Access the full text.
Sign up today, get DeepDyve free for 14 days.
Hindawi Journal of Healthcare Engineering Volume 2019, Article ID 3817124, 12 pages https://doi.org/10.1155/2019/3817124 Research Article Brain-Computer Interface Channel-Selection Strategy Based on Analysis of Event-Related Desynchronization Topography in Stroke Patients 1 1 2 1 2 Chong Li , Tianyu Jia , Quan Xu, Linhong Ji , and Yu Pan Division of Intelligent and Bio-mimetic Machinery, e State Key Laboratory of Tribology, Tsinghua University, Beijing, China Department of Physical Medicine and Rehabilitation, Beijing Tsinghua Changgung Hospital, School of Clinical Medicine, Tsinghua University, Beijing, China Correspondence should be addressed to Linhong Ji; email@example.com and Yu Pan; firstname.lastname@example.org Received 17 April 2019; Revised 12 June 2019; Accepted 13 August 2019; Published 28 August 2019 Guest Editor: Avinash Parnandi Copyright©2019ChongLietal.+is isan openaccessarticledistributedunderthe CreativeCommonsAttributionLicense, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. In the last decade, technology-assisted stroke rehabilitation has been the focus of research. Electroencephalogram- (EEG-) based brain-computer interface (BCI) has a great potential for motor rehabilitation in stroke patients since the closed loop between motor intention and the actual movement established by BCI can stimulate the neural pathways of motor control. Due to the deﬁcits in the brain, motor intention expression may shift to other brain regions during and even after neural reorganization. +e objective of this paper was to study the event-related desynchronization (ERD) topography during motor attempt tasks of the paretic hand in stroke patients and compare the classiﬁcation performance using diﬀerent channel-selection strategies in EEG- based BCI.Fifteenstrokepatientswererecruitedinthisstudy.Acue-basedexperimentalparadigmwasappliedintheexperiment, in which each patient was required to open the palm of the paretic or the unaﬀected hand. EEG was recorded and analyzed to measure the motor intention and indicate the activated brain regions. Support vector machine (SVM) combined with common spatial pattern (CSP) algorithm was used to calculate the oﬄine classiﬁcation accuracy between the motor attempt of the paretic hand and the resting state applying diﬀerent channel-selection strategies. Results showed individualized ERD topography during the motor attempt of the paretic hand due to the deﬁcits caused by stroke. Statistical analysis showed a signiﬁcant increase in the classiﬁcation accuracy by analyzing the channels showing ERD than analyzing the channels from the contralateral sensorimotor cortex (SM1). +e results indicated that for stroke patients whose aﬀected motor cortex is extensively damaged, the compensated brain regions should be considered for implementing EEG-based BCI for motor rehabilitation as the closed loop between the altered activated brain regions and the paretic hand can be stimulated more accurately using the individualized channel- selection strategy. patients and their families, especially with impaired upper 1. Introduction extremity, because lack of arm-movement control aﬀects According to the estimations, some 16 million people per independent daily living. year experience stroke, from which about two-thirds survive Eﬀective treatment and therapy for stroke rehabilitation worldwide . Stroke remains the most common cause of has been the focus of research. Recent studies have dem- disability for adults. Stroke survivors suﬀer various deﬁcits onstrated that electroencephalogram- (EEG-) based brain- that generate disability in motor, perceptual, and cognitive computer interface (BCI) has a great potential for motor functioning . Among these disabilities, motor deﬁcits rehabilitation instrokepatients[4–6],whichishypothesized have a large impact on managing everyday activities . that closing the loop between cortical activity (imagined or Hemiplegia caused by stroke brings terrible burden on attempted motor intention) and actual movement can 2 Journal of Healthcare Engineering restore functional corticospinal and corticomuscular con- (3) 25 to 80years of age nections.InapplicationswithEEG-basedBCIforhealthy (4) Able to understand the physician’s orders subjects, motor intention of the unilateral arm or hand can be indicated by the decline in the power of the sensorimotor rhythm, such as event-related desynchronization (ERD) 2.1.2. Exclusion Criteria [8–10], in the contralateral sensorimotor cortex (SM1) . (1) Aphasia, severe cognitive impairment, and severe Similarly, in stroke rehabilitation with EEG-based BCI, SM1 depression is chosen to detect ERD [12–15]. (2) Unstable condition, including orthostatic hypoten- +e principle behind this approach is to encourage the sion, sepsis, and epilepsy use of SM1 in the aﬀected hemisphere to induce the re- organization of the original neural circuits for motor control (3) Unilateralspatialneglectorseverevisualimpairment . However, if the aﬀected motor cortex is extensively (4) Ashworth of aﬀected shoulder, elbow, or wrist and damaged, then vicariation of function will appear with in- hand larger than 2 creased levels of activity in the unaﬀected hemisphere . (5) Heart, lung, liver, kidney, and other vital organs Functional magneticresonanceimaging (fMRI)studies have function decline or failure revealed widespread changes in the patterns of brain acti- vation during simple movements of the aﬀected hand after All subjects exhibited hemiparesis, which was shown by stroke [18, 19], and recruitment in the unaﬀected hemi- Fugl-Meyer assessment (FMA) (scores range from 0 to 100, sphere can also contribute to a good recovery for patients with higher scores representing better function). +e sub- with deﬁcits in the aﬀected motor cortex . jects recruited by this study had variable types of stroke and +erefore, for stroke patients whose movement of the large FMA range. +is was because the original idea of this paretic hand is compensated by other brain region(s), the study was to analyze ERD topography of stroke patients in startlocationoftheloopbetweenmotorintentionandactual diﬀerent conditions and compare the classiﬁcation accuracy movement is expected to be diﬀerent not only from that of using diﬀerent channel-selection strategies in order to show healthy subjects but also from other patients due to brain that an individualized channel-selection strategy in BCI reorganization. +e objective of this paper, therefore, was to rehabilitation may be needed. Detailed information of the investigate ERD topography of stroke patients in diﬀerent patients (S1–S15) is shown in Table 1. +ey gave written conditions and compare the classiﬁcation accuracy of the consent prior to participating in the study. +is study was motorattemptoftheparetichandusingEEGdatafromSM1 conducted according to the principles expressed in the or from the channels showing ERD. +e closed loop with an Declaration of Helsinki and approved by the Institutional individualized start location according to each patient’s Ethical Committee. condition may be able to stimulate the neural circuit of motor control more accurately for stroke patients, especially theoneswhodevelopalteredbrainregionsforcompensating 2.2. Experimental Paradigm. +e experiment was conducted aﬀected motor function . between patients and theexperimenter in a quiet dark room. In the experiment, EEG was recorded when stroke pa- Each patient was seated in a comfortable chair with arms tients were required to make movement attempts with their supported, and the experimenter explained the task to each paretic hand. +e acquired EEG was analyzed to identify the patientbeforehand.Acue-basedparadigmwasappliedusing motor intention of the paretic hand. Section 2 describes the OpenViBE V2.0.1. A 15-inch computer monitor was placed conduct of the experiment, and then, Sections 3 and 4 about 1meter in front of the subjects. Each patient was present and discuss the results of the experiment, requested to stare at the center of the monitor and respond respectively. to the cue. Two sessions were conducted for each patient, and each run consisted 40 trials. As shown in Figure 1, each trial started with a ﬁxation cross for 3seconds, at which the 2. Methods patient was required to stare to avoid excessive eye move- 2.1. Subjects. +e study recruited 15 stroke patients. To be ments. After this, an arrow cue appeared on the monitor able to respond precisely to the experiment task, the patient whenthemotorattempttaskneededtobeperformed.Motor should be in a stable condition and capable of communi- attempt has been proved to have better performance than cating with the experimenter. Since the best time for re- motor imagery in BCI . In total, 20 arrow cues and 20 habilitation is within 6months after onset, this experiment blank controls were demonstrated in each session in a selected patients that are less than 6months after stroke. +e randomized order. In the ﬁrst session, the patient was asked inclusion and exclusion criteria were described as follows. to attempt opening the palm of the paretic hand when the arrow pointed to the pareticside even though themovement of the paretic hand could not be truly executed; when the 2.1.1. Inclusion Criteria arrow did not appear (blank control), the patient need not (1) Diagnosed with stroke by magnetic resonance im- doanythingbutstareatthecrosscenter.+eﬁrstsessionwas aging (MRI) or computed tomography (CT) ex- used to classify between the motor intention of the paretic amination for the ﬁrst time handandtherestingstate,whichwasneededinBCItherapy, (2) Less than 6months after onset and no spasticity rather than to classify motor intention of the aﬀected hand Journal of Healthcare Engineering 3 Table 1: Patient information and clinical evaluation. Subject Age Lesion site Lesion type Paretic side TFO (days) FMA S1 39 Left basal ganglia region Hemorrhage Right 18 60 S2 66 Left basal ganglia region Infarction Right 20 25 S3 32 Left pons Infarction Right 36 87 S4 58 Left pons Infarction Right 28 35 S5 47 Left basal ganglia region Hemorrhage Right 8 10 S6 78 Left pons Infarction Right 53 21 S7 62 Left middle cerebral artery area Infarction Right 129 49 S8 64 Brainstem Infarction Right 13 88 S9 49 Left basal ganglia region Middle cerebral artery area Hemorrhage Right 73 45 S10 73 Right basal ganglia region Infarction Left 51 56 Bilateral basal ganglia regions S11 66 Infarction Left 46 79 Parietal lobe S12 61 Left occipital-parietal lobe Hemorrhage Right 72 93 S13 43 Left basal ganglia region Hemorrhage Right 170 22 S14 60 Bilateral frontal-temporo-parietal lobe Infarction Right 82 24 S15 30 Right temporo-parietal lobe Hemorrhage Left 50 45 Average±SD 57±45 49±27 TFO, time from onset; FMA, Fugl-Meyer assessment. 2.3. EEG Recordings and Data Processing. During the ex- periments, EEG data were recorded with ANTeego rt from Right hand movement 64 Ag/AgCl electrodes, positioned according to the in- ternational 10/20 system and streamed through OpenViBE V2.0.1, with CPz as reference and AFz as ground, digitally 3s 1s 3s 5-6s sampled at 500Hz with 24-bit resolution. Electrode im- pedances were kept below 5KΩ. EEG recordings from all 64 Resting state Display cue Motor attempt Blank channels were raw data without any preprocessing by the acquisition software. +e EEG data were preprocessed in EEGLAB 14.1.2b Left hand (EEGLAB toolbox, Swartz Center for Computational Neu- movement rosciences, La Jolla, CA; https://sccn.ucsd.edu/eeglab). +e 32nd channel EOG was excluded from the EEG data. +e raw data were referenced using an average reference [8, 23]. Figure 1: Each trial started with a ﬁxation cross displayed in the centerofthescreen,whichlastedforthewholetrialfromtheﬁrstto +en, the data were ﬁltered using a ﬁnite impulse response seventh second. +e cue for motor attempt appeared from the (FIR) bandpass ﬁlter (1–40Hz), and the baseline was fourthtoﬁfthsecondasaredarrowpointingtotherightorleftside removed. specifying the task to be performed. When no arrow appeared, it +e preprocessed EEG data in each session were divided was the blank control, during which the subject need not do into 40 trials based on the trial marks, with each epoch anythingbutstareatthecrosscenter.Patientswererequiredtostop containing 220500 data points for all 63 channels. 40 epochs the task when the cue disappeared at the end of the seventh second, of data were categorized into two groups: (i) motor attempt and only the ﬁxation cross was shown till the end of the trial. To group and (ii) blank control group. Each trial was detected avoid adaptation to the timing, cues were presented in randomized one by one. If artefacts were visually detected in the trial, sequence with randomized intervals between 5seconds and then the trial was processed using independent component 6seconds between each trial. analysis (ICA) in EEGLAB 14.1.2b using runica ICA algo- rithm . If the artefacts can still be observed, the cor- and the unaﬀected hand. So, in the second session, ERD responding trial was excluded from the following analysis. Time-frequency analysis of each epoch was conducted using topographyduringmovementoftheunaﬀectedhandwasalso collected only for comparison. In the second session, the Morlet Wavelet  in alpha (8–13Hz) and beta (13–30Hz) range, respectively, with a step of 0.5Hz. +en, power patient was asked to attempt opening the palm of the un- aﬀected hand when the arrow pointed to the unaﬀected side; spectral density (PSD) of alpha and beta band for all when the arrow did not appear (blank control), the patient channels in the same group was averaged across epochs to need not do anything but stare at the cross center. To avoid maximize features and minimize noises. Based on the ERD adaptation to the timing , cues were presented in ran- calculation , modiﬁed ERDratio was calculated: for each domized sequence with randomized intervals between channel, integration across time and frequency in each 5seconds and 6seconds between each trial. Each trial lasted frequency band was calculated in resting state and moving 7seconds, and each session lasted about 9minutes. A state,respectively.+etwostateswereseparatedbytheevent 5minutes’restwasgivenaftertheﬁrstsessiontoavoidfatigue. marks from the cues. With the results of the integration of 4 Journal of Healthcare Engineering both states, the ERD ratio was calculated using the following example of S4 is shown in Figure 3, in which no apparent equation in order to quantify the changes of the spectral decline was found in power spectral density during the power in a certain frequency range at a channel c: experiment. For S1, S2, S3, S5, S6, S7, S9, S12, and S13, ERD was found around C3 and/or C4; while for S8, ERD focused h E 1/R · spe E − S f�l t�S c 1 1 2 on temporal areas C4 and FC4; for S10, ERD focused on the ERDratio � · , (1) E − S 2 2 1/R · spe centralparietalarea;forS15,ERDfocusedonthecentraland f�l t�S c frontal central areas. where frequency ranges from l to h(Hz), sampling rate is MorepatternsofERDtopographywereidentiﬁedduring R(Hz), S and E are the start and end time points of each the motor attempt of the paretic hand (Figure 2). +e state:1 indicates the resting state and 2 indicates the moving subjects were divided into 5 groups based on the following state, spe (μV /Hz · ms) is the power spectral density. criteria: +e ERD/ERS are not phase locked to the event, and (i) Whether ERD appears during hand movement or ERD/ERS are highly frequency band-speciﬁc signals . not +erefore, in our study, the frequency band with lower ERD (ii) Whether ERD appears on the ipsilateral or the ratio was selected to represent motor intention. +en, the contralateral side correspondingcalculatedERDratioofeach channelforeach subject was plotted as ratio maps. (iii) Whether ERD is identiﬁed on speciﬁc sites or in a wide range SpeciﬁcgroupingofthepatientsisshowninFigure4and 2.4. Classiﬁcation Accuracy. +e accuracy of classifying listed as follows: between motor attempt of the paretic hand and the resting state was calculated using MATLAB (MathWorks Inc., (i) ERD-blind group: for S4, S11, and S14, no ERD was USA). In this paper, we applied three strategies to select the foundinalphaand/orbetafrequencybandsoverthe EEG channels for analyzing the classiﬁcation performance. aﬀected and/or the unaﬀected hemispheres. +e ﬁrst strategy, SM1-4, was to select 4 channels on the (ii) ERD-disappearance group: S2 and S15 were not contralateral SM1, either Cz, C1, C3, C5, or Cz, C2, C4, C6, ERD blind as ERD was identiﬁed during their according to the side of the aﬀected hand . +e second movement of the unaﬀected hand. So the disap- strategy (SM1-5) was similar with the ﬁrst one which se- pearance of ERD during the motor attempt of the lected 5channels on thecontralateral side,either C1,C3, C5, paretic hand may be due to the deﬁcits in the af- FC3,CP3,orC2,C4,C6,FC4,CP4.Itwasbelievedthese fected hemisphere. Time-frequency analysis of EEG two classical channel-selection practices could close the loop during the paretic hand motor attempt and un- between the original brain region of motor control and the aﬀected hand motor attempt for S2 is shown in actual movement. +e third strategy, ERD selection, was to Figure 5. select 4 channels with the lowest ERD ratio to formulate the (iii) ERD-proliferation group: for S1, S3, S6, and S9, individualized loop of motor intention and motor attempt ERD was detected in a wide range of electrodes according to each patient’s condition. +e selected 40 trials covering brain regions with a stronger focus on the of EEG data were divided into two groups, namely, the unaﬀected hemispheres. For S3, ERD was found on motor attempt of the paretic hand and resting state. In the bilateral hemispheres with a stronger focus on addition to the data processing mentioned above, spatial the sensorimotor cortex area. ﬁltering was also conducted using the common spatial pattern (CSP) in order to extract features for classiﬁcation (iv) ERD-SM1 group: for S5, S10, and S12, ERD focused . +e dimension of the feature matrix is two times as on the C3 or C4 of the aﬀected hemisphere, whereas many as numbers of the selected channels. After the feature for S7, ERD focused on the bilateral C3 and C4. extraction, the selected features were then classiﬁed in a (v) Others: for S8, slight ERD was detected in C4. For classiﬁerbasedonsupportvectormachine(SVM)algorithm. S13, slight ERD was detected in C1 and Cz. +e linear kernel was applied [30, 31]. +e accuracy of a 5- fold cross-validation test, using 80% data sets for training and 20% for testing, was calculated. ANOVA with repeated 3.2. Channel-Selection Strategies and Classiﬁcation measures was applied for statistical analysis to analyze the Performance. +e accuracy of classifying motor attempt of diﬀerenceintheclassiﬁcationaccuracybetweenthechannel- the paretic hand and resting state is shown in Table 2, selection strategies. applying diﬀerent channel-selection strategies. Since no ERD was detected for S2, S4, S11, S14, and S15 during the motor attempt of the paretic hand, the classiﬁcation accu- 3. Results racy was not calculated. +e classiﬁcation accuracy for all 3.1. ERD Topography. As shown in Figure 2, during the other subjects was above chance-level (50%). +e classiﬁ- motor attempt of the unaﬀected hand, ERD was detected in cation accuracy using ERD selection was at least 7.5% higher alpha and/or beta frequency bands in S1, S2, S3, S5, S6, S7, than that using SM1-4 for S1, S3, S5, S6, and S9. For S7, S8, S10, S12, and S13, the results yielded by SM1-4 was 2.5% or S8, S9, S10, S12, S13, and S15. ERD was not detected in S4, S11, and S14, who were regarded as ERD blind . +e 5% lower than that by ERD selection. +e classiﬁcation Journal of Healthcare Engineering 5 Subject 1 Subject 2 Subject 3 Subject 4 Subject 5 (8–13 Hz) (13–30 Hz) (8–13 Hz) (13–30 Hz) (8–13 Hz) Paretic hand movement Unaffected hand movement (a) (b) (c) (d) (e) Subject 6 Subject 7 Subject 8 Subject 9 Subject 10 (8–13 Hz) (8–13 Hz) (8–13 Hz) (8–13 Hz) (13–30 Hz) Paretic hand movement Unaffected hand movement (f ) (g) (h) (i) (j) Subject 11 Subject 12 Subject 13 Subject 14 Subject 15 (13–30 Hz) (13–30 Hz) (13–30 Hz) (13–30 Hz) (8–13 Hz) Paretic hand movement Unaffected hand movement (k) (l) (m) (n) (o) 0.75 0.8 0.85 0.9 0.95 1.0 1.05 Figure 2: ERD ratio-maps: ERD activation during paretic hand movement and unaﬀected hand movement relative to resting state overlaid on topography, respectively, for 15 stroke patients. +e cross indicates the side of the paretic hand. Blue regions indicate the involved areas when ERD occurs during mental tasks. accuracy using ERD selection was at least 10% higher than selection: 71.88±3.75%). One-way ANOVA with repeated thatusingSM1-5forS1,S3,andS10.+eaccuracygenerated measure determined that mean classiﬁcation accuracy dif- by SM1-5 was 2.5% or 5% lower than that by ERD selection fered signiﬁcantly between diﬀerent channel-selection for S6,S7, S8,S9, S12,andS13.Resultswere equalfor S5.For strategies (F (2, 18) �19.442, p<0.0005). Post hoc tests the ERD-proliferation group, the accuracy of ERD selection using the Bonferroni correction revealed that there was a strategy was 79.38±10.28%, higher than that of SM1-4 signiﬁcant diﬀerence in the classiﬁcation accuracy between (70.00±9.35%) or SM1-5 (71.25±8.54%) selection strategy. the SM1-4 channel-selection strategy (69.50±8.48%) and For the ERD-SM1 group, the accuracy among these three ERD Selection (75.75±8.98%) (p � 0.001) and between the channel-selection strategies was slightly diﬀerent (SM1-4 SM1-5 channel-selection strategy (71.25±7.84%) and ERD selection: 67.50±3.54%, SM1-5: 70.00±4.56%, ERD Selection (75.75±8.98%) (p � 0.015). However, there was Time (s) Time (s) 6 Journal of Healthcare Engineering Time-frequency analysis in C4 Time-frequency analysis in C3 1 25 2 20 20 3 5 15 15 6 1.05 1.05 0.95 0.95 0.9 0.9 0.85 0.85 0.8 0.8 0.75 0.75 Paretic hand movement Unaffected hand movement (a) (b) Figure 3: Time-frequency graph and ERD topography during motor attempt for S4. During the motor attempt of the paretic or unaﬀected hand, based on Morlet Wavelet time-frequency analysis, no apparent decline was found in power spectral density during the experiment. Correspondingly, no ERD was shown in the ERD ratio-maps. (a) Paretic hand movement. (b) Unaﬀected hand movement. no signiﬁcant diﬀerence in the classiﬁcation accuracy be- in the aﬀected hemisphere. While for S1 with higher FMA tween the SM1-4 channel-selection strategy (69.50±8.48%) than that of S6 and S9, there was a stronger focus of ERD and SM1-5 channel-selection strategy (71.25±7.84%) around CP1, CPz, and TP8. For S3 who was also evaluated with a rather high FMA, ERD was found on bilateral (p � 0.134). +erefore, we can conclude that ERD channel- selectionstrategygeneratedastatisticallysigniﬁcantincrease hemisphereswithastrongerfocusonSM1.Itcanbeinferred for S3, recruitment of compensatory brain regions may in classiﬁcation accuracy than the other two classical channel-selection strategies. narrow down and ultimately activation area of motor control may focus on the contralateral SM1 again . (iv) ERD-SM1 group: For S7, S10, and S12, ERD was found 4. Discussion around SM1 in the aﬀected hemisphere, which means SM1 4.1. ERD Topography and Clinical Assessment. +e results wasstillcapableofexpressingmotorintention.Accordingto showedERDtopographyvariedduringthemotorattemptof thefactthatthese3subjectshadhigh FMA,wecaninferthat the paretic hand due to the deﬁcits caused by stroke . SM1 was preserved or has recovered from the brain lesion. Based on the pattern of ERD topography and clinical as- However, for S5, the day of experiment was only 8days after stroke onset. Although motor intention was detected on sessment, the patients’ condition can be further inferred. (i) ERD-blind group: ERD blind subjects were not taken into SM1, the neural pathway between cortical activity andactual movement could be seriously aﬀected, which may be the consideration. (ii) ERD-disappearance group: ERD was found during movement of the unaﬀected hand but not cause of a low FMA. (v) Others: S13 was diagnosed with found during the motor attempt of the paretic hand. It can hemorrhage in the left basal ganglia, which may have a be inferred that the expression of motor intention of the serious eﬀect on motor function, leading to a rather low paretic hand may be inﬂuenced by the lesion, which can also FMA. +e electrode-detecting motor intention shifted to a be demonstrated by a rather low FMA. (iii) ERD-pro- narrow area as shown in the ERD topography. Although the liferation group: For S6 and S9, ERD was shown in a wide electrode-detecting motor intention shifted to C4 on the range of electrodes covering brain regions with areas unaﬀected hemisphere for S8, the infarction may have less compensatingfortheexpressionsofmotorintentionbySM1 severe eﬀects on motor control as demonstrated by a higher Frequency (Hz) Frequency (Hz) Power spectral density (µV /Hz·ms) Power spectral density (µV /Hz·ms) Journal of Healthcare Engineering 7 S4 S11 S14 (i) ERD-blind group S2 S15 (ii) ERD-disappearance group S1 S3 S6 S9 (iii) ERD-proliferation group S5 S7 S10 S12 (iv) ERD-SM1 group S8 S13 (v) Others 0.75 0.8 0.85 0.9 0.95 1.0 1.05 Figure 4: Grouping of the patients: ERD ratio-maps during paretic hand movement relative to resting state. +e cross indicated the side of the paretic hand. Blue regions indicate the involved brain areas when ERD occurs during mental tasks. FMA than the hemorrhage for S13 . Brain re- classical SM1-4 channel-selection performance was com- pared with the individualized ERD channel-selection organization is inﬂuenced by numerous factors, such as lesion site, lesion type, and TFO. ERD topography can be strategy: (i) ERD-blind group and (ii) ERD-disappearance used as an additional supplement for clinical evaluation for group: patients in the ERD-blind group and ERD-disap- stroke patients’ condition, especially the progress of brain pearancegroupwerenottakenintoconsideration.(iii)ERD- reorganization, as it can indicate the brain regions proliferation group: the classiﬁcation accuracy using ERD expressing motor intention during motor attempt of the selection was at least 7.5% higher than that using SM1-4. In paretic hand. these subjects, ERD was detected in a wide range of elec- trodes covering brain regions including the unaﬀected hemisphere which compensated for the aﬀected motor function.+erefore,moreeﬀectiveclassiﬁcationfeaturescan 4.2. Channel-Selection Strategies and Classiﬁcation Performance. +e aim of this study focused on investigating beextractedfromthechannelsshowingERD,whichresulted in a much higher classiﬁcation accuracy applying in- the expression of motor intention, which is an important element of the BCI neural circuit . +e oﬄine motor dividualized channel selection strategy. (iv) ERD-SM1 intention classiﬁcation accuracy showed the performance of group:ForS7,S10,andS12,ERDwasshownaroundC3and/ the classiﬁer and thus can predict the online BCI classiﬁ- or C4 of the aﬀected hemisphere. +e motor intention of cation performance. +us, the oﬄine motor intention these 3 subjects was mainly expressed on SM1. Conse- classiﬁcation accuracy was calculated and taken into quently, classiﬁcation accuracy analyzed by the in- researching along with ERD topography and clinical as- dividualized channel selection strategy was similar with that sessment. Table 2 shows the oﬄine classiﬁcation accuracy analyzed by classical channel selection strategy. Even so, generated by diﬀerent channel-selection strategies. +e ﬁrst individualized channel selection can detect the principal Time (s) Time (s) 8 Journal of Healthcare Engineering Time-frequency analysis ERD in C4 1.05 0.95 0.9 0.85 18 5 0.8 14 0.75 1 234567 1 234567 Time (s) Time (s) (a) Time-frequency analysis ERD in C3 1.05 9 1 8 0.95 0.9 20 0.85 18 4 0.8 0.75 1 23456 1 234567 Time (s) Time (s) (b) Figure5:Time-frequencygraphandERDtopographyduringmotorattemptforS2.(a)Unaﬀectedband(left)movement:Duringtheﬁrstto third second, the subject was in resting state. During the fourth to seventh second, the subject was conducting the required task. During the motor attempt of the unaﬀected left hand, apparent decline was identiﬁed in power spectral density on contralateral C4 compared with the resting state, based on Morlet Wavelet time-frequency analysis. (b) Aﬀected band (right) movement: During the ﬁrst to third second, the subject was in resting state. From the fourth to seventh second, the subject was required to conduct the motor attempt task. During the motorattemptofthepareticrighthand,no apparentdeclinewasfoundin powerspectral density oncontralateral C3comparedwithresting state, based on Morlet Wavelet time-frequency analysis. Correspondingly, no ERD was shown in the ERD ratio-map. channels representing ERD more accurately which results in the features could be extracted more eﬀectively, which slightly higher accuracy. However, for S5, ERD was detected resulted in higher accuracy (7.5% increase) using ERD se- limitedlyonC3,CP1,FC3,andC1withoutawiderange,and lectioninspiteofoverlapping2channelsinaﬀectedSM1.(v) Frequency (Hz) Frequency (Hz) Frequency (Hz) Frequency (Hz) Power spectral density Power spectral density (µV /Hz·ms) (µV /Hz·ms) Power spectral density Power spectral density (µV /Hz·ms) (µV /Hz·ms) Journal of Healthcare Engineering 9 Table 2: Channel-selection strategies and classiﬁcation performance. Group Subject SM1-4 selection Accuracy SM1-5 selection Accuracy ERD selection Accuracy 4 None None None None None None ERD-blind group 11 None None None None None None 14 None None None None None None Average±SD None None None 2 None None None None None None ERD-disappea-rance group 15 None None None None None None Average±SD None None None 1 Cz C1 C3 C5 82.5% C1 C3 C5 FC3 CP3 82.5% CP1 Pz C1 TP8 92.5% 3 Cz C1 C3 C5 67.5% C1 C3 C5 FC3 CP3 67.5% FC5 C4 C5 C1 80.0% ERD-proliferation group 6 Cz C1 C3 C5 60.0% C1 C3 C5 FC3 CP3 62.5% C4 CP2 AF8 C2 67.5% 9 Cz C1 C3 C5 70.0% C1 C3 C5 FC3 CP3 72.5% C4 CP2 CP6 CP4 77.5% Average±SD 70.00±9.35% 71.25±8.54% 79.38±10.28% 5 Cz C1 C3 C5 67.5% C1 C3 C5 FC3 CP3 75.0% C3 CP1 FC3 C1 75.0% 7 Cz C1 C3 C5 72.5% C1 C3 C5 FC3 CP3 72.5% C1 C2 C3 C4 75.0% ERD-SM1 group 10 Cz C2 C4 C6 65.0% C2 C4 C6 FC4 CP4 67.5% C4 AF3 F5 CP4 70.0% 12 Cz C1 C3 C5 65.0% C1 C3 C5 FC3 CP3 65.0% C3 FC3 C1 CP3 67.5% Average±SD 67.50±3.54% 70.00±4.56% 71.88±3.75% 8 Cz C1 C3 C5 85.0% C1 C3 C5 FC3 CP3 85.0% F7 FC2 C4 F5 87.5% Others 13 Cz C1 C3 C5 60.0% C1 C3 C5 FC3 CP3 62.5% Cz CP1 C1 C2 65.0% Average±SD 72.50±17.68% 73.75±15.91% 76.25±15.91% Overall average±SD 69.50±8.48% 71.25±7.84% 75.75±8.98% SM1-4 selection, 4 channels selected from the sensorimotor cortex. SM1-5 selection, 5 channels selected from the sensorimotor cortex. ERD selection, 4 channels selected with the lowest ERD ratio. SD, standard deviation. 10 Journal of Healthcare Engineering or from the channels showing ERD. However, no deﬁnite Others: ERD was detected in a narrow area consisting of one or two EEG channels, which means the principal feature of relationship between lesion type and ERD expression can be concluded. As mentioned above, in this study, a heteroge- motor intention was expressed in a limited area. +e motor intention features can be detected by ERD selection but with neous group of patients was recruited with the level of se- limited improvement on classiﬁcation accuracy. +us, the verity ranging from moderate to very severe and with classiﬁcation accuracy using ERD selection was 2.5% higher diﬀerent lesion sites. It is diﬃcult to ﬁnd a correlation be- for S8 and 5% higher for S13 than using SM1 channel-se- tween the patient’s ERD topography and FMA, since FMA lection strategy. +e ERD channel-selection strategy was varied from subject to subject even in the same group due to based on the motor intention features which varied from variousfactors,includingTFO,lesionsite,lesiontype,etc.In patient topatient, while classical channel-selection strategies further research, more patients with the same lesion type werebasedonthemotorcontrolforhealthysubjects[13,14]. should be selected, and the quantitative correlation between the ERD topography and these inﬂuencing factors can be +erefore, the comparison result between SM1-5 channel selection strategy and ERD selection was similar except for investigated. Furthermore, the actual rehabilitation eﬀect using EEG-based BCI with the proposed individualized S5. For S5, SM1-5 and ERD selection generated equal classiﬁcation accuracy since the analyzed EEG channels channel-selection strategy also needs to be validated in the overlapped for these two strategies. future. +e results indicate that motor intention expression is not limited within SM1 for stroke patients. Depending on 5. Conclusion the severity of the deﬁcits in the aﬀected motor cortex, two +is study analyzed ERD topography in stroke patients main patterns of cortical reorganization have been identiﬁed during the motor attempt of the paretic hand and compared in a longitudinal study; for stroke patients whose aﬀected the classiﬁcation performance using diﬀerent channel-se- motor cortex is extensively damaged, persistent recruitment lection strategies collecting EEG from diﬀerent channels. of additional ipsilateral and contralateral brain regions was Results showed that the classiﬁcation accuracy analyzing the found during movement of the paretic hand . +erefore, channels showing ERD is higher than that analyzing EEG the compensatory region(s) should be taken into consid- fromSM1,which,inaddition,maynotbeabletorecoverthe eration for BCI implementation for patients whose aﬀected original motor control ability in stroke patients with severe motor ability has little potential to recover on the original damages.Webelievetheﬁndingscanexplainthereasonwhy site. +e altered pattern of activated brain regions after the accuracy of classiﬁcation is rather low for some stroke stroke may have the potential to predict the regions rep- patients so that they cannot be recruited in the BCI training. resenting motor functions after neural reorganization. +emainhypothesisbehindstrokerehabilitationwithBCIis +erefore, the loop between the altered activated brain re- that closing the loop between motor intention and actual gions and the paretic hand can be stimulated using in- movement can restore functional corticospinal and corti- dividualized channel-selection strategy. comuscular connections. +e results in this study indicated On the contrary, for the patients whose aﬀected motor that the closed loop should be individualized according to cortex is damaged within limited extent, another pattern of patient’s deﬁcits and condition to achieve a better re- cortical reorganization appears, in which, after initial re- habilitation outcome. +e actual rehabilitation eﬀects of cruitment of additional ipsilateral and contralateral brain larger amount of patients using EEG-based BCI with this regions, brain activation during movement of the paretic channel-selection strategy need to be validated in the future. hand gradually develops toward a pattern of activation re- stricted to the contralateral sensorimotor cortex; however, Data Availability this trend of focusing does not imply recovery . It means although the original brain region is still capable of repre- +e EEG data used to support the ﬁndings of this study are senting motor intention, the neural circuit between the available from the corresponding author upon request. cortex and paretic side of the body may be damaged. +erefore, for the patients whose neural reorganization Conflicts of Interest evolves during the whole rehabilitation process, channel selection in EEG-based BCI rehabilitation can be updated as +e authors disclose that there are no actual or potential the expression of motor intention focuses to the contra- conﬂicts of interest including any ﬁnancial, personal, or lateral sensorimotor cortex. With an individualized channel other relationships with other people or organizations that selection strategy, BCI rehabilitation training can respond could inappropriately inﬂuence (bias) their work. more accurately to the patient’s motor intention in order to stimulate the neural circuit, which is the principle of BCI- Authors’ Contributions based rehabilitation. Chong Li and Tianyu Jia contributed equally to this work. 4.3. Limitations. +e objective of this paper was to in- Acknowledgments vestigate ERD topography of stroke patients in diﬀerent conditions, and compare the classiﬁcation accuracy of the +e authors are grateful to the patients who participated in motorattemptoftheparetichandusingEEGdatafromSM1 this study and the therapists who helped to conduct the Journal of Healthcare Engineering 11 Cybernetics, Part C (Applications and Reviews), vol. 42, no. 6, experiment. +is work was supported by the National pp. 1169–1179, 2012. Natural Science Foundation of China (Grant 51805288 and  Y. Bergamasco, L.Spyrou, D. +ijssenet al., “CombinedEEG- U1613207) and China Postdoctoral Science Foundation fNIRS decoding of motor attempt and imagery for brain (Grant 2017M620762). switch control: an oﬄine study in patients with tetraplegia,” IEEE Transactions on Neural Systems and Rehabilitation References Engineering, vol. 22, no. 2, pp. 222–229, 2014.  U. Chaudhary, N. Birbaumer, and A. Ramos-Murguialday,  K. Strong, C. Mathers, and R. Bonita, “Preventing stroke: “Brain-computer interfaces for communication and re- saving lives around the world,” e Lancet Neurology, vol. 6, habilitation,” Nature Reviews Neurology, vol. 12, no. 9, no. 2, pp. 182–187, 2007. pp. 513–525, 2016.  J. Carr and R. Shepherd, “Clinical physiotherapy specialisa-  G. Di Pino, G. Pellegrino, G. Assenza et al., “Modulation of tion in Australia: some current views,” Australian Journal of brain plasticity in stroke: a novel model for neuro- Physiotherapy, vol. 42, no. 1, pp. 9–14, 1996. rehabilitation,” Nature Reviews Neurology, vol. 10, no. 10,  H. M. Feys, W. J. De Weerdt, B. E. Selz et al., “Eﬀect of a pp. 597–608, 2014. therapeutic intervention for the hemiplegic upper limb in the  P. M. Rossini, C. Caltagirone, A. Castriota-Scanderbeg et al., acute phase after stroke,” Stroke, vol. 29, no. 4, pp. 785–792, “Hand motor cortical area reorganization in stroke,” Neu- roReport, vol. 9, no. 9, pp. 2141–2146, 1998.  K. K. Ang, Q. Zhang, D. I. Rosenthal et al., “Randomized  R.Pineiro,S.Pendlebury,H.Johansen-Berg,andP.M.Matthews, phase III trial of concurrent accelerated radiation plus cis- “Functional MRI detects posterior shifts in primary sensorimotor platin with or without cetuximab for stage III to IV head and cortex activation after stroke,” Stroke, vol. 32, no. 5, pp. 1134– neck carcinoma: RTOG 0522,” Journal of Clinical Oncology, 1139, 2001. vol. 32, no. 27, pp. 2940–2950, 2014.  A.Feydy,R.Carlier,A.Roby-Bramietal.,“Longitudinalstudy  T. Ono, K. Shindo, K. Kawashima et al., “Brain-computer of motor recovery after stroke,” Stroke, vol. 33, no. 6, interface with somatosensory feedback improves functional pp. 1610–1617, 2002. recovery from severe hemiplegia due to chronic stroke,”  E. Lopez-Larraz, J. M. Antelis, L. Montesano, A. Gil-Agudo, Frontiers in Neuroengineering, vol. 7, no. 19, 2014. and J. Minguez, “Continuous decoding of motor attempt and  N.N.Johnson,J.Carey,B.J.Edelmanetal.,“CombinedrTMS motor imagery from EEG activity in spinal cord injury pa- and virtual reality brain-computer interface training for tients,” in Proceedings of the 2012 Annual International motor recovery after stroke,” Journal of Neural Engineering, Conference of the IEEE Engineering in Medicine and Biology vol. 15, no. 1, Article ID 016009, 2018. Society, pp. 1798–1801, San Diego, CA, USA, August 2012.  A. Ramos-Murguialday, D. Broetz, M. Rea et al., “Brain-  W. Roseboom, D. Linares, and S. Y. Nishida, “Sensory ad- machine interface in chronic stroke rehabilitation: a con- aptation for timing perception,” Proceedings of the Royal trolledstudy,” Annals of Neurology,vol.74,no.1,pp.100–108, Society B: Biological Sciences, vol. 282, no. 1805, Article ID 20142833, 2015.  K. Wang, Z. Wang, Y. Guo et al., “A brain-computer interface  O. Bertrand, F. Perrin, and J. Pernier, “A theoretical justiﬁ- driven by imagining diﬀerent force loads on a single hand: an cation of the average reference in topographic evoked po- online feasibility study,” Journal of Neuroengineering and tential studies,” Electroencephalography and Clinical Rehabilitation, vol. 14, no. 1, 2017. Neurophysiology/Evoked Potentials Section, vol. 62, no. 6,  W.Ming,S. Qiu, K. Wang et al., “Enhancing performance of a pp. 462–464, 1985. motor imagery based brain-computer interface by in-  A.Delorme andS.Makeig, “EEGLAB:anopensourcetoolbox corporating electrical stimulation-induced SSSEP,” Journal of for analysis of single-trial EEG dynamics including in- Neural Engineering, vol. 14, no. 2, Article ID 026002, 2017. dependent component analysis,” Journal of Neuroscience  Y. Bian, H. Qi, L. Zhao, D. Ming, T. Guo, and X. Fu, “Im- Methods, vol. 134, no. 1, pp. 9–21, 2004. provements in event-related desynchronization and classiﬁ-  M. Mørup, L. K. Hansen, J. Parnas, and S. M. Arnfred, cation performance of motor imagery using instructive Decomposing the Time-Frequency Representation of EEG dynamic guidance and complex tasks,” Computers in Biology Using Non-Negative Matrix and Multi-Way Factorization, and Medicine, vol. 96, pp. 266–273, 2018. Technical University of Denmark, Lyngby, Denmark, 2006.  G. Pfurtscheller and A. Berghold, “Patterns of cortical acti-  T. Ono, Y. Tomita, M. Inose et al., “Multimodal sensory vation during planning of voluntary movement,” Electroen- feedback associated with motor attempts alters BOLD re- cephalography and Clinical Neurophysiology, vol. 72, no. 3, sponses to paralyzed hand movement in chronic stroke pa- pp. 250–258, 1989. tients,” Brain Topography, vol. 28, no. 2, pp. 340–351, 2015.  S. Shahid, R. K. Sinha, and G. Prasad, “Mu and beta rhythm  G. Pfurtscheller and F. H. Lopes da Silva, “Event-related EEG/ modulations in motor imagery related post-stroke EEG: a MEG synchronization and desynchronization: basic princi- study under BCI framework for post-stroke rehabilitation,” ples,” Clinical Neurophysiology, vol.110, no.11, pp.1842–1857, BMC Neuroscience, vol. 11, no. 1, p. 127, 2010.  W.K.Tam,K.Y.Tong,F.Meng,andS.Gao,“Aminimalsetof  L. Zhang, C. Zhang, H. Higashi, J. Cao, and T. Tanaka, electrodes for motor imagery BCI to control an assistive device in chronic stroke subjects: a multi-session study,” IEEE “Common spatial pattern using multivariate EMD for EEG classiﬁcation,” in Proceedings of the 2011 Asia-Paciﬁc Signal Transactions on Neural Systems and Rehabilitation Engi- neering, vol. 19, no. 6, pp. 617–627, 2011. and Information Processing Association Annual Summit and  A. Frisoli, C. Loconsole, D. Leonardis et al., “A new gaze-BCI- Conference, pp. 244–248, Xian, China, November 2011.  H. Ramoser, J. Muller-Gerking, and G. Pfurtscheller, “Opti- driven control of an upper limb exoskeleton for rehabilitation in real-world tasks,” IEEE Transactions on Systems, Man, and mal spatial ﬁltering of single trial EEG during imagined hand 12 Journal of Healthcare Engineering movement,” IEEE Transactions on Rehabilitation Engineering, vol. 8, no. 4, pp. 441–446, 2000.  D. J. Krusienski, E. W. Sellers, F. Cabestaing et al., “A comparison of classiﬁcation techniques for the P300 speller,” Journal of Neural Engineering, vol. 3, no. 4, pp.299–305, 2006.  Y. Cao, X. An, Y. Ke et al., “+e eﬀects of semantic con- gruency: a research of audiovisual P300-speller,” Biomedical Engineering Online, vol. 16, no. 1, 2017.  B. Z. Allison and C. Neuper, “Could anyone use a BCI?,” in Brain-Computer Interfaces: Applying Our Minds to Human- Computer Interaction, D. S. Tan and A. Nijholt, Eds., pp. 35–54, Springer, London, UK, 2010.  K. K. Andersen, T. S. Olsen, C. Dehlendorﬀ, and L. P. Kammersgaard, “Hemorrhagic and ischemic strokes compared,” Stroke, vol. 40, no. 6, pp. 2068–2072, 2009.  D. Bullock and J. Contreras-Vidal, “How spinal neural net- works reduce discrepancies between motor intention and motor realization,” in Variability and Motor Control, K. Newell and D. Corcos, Eds., pp.183–221, Human Kinetics Press, Champaign, IL, USA, 1993. International Journal of Advances in Rotating Machinery Multimedia Journal of The Scientific Journal of Engineering World Journal Sensors Hindawi Hindawi Publishing Corporation Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 http://www www.hindawi.com .hindawi.com V Volume 2018 olume 2013 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Journal of Control Science and Engineering Advances in Civil Engineering Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 Submit your manuscripts at www.hindawi.com Journal of Journal of Electrical and Computer Robotics Engineering Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 VLSI Design Advances in OptoElectronics International Journal of Modelling & Aerospace International Journal of Simulation Navigation and in Engineering Engineering Observation Hindawi Hindawi Hindawi Hindawi Volume 2018 Volume 2018 Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com www.hindawi.com www.hindawi.com Volume 2018 International Journal of Active and Passive International Journal of Antennas and Advances in Chemical Engineering Propagation Electronic Components Shock and Vibration Acoustics and Vibration Hindawi Hindawi Hindawi Hindawi Hindawi www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018 www.hindawi.com Volume 2018
Journal of Healthcare Engineering – Hindawi Publishing Corporation
Published: Aug 28, 2019
Access the full text.
Sign up today, get DeepDyve free for 14 days.