods for epilepsy seizure detection from EEGsignalsrdquo Biomedical EngineeringBiomedizinische Technikvol 64 no 5 pp 507ndash517 2019

[27] R B Pachori ldquoAnalysis of normal and epileptic seizure EEGsignals using empirical mode decompositionrdquo ComputerMethods and Programs in Biomedicine vol 104 no 3pp 373ndash381 2011

[28] R J Martis U R Acharya H J Tan et al ldquoApplication ofempirical mode decomposition (EMD) for automated detec-tion of epilepsy using EEG signalsrdquo International Journal ofNeural Systems vol 22 no 6 2012

[29] V Bajaj and R B Pachri ldquoEpileptic seizure detection based onthe instantaneous area of analytic intrinsic mode functions ofEEG signalsrdquo Biomedical Engineering Letters vol 3 no 1pp 17ndash21 2013

[30] R B Pachori R Sharma and S Patidar ldquoClassification ofnormal and epileptic seizure EEG signals based on empiricalmode decompositionrdquo in Complex system modelling andcontrol through intelligent soft computations pp 367ndash388Springer Cham 2015

[31] C Guerrero-Mosquera A M Trigueros J I Franco andAacute Navia-Vaacutezquez ldquoNew feature extraction approach forepileptic EEG signal detection using time-frequency distri-butionsrdquo Medical ampBiological Engineering amp Computingvol 48 no 4 pp 321ndash330 2010

[32] E Ebrahimzadeh S M Alavi A Bijar and A Pakkhesal ldquoAnovel approach for detection of deception using SmoothedPseudo Wigner-Ville Distribution (SPWVD)rdquo Journal of Bio-medical Science and Engineering vol 6 no 1 pp 8ndash18 2013

[33] H Polat M U Aluccedillu and M S Oumlzerdem ldquoEvaluation ofpotential auras in generalized epilepsy from EEG signals usingdeep convolutional neural networks and time-frequency rep-resentationrdquo Biomedical Engineering Biomedizinische Tech-nik 2019

[34] M A Al-Manie and W J Wang ldquoA modified S-transform forEEG signals analysisrdquo Internation Journal of ComputationalMethods vol 12 no 3 2015

[35] N Kannathal M L Choo U R Acharya and P K SadasivanldquoEntropies for detection of epilepsy in EEGrdquo ComputerMethods and Programs in Biomedicine vol 80 no 3pp 187ndash194 2005

[36] A Arnold R Nallapati and W W Cohen ldquoA comparativestudy of methods for transductive transfer learningrdquo in Pro-ceedings of the 7th IEEE International Conference on DataMining Workshops(ICDMWrsquo07) pp 77ndash82 IEEE ComputerSociety Washington DC USA 2007

[37] S Xiang F Nie G Meng C Pan and C Zhang ldquoDiscrimina-tive least squares regression for multiclass classification andfeature selectionrdquo IEEE Transactions on Neural Networksand Learning Systems vol 23 no 11 pp 1738ndash1754 2012

[38] Y Jiang Z Deng F-L Chung et al ldquoRecognition of epilepticEEG signals using a novel multiview TSK fuzzy systemrdquo IEEETransactions on Fuzzy Systems vol 25 no 1 pp 3ndash20 2017

[39] Y Jiang D Wu Z Deng et al ldquoSeizure classification fromEEG signals using transfer learning semi-supervised learningand TSK fuzzy systemrdquo IEEE Transactions on Neural Systemsand Rehabilitation Engineering vol 25 no 12 pp 2270ndash22842017


An Automatic Epilepsy Detection Method Based on Improved Inductive Transfer Learning

1 Introduction

2 Related Work

21 The Related Technology of Epilepsy EEG Signal Detection

22 The Epilepsy EEG Signal Classification Based on Transfer Learning

3 The Inductive Transfer Learning Algorithm Based on DLSR

31 The Least Squares Regression

32 The Discriminate Least Squares Regression



41 The Environment and Parameter Settings

42 Experimental Dataset

43 Experimental Results and Analysis

5 Conclusion

Data Availability


In view of this it is of great value to study the epilepsyautomatic detection algorithm based on EEG signals anddevelop an efficient and accurate epilepsy automatic detec-tion system

In fact the study of automatic epileptic detection basedon EEG signals has attracted extensive attention fromscholars and experts at home and abroad since the 1970sTo predict the onset or preonset of epilepsy in the processof seizure detection machine learning and pattern classifica-tion algorithms are generally applied to classify the EEG sig-nals after extracting the characteristics of the time domainfrequency domain time frequency domain or nonlineardomain of the EEGWith the development of computer tech-nology and digital signal processing technology more andmore methods are widely used in the study of seizure detec-tion methods and have achieved certain research results suchas the Bayesian classifier [4] artificial neural network [5ndash9]support vector machine (SVM) [10ndash13] and fuzzy reasoning[14 15] For example Obeyli extracted the Lyapunov expo-nential features of EEG signals and used probabilistic neu-ral networks to classify EEG signals so as to achieve highclassification results [9] Chan et al extracted the time-frequency features of five subbands in the wavelet trans-form domain of epilepsy EEG signals and then used sup-port vector machines and cluster regression models torecognize the onset of seizures Aarabi et al [14] extractedthe features such as sample entropy dominant frequencyaverage amplitude and amplitude variation coefficient ofintracranial EEG da4ta of patients with epilepsy and usedthe established fuzzy inference rules to fuse the EEG featureinformation for seizure detection Although many of theabove intelligent classification methods have shown theeffectiveness of epilepsy EEG signal classification they stillface a challenge that is it is very hard to get enough EEGdata for epilepsy to train the model in real life Thereforeit has important practical value to explore how to usethe knowledge acquired from related fields to enhancethe classification performance of EEG data in the currentscenario [16]

To solve the above challenges a novel inductive transferlearning method based on discriminant least squares regres-sion (TDLSR) was proposed Meanwhile it was applied tospecific medical application scenarios namely epilepsyEEG signal classification so as to relieve the effect of severedata shortage on the performance of the algorithms Transferlearning is an effective way to transfer knowledge fromrelated fields and is helpful of obtaining more informationin the absence of sufficient data or information It focuseson how to use the useful information from similar but differ-ent source domains to improve the classification result of theclassifier in the target domain When studying epilepsy EEGsignal classification the inductive transfer learning methodnaturally becomes the first choice because of insufficientlabeled epilepsy EEG samples in the target domain some-times What is more since the discriminant LSR is still basedon least squares regression (LSR) [17] which can explain theimportance of each feature in the prediction model based onthe original data space we introduced the inductive transferlearning method based on DLSR to use for epilepsy EEG sig-

nal classification In summary the innovations of this workare summarized as follows

Point 1 The improved method is on the basis of theinductive transfer learning but it has some difference fromthe traditional inductive transfer learning The latter directlytransfers the samples or features used in the source domain tothe target domain for transfer learning while the former usesa knowledge lever mechanism that transfers some knowledgefrom the source domain to the target domain Then the secu-rity of the data in the source domain can be well protectedand the data in the target domain and the knowledge in thesource domain can be used simultaneously so that the classi-fication effect is better

Point 2 The improved method expands DLSR that can bewell applied to classification scenes into a novel method withcertain transfer learning ability so that it can be used in morecomplex scenes

Point 3 The improved method inherits the characteris-tics of DLSR that is it can be better applied to classificationscenarios by expanding the interval between different catego-ries And it can transfer knowledge from the source domainthus ensuring the rationality of its training model

Finally to better illustrate the basic idea of this study thestructure of the paper is as follows

Part 1 Introduction to the research background statusand significance of the thesis

Part 2 The related work including the related technologyof epilepsy EEG signal detection and the epilepsy EEG signalclassification based on transfer learning is summarized inadvance so that the following sections becomemore readable

Part 3 The notations of the inductive transfer learningalgorithm based on DLSR was introduced in detail

Part 4 The reliability and validity of TDLSR algorithm indetection of epilepsy EEG signals based on a series of exper-imental were verified

2 Related Work

Automatic epilepsy detection is based on signal processingtechnology and pattern recognition It analyzes EEG data toidentify the location and duration of seizures Usually theEEG signals collected during seizures was called seizureEEG and the EEG signals collected in the nonseizure arecalled nonseizure EEG The problem of automatic detectionof epilepsy EEG is to effectively judge the above two typesof EEG signals and identify seizures The related detectiontechnologies are introduced as follows

21 The Related Technology of Epilepsy EEG Signal DetectionBecause the EEG signal of epilepsy is easily interfered bymany factors it is very random and it is a nonstationarysignal and its rule is generally difficult to grasp There-fore researchers often use quantitative analysis to extractcharacteristic information of epilepsy EEG signals Theexisting methods of automatic epilepsy detection includethe following

(1) Time-domain analysis time-domain analysis is oneof the earliest methods used in signal analysis It












Time

Amplitude

(a)

Time

Frequency

(b)

Time

Frequency

(c)


Same distribution

Trainingdata

Testingdata









arg minZ





Trainingdata

Testingdata



EEG signal

Signal data



Feature extraction

Transfer learning

Sourcedomain

Useful knowledge


Targetdomain



Symbol Description


and xi isin Rd




Y


elements are 0












= 0


⟹ XTX + λId

Z = XTY

eth3THORN



XTY eth4THORN




Bij=+1 if yi = 1

minus1 otherwise

(

eth5THORN


minstWge0





XZ + epT minus L 2




= 0



N

eth8THORN



= 0


+ 2λZ = 0


Nminus L

+ λZ = 0

⟹XT IN minuseeT

N


eeT

N

L + λZ = 0


XTHL

eth9THORN


arg minW

stWge0



minWij

Gij minus BijWij

2

stWijge0

eth11THORN



Gij minus BijWij

2 = BijGij minusWij

2 eth12THORN


Wij =max BijGij 0

eth13THORN







minZp


stWge0


eth15THORN






Y





target domain p



Output Z and p

N


F






XZ + epT minus L 2


eth16THORN



= 0



N

eth17THORN



= 0




Nminus L

+ λZ


⟹XT IN minuseeT

N


eeT

N

L



XTHL + ηZS

eth18THORN


arg minW

stWge0










































D1 087 084 089 090 082

D2 079 072 081 080 074

D3 071 062 072 072 068

D4 071 064 070 072 066

D5 082 078 083 084 078

D6 073 060 072 074 061

D7 070 060 070 072 059

D8 063 053 065 064 050



D1 090 092 089 092 096

D2 082 085 082 084 090

D3 072 075 073 076 083

D4 072 075 071 078 082

D5 085 086 084 087 091

D6 075 078 075 080 083

D7 071 078 074 078 083

D8 068 070 066 070 073










D1 D2 D3 D4 D5 D6 D7 D8

05

06

07

08

09

Datasets

Acc

urac

y (

)

LSRSRCRLR

DLSRSVM







5 Conclusion





D1 083 080 084 085 077

D2 074 067 076 076 070

D3 066 059 068 069 065

D4 066 061 067 068 063

D5 078 074 080 080 074

D6 065 056 069 069 058

D7 064 057 067 071 056

D8 060 050 062 061 047



D1 087 086 089 089 093

D2 079 079 082 081 087

D3 069 070 072 073 080

D4 069 068 072 075 079

D5 082 081 083 084 088

D6 072 072 075 077 080

D7 068 071 075 075 080

D8 065 063 067 067 070

D1 D2 D3 D4 D5 D6 D7 D8

065

07

075

08

085

09

095

1

Datasets

Acc

urac

y (

)


LMPROJTDLSR




Data Availability




References











































1 Introduction

2 Related Work











5 Conclusion

Data Availability












Time

Amplitude

(a)

Time

Frequency

(b)

Time

Frequency

(c)


Same distribution

Trainingdata

Testingdata









arg minZ





Trainingdata

Testingdata



EEG signal

Signal data



Feature extraction

Transfer learning

Sourcedomain

Useful knowledge


Targetdomain



Symbol Description


and xi isin Rd




Y


elements are 0












= 0


⟹ XTX + λId

Z = XTY

eth3THORN



XTY eth4THORN




Bij=+1 if yi = 1

minus1 otherwise

(

eth5THORN


minstWge0





XZ + epT minus L 2




= 0



N

eth8THORN



= 0


+ 2λZ = 0


Nminus L

+ λZ = 0

⟹XT IN minuseeT

N


eeT

N

L + λZ = 0


XTHL

eth9THORN


arg minW

stWge0



minWij

Gij minus BijWij

2

stWijge0

eth11THORN



Gij minus BijWij

2 = BijGij minusWij

2 eth12THORN


Wij =max BijGij 0

eth13THORN







minZp


stWge0


eth15THORN






Y





target domain p



Output Z and p

N


F






XZ + epT minus L 2


eth16THORN



= 0



N

eth17THORN



= 0




Nminus L

+ λZ


⟹XT IN minuseeT

N


eeT

N

L



XTHL + ηZS

eth18THORN


arg minW

stWge0










































D1 087 084 089 090 082

D2 079 072 081 080 074

D3 071 062 072 072 068

D4 071 064 070 072 066

D5 082 078 083 084 078

D6 073 060 072 074 061

D7 070 060 070 072 059

D8 063 053 065 064 050



D1 090 092 089 092 096

D2 082 085 082 084 090

D3 072 075 073 076 083

D4 072 075 071 078 082

D5 085 086 084 087 091

D6 075 078 075 080 083

D7 071 078 074 078 083

D8 068 070 066 070 073










D1 D2 D3 D4 D5 D6 D7 D8

05

06

07

08

09

Datasets

Acc

urac

y (

)

LSRSRCRLR

DLSRSVM







5 Conclusion





D1 083 080 084 085 077

D2 074 067 076 076 070

D3 066 059 068 069 065

D4 066 061 067 068 063

D5 078 074 080 080 074

D6 065 056 069 069 058

D7 064 057 067 071 056

D8 060 050 062 061 047



D1 087 086 089 089 093

D2 079 079 082 081 087

D3 069 070 072 073 080

D4 069 068 072 075 079

D5 082 081 083 084 088

D6 072 072 075 077 080

D7 068 071 075 075 080

D8 065 063 067 067 070

D1 D2 D3 D4 D5 D6 D7 D8

065

07

075

08

085

09

095

1

Datasets

Acc

urac

y (

)


LMPROJTDLSR




Data Availability




References











































1 Introduction

2 Related Work











5 Conclusion

Data Availability








arg minZ





Trainingdata

Testingdata



EEG signal

Signal data



Feature extraction

Transfer learning

Sourcedomain

Useful knowledge


Targetdomain



Symbol Description


and xi isin Rd




Y


elements are 0












= 0


⟹ XTX + λId

Z = XTY

eth3THORN



XTY eth4THORN




Bij=+1 if yi = 1

minus1 otherwise

(

eth5THORN


minstWge0





XZ + epT minus L 2




= 0



N

eth8THORN



= 0


+ 2λZ = 0


Nminus L

+ λZ = 0

⟹XT IN minuseeT

N


eeT

N

L + λZ = 0


XTHL

eth9THORN


arg minW

stWge0



minWij

Gij minus BijWij

2

stWijge0

eth11THORN



Gij minus BijWij

2 = BijGij minusWij

2 eth12THORN


Wij =max BijGij 0

eth13THORN







minZp


stWge0


eth15THORN






Y





target domain p



Output Z and p

N


F






XZ + epT minus L 2


eth16THORN



= 0



N

eth17THORN



= 0




Nminus L

+ λZ


⟹XT IN minuseeT

N


eeT

N

L



XTHL + ηZS

eth18THORN


arg minW

stWge0










































D1 087 084 089 090 082

D2 079 072 081 080 074

D3 071 062 072 072 068

D4 071 064 070 072 066

D5 082 078 083 084 078

D6 073 060 072 074 061

D7 070 060 070 072 059

D8 063 053 065 064 050



D1 090 092 089 092 096

D2 082 085 082 084 090

D3 072 075 073 076 083

D4 072 075 071 078 082

D5 085 086 084 087 091

D6 075 078 075 080 083

D7 071 078 074 078 083

D8 068 070 066 070 073










D1 D2 D3 D4 D5 D6 D7 D8

05

06

07

08

09

Datasets

Acc

urac

y (

)

LSRSRCRLR

DLSRSVM







5 Conclusion





D1 083 080 084 085 077

D2 074 067 076 076 070

D3 066 059 068 069 065

D4 066 061 067 068 063

D5 078 074 080 080 074

D6 065 056 069 069 058

D7 064 057 067 071 056

D8 060 050 062 061 047



D1 087 086 089 089 093

D2 079 079 082 081 087

D3 069 070 072 073 080

D4 069 068 072 075 079

D5 082 081 083 084 088

D6 072 072 075 077 080

D7 068 071 075 075 080

D8 065 063 067 067 070

D1 D2 D3 D4 D5 D6 D7 D8

065

07

075

08

085

09

095

1

Datasets

Acc

urac

y (

)


LMPROJTDLSR




Data Availability




References











































1 Introduction

2 Related Work











5 Conclusion

Data Availability




= 0


⟹ XTX + λId

Z = XTY

eth3THORN



XTY eth4THORN




Bij=+1 if yi = 1

minus1 otherwise

(

eth5THORN


minstWge0





XZ + epT minus L 2




= 0



N

eth8THORN



= 0


+ 2λZ = 0


Nminus L

+ λZ = 0

⟹XT IN minuseeT

N


eeT

N

L + λZ = 0


XTHL

eth9THORN


arg minW

stWge0



minWij

Gij minus BijWij

2

stWijge0

eth11THORN



Gij minus BijWij

2 = BijGij minusWij

2 eth12THORN


Wij =max BijGij 0

eth13THORN







minZp


stWge0


eth15THORN






Y





target domain p



Output Z and p

N


F






XZ + epT minus L 2


eth16THORN



= 0



N

eth17THORN



= 0




Nminus L

+ λZ


⟹XT IN minuseeT

N


eeT

N

L



XTHL + ηZS

eth18THORN


arg minW

stWge0










































D1 087 084 089 090 082

D2 079 072 081 080 074

D3 071 062 072 072 068

D4 071 064 070 072 066

D5 082 078 083 084 078

D6 073 060 072 074 061

D7 070 060 070 072 059

D8 063 053 065 064 050



D1 090 092 089 092 096

D2 082 085 082 084 090

D3 072 075 073 076 083

D4 072 075 071 078 082

D5 085 086 084 087 091

D6 075 078 075 080 083

D7 071 078 074 078 083

D8 068 070 066 070 073










D1 D2 D3 D4 D5 D6 D7 D8

05

06

07

08

09

Datasets

Acc

urac

y (

)

LSRSRCRLR

DLSRSVM







5 Conclusion





D1 083 080 084 085 077

D2 074 067 076 076 070

D3 066 059 068 069 065

D4 066 061 067 068 063

D5 078 074 080 080 074

D6 065 056 069 069 058

D7 064 057 067 071 056

D8 060 050 062 061 047



D1 087 086 089 089 093

D2 079 079 082 081 087

D3 069 070 072 073 080

D4 069 068 072 075 079

D5 082 081 083 084 088

D6 072 072 075 077 080

D7 068 071 075 075 080

D8 065 063 067 067 070

D1 D2 D3 D4 D5 D6 D7 D8

065

07

075

08

085

09

095

1

Datasets

Acc

urac

y (

)


LMPROJTDLSR




Data Availability




References











































1 Introduction

2 Related Work











5 Conclusion

Data Availability







minZp


stWge0


eth15THORN






Y





target domain p



Output Z and p

N


F






XZ + epT minus L 2


eth16THORN



= 0



N

eth17THORN



= 0




Nminus L

+ λZ


⟹XT IN minuseeT

N


eeT

N

L



XTHL + ηZS

eth18THORN


arg minW

stWge0










































D1 087 084 089 090 082

D2 079 072 081 080 074

D3 071 062 072 072 068

D4 071 064 070 072 066

D5 082 078 083 084 078

D6 073 060 072 074 061

D7 070 060 070 072 059

D8 063 053 065 064 050



D1 090 092 089 092 096

D2 082 085 082 084 090

D3 072 075 073 076 083

D4 072 075 071 078 082

D5 085 086 084 087 091

D6 075 078 075 080 083

D7 071 078 074 078 083

D8 068 070 066 070 073










D1 D2 D3 D4 D5 D6 D7 D8

05

06

07

08

09

Datasets

Acc

urac

y (

)

LSRSRCRLR

DLSRSVM







5 Conclusion





D1 083 080 084 085 077

D2 074 067 076 076 070

D3 066 059 068 069 065

D4 066 061 067 068 063

D5 078 074 080 080 074

D6 065 056 069 069 058

D7 064 057 067 071 056

D8 060 050 062 061 047



D1 087 086 089 089 093

D2 079 079 082 081 087

D3 069 070 072 073 080

D4 069 068 072 075 079

D5 082 081 083 084 088

D6 072 072 075 077 080

D7 068 071 075 075 080

D8 065 063 067 067 070

D1 D2 D3 D4 D5 D6 D7 D8

065

07

075

08

085

09

095

1

Datasets

Acc

urac

y (

)


LMPROJTDLSR




Data Availability




References











































1 Introduction

2 Related Work











5 Conclusion

Data Availability


Y





target domain p



Output Z and p

N


F






XZ + epT minus L 2


eth16THORN



= 0



N

eth17THORN



= 0




Nminus L

+ λZ


⟹XT IN minuseeT

N


eeT

N

L



XTHL + ηZS

eth18THORN


arg minW

stWge0










































D1 087 084 089 090 082

D2 079 072 081 080 074

D3 071 062 072 072 068

D4 071 064 070 072 066

D5 082 078 083 084 078

D6 073 060 072 074 061

D7 070 060 070 072 059

D8 063 053 065 064 050



D1 090 092 089 092 096

D2 082 085 082 084 090

D3 072 075 073 076 083

D4 072 075 071 078 082

D5 085 086 084 087 091

D6 075 078 075 080 083

D7 071 078 074 078 083

D8 068 070 066 070 073










D1 D2 D3 D4 D5 D6 D7 D8

05

06

07

08

09

Datasets

Acc

urac

y (

)

LSRSRCRLR

DLSRSVM







5 Conclusion





D1 083 080 084 085 077

D2 074 067 076 076 070

D3 066 059 068 069 065

D4 066 061 067 068 063

D5 078 074 080 080 074

D6 065 056 069 069 058

D7 064 057 067 071 056

D8 060 050 062 061 047



D1 087 086 089 089 093

D2 079 079 082 081 087

D3 069 070 072 073 080

D4 069 068 072 075 079

D5 082 081 083 084 088

D6 072 072 075 077 080

D7 068 071 075 075 080

D8 065 063 067 067 070

D1 D2 D3 D4 D5 D6 D7 D8

065

07

075

08

085

09

095

1

Datasets

Acc

urac

y (

)


LMPROJTDLSR




Data Availability




References











































1 Introduction

2 Related Work











5 Conclusion

Data Availability



































D1 087 084 089 090 082

D2 079 072 081 080 074

D3 071 062 072 072 068

D4 071 064 070 072 066

D5 082 078 083 084 078

D6 073 060 072 074 061

D7 070 060 070 072 059

D8 063 053 065 064 050



D1 090 092 089 092 096

D2 082 085 082 084 090

D3 072 075 073 076 083

D4 072 075 071 078 082

D5 085 086 084 087 091

D6 075 078 075 080 083

D7 071 078 074 078 083

D8 068 070 066 070 073










D1 D2 D3 D4 D5 D6 D7 D8

05

06

07

08

09

Datasets

Acc

urac

y (

)

LSRSRCRLR

DLSRSVM







5 Conclusion





D1 083 080 084 085 077

D2 074 067 076 076 070

D3 066 059 068 069 065

D4 066 061 067 068 063

D5 078 074 080 080 074

D6 065 056 069 069 058

D7 064 057 067 071 056

D8 060 050 062 061 047



D1 087 086 089 089 093

D2 079 079 082 081 087

D3 069 070 072 073 080

D4 069 068 072 075 079

D5 082 081 083 084 088

D6 072 072 075 077 080

D7 068 071 075 075 080

D8 065 063 067 067 070

D1 D2 D3 D4 D5 D6 D7 D8

065

07

075

08

085

09

095

1

Datasets

Acc

urac

y (

)


LMPROJTDLSR




Data Availability




References











































1 Introduction

2 Related Work











5 Conclusion

Data Availability










D1 D2 D3 D4 D5 D6 D7 D8

05

06

07

08

09

Datasets

Acc

urac

y (

)

LSRSRCRLR

DLSRSVM







5 Conclusion





D1 083 080 084 085 077

D2 074 067 076 076 070

D3 066 059 068 069 065

D4 066 061 067 068 063

D5 078 074 080 080 074

D6 065 056 069 069 058

D7 064 057 067 071 056

D8 060 050 062 061 047



D1 087 086 089 089 093

D2 079 079 082 081 087

D3 069 070 072 073 080

D4 069 068 072 075 079

D5 082 081 083 084 088

D6 072 072 075 077 080

D7 068 071 075 075 080

D8 065 063 067 067 070

D1 D2 D3 D4 D5 D6 D7 D8

065

07

075

08

085

09

095

1

Datasets

Acc

urac

y (

)


LMPROJTDLSR




Data Availability




References











































1 Introduction

2 Related Work











5 Conclusion

Data Availability






5 Conclusion





D1 083 080 084 085 077

D2 074 067 076 076 070

D3 066 059 068 069 065

D4 066 061 067 068 063

D5 078 074 080 080 074

D6 065 056 069 069 058

D7 064 057 067 071 056

D8 060 050 062 061 047



D1 087 086 089 089 093

D2 079 079 082 081 087

D3 069 070 072 073 080

D4 069 068 072 075 079

D5 082 081 083 084 088

D6 072 072 075 077 080

D7 068 071 075 075 080

D8 065 063 067 067 070

D1 D2 D3 D4 D5 D6 D7 D8

065

07

075

08

085

09

095

1

Datasets

Acc

urac

y (

)


LMPROJTDLSR




Data Availability




References











































1 Introduction

2 Related Work











5 Conclusion

Data Availability



Data Availability




References











































1 Introduction

2 Related Work











5 Conclusion

Data Availability


















1 Introduction

2 Related Work











5 Conclusion

Data Availability


An Automatic Epilepsy Detection Method Based on Improved...

Documents

Transcript of An Automatic Epilepsy Detection Method Based on Improved...