{"title": "Impact of an Energy Normalization Transform on the Performance of the LF-ASD Brain Computer Interface", "book": "Advances in Neural Information Processing Systems", "page_first": 725, "page_last": 732, "abstract": "", "full_text": "Impact of an Energy Normalization \nTransform on the Performance of the \nLF-ASD Brain Computer Interface \n\n \n \n\n \n\nZhou Yu1 \n\nSteven G. Mason2 \n\nGary E. Birch1,2 \n\n1  Dept. of Electrical and Computer Engineering \n\nUniversity of British Columbia \n2356 Main Mall \nVancouver, B.C. Canada V6T 1Z4 \n\n2  Neil Squire Foundation \n\n220-2250 Boundary Road \nBurnaby, B.C. Canada V5M  3Z3 \n\nAbstract \n\nThis  paper  presents  an  energy  normalization  transform  as  a \nmethod to reduce system errors in the LF-ASD brain-computer \ninterface.  The energy normalization transform has two major \nbenefits to the system performance. First, it can increase class \nseparation  between  the  active  and  idle  EEG  data.    Second,  it \ncan  desensitize  the  system  to  the  signal  amplitude  variability. \nFor  four  subjects  in  the  study,  the  benefits  resulted  in  the \nperformance  improvement  of  the  LF-ASD  in  the  range  from \n7.7% to 18.9%, while for the fifth subject, who had the highest \nnon-normalized  accuracy  of  90.5%,  the  performance  did  not \nchange notably with normalization. \n\n1  Introduction \n\nto  develop  a  direct  Brain-Computer  Interface  (BCI). \n\nIn  an  effort  to  provide  alternative  communication  channels  for  people  who  suffer \nfrom  severe  loss  of  motor  function,  several  researchers  have worked  over  the past \ntwo  decades \n  Since \nelectroencephalographic (EEG) signal has good time resolution and is non-invasive, \nit is commonly used for data source of a BCI. A BCI system converts the input EEG \ninto  control  signals,  which  are  then  used  to  control  devices  like  computers, \nenvironmental control system and neuro-prostheses. \nMason  and  Birch  [1]  proposed  the  Low-Frequency  Asynchronous  Switch  Design \n(LF-ASD) as a BCI which detected imagined voluntary movement-related potentials \n(IVMRPs) in spontaneous EEG. The principle signal processing components of the \nLF-ASD are shown in Figure 1. \n\n\fLPF \n\n                   \nsIN                  sLPF                                sFE                                   sFC\n \n\n \nFigure 1: The original LF-ASD design. \n\nFeature  \nClassifier \n\nFeature \nExtractor\n\n \n\n \n\n \n\n \n\n \n\n \n\n \n\nThe  input  to  the  low-pass  filter  (LPF),  denoted  as  SIN  in  Figure  1,  are  six  bipolar \nEEG  signals  recorded  from  F1-FC1,  Fz-FCz,  F2-FC2,  FC1-C1,  FCz-Cz  and  \nFC2-C2 sampled at 128 Hz. The cutoff frequency of the LPF implemented by Mason \nand Birch was 4 Hz. The Feature Extractor of the LF-ASD extracts custom features \nrelated  to  IVMRPs.  The  Feature  Classifier  implements  a  one-nearest-neighbor  (1-\nNN)  classifier,  which  determines  if  the  input  signals  are  related  to  a  user  state  of \nvoluntary  movement  or  passive  (idle)  observation.    The  LF-ASD  was  able  to \nachieve True Positive (TP) values in the range of 44%-81%, with the corresponding \nFalse Positive (FP) values around 1% [1]. \nAlthough  encouraging,  the  current  error  rates  of  the  LF-ASD  are  insufficient  for \nreal-world  applications.  This  paper  proposes  a  method  to  improve  the  system \nperformance. \n\n2  Design and Rationale \n\nThe  improved  design  of  the  LF-ASD  with  the  Energy  Normalization  Transform \n(ENT) is provided in Figure 2. \n\nSIN                             SN                        SNLPF        \n\nFeature \nExtractor\n\n     SNFE                SNFC \nFeature  \nClassifier\n\n \n\n  ENT \n\nLPF \n\n \n\nFigure 2:  The improved LF-ASD with the Energy Normalization Transform. \n\nThe design of the Feature Extractor and Feature Classifier were the same as shown \nin Figure 1. The Energy Normalization Transform (ENT) is implemented as \n\nS\n\nN\n\n(\n\nn\n\n)\n\n=\n\ns\n\n=\n\n(\n\ns\n\n\u2212=\n\nS\nS\n\nN\n\nw\n\u2211\nw\n\n(\n\nN\n\n\u2212\n\n2/)1\n\n\u2212\n\n2/)1\n\n(\n\nn\n\n)\n\n2\n\n(\n\nn\n\n\u2212\n\ns\n\n)\n\nw\n\nN\n\nIN\n\nIN\n\n \n\nwhere WN (normalization  window size) is  the only parameter in the equation. The \noptimal  parameter  value  was  obtained  by  exhaustive  search  for  the  best  class \nseparation  between  active  and  idle  EEG  data.  The  method  of  obtaining  the  active \nand idle EEG data is provided in Section 3.1. \nThe  idea  to  use  energy  normalization  to  improve  the  LF-ASD  design  was  based \nprimarily  on  an  observation  that  high  frequency  power  decreases  significantly \naround movement. For example, Jasper and Penfield [3] and Pfurtscheller et al, [4] \nreported EEG power decrease in the mu (8-12 Hz) and beta rhythm (18-26 Hz) when \npeople are involved in motor related activity. Also Mason [5] found that the power \nin  the  frequency  components  greater  than  4Hz  decreased  significantly  during \nmovement-related potential periods, while power in the frequency components less \nthan  4Hz  did  not.  Thus  energy  normalization,  which  would  increase  the  low \nfrequency  power  level,  would  strengthen  the  0-4  Hz  features  used  in  the  LF-ASD \nand hence reduce errors. In addition, as a side benefit, it can automatically adjust the \nmean scale of the input signal and desensitize the system to change in EEG power, \nwhich is known to vary over time [2]. Therefore, it was postulated that the addition \nof  ENT  into  the  improved  design  would  have  two  major  benefits.  First,  it  can \n\n\f \n\nincrease the EEG power around motor potentials, consequently increasing the class \nseparation and feature strength. Second, it can desensitize the system to amplitude \nvariance of the input signal. \nIn  addition,  since  the  system  components  of  the  modified  LF-ASD  after  the  ENT \nwere  the  same  as  in  the  original  design,  a  major  concern  was  whether  or  not  the \nENT distorted the features used by the LF-ASD. Since the features used by the LF-\nASD are generated from the 0-4 Hz band, if the ENT does not distort the phase and \nmagnitude spectrum in this specific band, it would not distort the features related to \nmovement potential detection in the application. \n\n3  Evaluation \n\n3.1  Test data \n\nTwo  types  of  EEG  data  were  pre-recorded  from  five  able-bodied  individuals  as \nshown in Figure 3.  Active Data Type and Idle Data Type. Active Data was recorded \nduring  repeated  right  index  finger  flexions  alternating  with  periods  of  no  motor \nactivity; Idle Data was recorded during extended periods of passive observation. \n\n \n\nFigure 3: Data Definition of M1, M2, Idle1 and Idle2. \n\nObservation windows centered at the time of the finger switch activations (as shown \nin Figure 4) were imposed in the active data to separate data related to movements \nfrom data during periods of idleness. For purpose of this study, data in the front part \nof  the  observation  window  was  defined  as  M1  and  data  in  the  rear  part  of  the \nwindow was defined as M2. Data falling out of the observation window was defined \nas  Idle2.    All  the  data  in  the  Idle  Data  Type  was  defined  as  Idle1  for  comparison \nwith Idle2. \n\nFigure 4:  Ensemble Average of EEG centered on finger activations. \n\n \n\n \n\n\f \n\n \nFigure 5:  Density distribution of Idle1, Idle2, M1 and M2. \n\nIt  was  noted,  in  terms  of  the  density  distribution  of  active  and  idle  data,  the \nseparation  between  M2  and  Idle2  was  the  largest  and  Idle1  and  Idle2  were  nearly \nidentical (see Figure 5). For the study, M2  and Idle2 were chosen to represent the \nactive  and  idle  data  classes  and  the  separation  between  M2  and  Idle2  data  was \ndefined by the difference of means (DOM) scaled by the amplitude range of Idle2. \n\n3.2  Optimal parameter determination \n\nThe  optimal  combination  of  normalization  window  size,  WN,  and  observation \nwindow size, WO was selected to be that which achieved the maximal DOM value. \nThis  was \nin  \nSection 4.1. \n\ndetermined \n\nexhaustive \n\nby \n\nsearch, \n\nand \n\ndiscussed \n\n3.3  Effect of ENT on the Low Pass Filter output \n\nAs  mentioned  previously,  it  was  postulated  that  the  ENT  had  two  major  impacts: \nincreasing  the  class  separation  between  active  and  idle  EEG  and  desensitizing  the \nsystem  to  the  signal  amplitude  variance.  The  hypothesis  was  evaluated  by \ncomparing  characteristics  of  SNLPF  and  SLPF  in  Figure  1  and  Figure  2.  DOM  was \napplied to measure the increased class separation. The signal with the larger DOM \nmeant larger class separation. In addition, the signal with smaller standard deviation \nmay result in a more stable feature set. \n\n3.4  Effect of ENT on the LF-ASD output \n\nThe  performances  of  the  original  and  improved  designs  were  evaluated  by \ncomparing  the  signal  characteristics  of  SNFC  in  Figure  2  to  SFC  in  Figure  1.  A \nReceiver  Operating  Characteristic  Curve  (ROC  Curve)  [6]  was  generated  for  the \noriginal  and \nthe  system \nperformance  over  a  range  of  TP  vs.  FP  values.  The  larger  area  under  ROC  Curve \nindicates better system performance. In real applications, a BCI with high-level FP \nrates could cause frustration for subjects. Therefore, in this work only the LF-ASD \nperformance when the FP values are less than 1% were studied. \n\nimproved  designs.  The  ROC  Curve  characterizes \n\n4  Results \n\n4.1  Optimal normalization window size (W N) \nThe  method  to  choose  optimal  WN  was  an  exhaustive  search  for  maximal  DOM \nbetween  active  and  idle  classes.  This  method  was  possibly  dependent  on  the \nobservation  window  size  (WO).  However,  as  shown  in  Figure  6a,  the  optimal  WN \nwas found to be independent of WO. Experimentally, the WO values were selected in \nthe  range  of  50-60  samples,  which  corresponded  to  largest  DOM  between  non-\nnormalized active and idle data. The optimal WN was obtained by exhaustive search \n\n\ffor  the  largest  DOM  through  normalized  active  and  idle  data.  The  DOM  vs.  WN \nprofile for Subject 1 is shown in Figure 6b. \n\n \n\na)    \n\n \n\n \n\nb) \n\n \n\nFigure 6:  Optimal parameter determination for Subject 1 in Channel 1 \n\na) DOM vs. WO;  b) DOM vs. WN. \n\nWhen using ENT, a small WN value may cause distortion to the feature set used by \nthe LF-ASD. Thus, the optimal WN was not selected in this range (< 40 samples). \nWhen  WN  is  greater  than  200,  the  ENT  has  lost  its  capability  to  increase  class \nseparation  and  the DOM  curve  gradually  goes  towards  the best  separation without \nnormalization. Thus, the optimal WN should correspond to the maximal DOM value \nwhen WN is in the range from 40 to 200. In Figure 6b, the optimal WN is around 51.  \n\n4.2  Effect of ENT on the Low Pass Filter output  \n\nWith ENT, the standard deviation of the low frequency EEG signal decreased from \naround  1.90  to  1.30  over  the  six  channels  and  over  the  five  subjects.  This  change \nresulted in more stable feature sets. Thus, the ENT desensitizes the system to input \nsignal variance. \n\na)           \n\n \n\n  b) \n\nFigure 7:  Density distribution of the active vs. idle class without   \n\n(a) and with (b) ENT, for Subject 1 in Channel 1. \n\n \n\nAs shown in Figure 7, by increasing the EEG power around motor potentials, ENT \ncan  increase  class  separations  between  active  and  idle  EEG  data.  The  class \nseparation  in (frontal)  Channels  1-3  across  all  subjects  increased  consistently  with \nthe proposed ENT. The same was  true for (midline)  Channels 4-6, for all subjects \nexcept  Subject  5,  whose  DOM  in  channel  5-6  decreased  by  2.3%  and  3.4% \nrespectively with normalization. That is consistent with the fact that his EEG power \nin  Channels  4-6  does  not  decrease.  On  average,  across  all  five  subjects,  DOM \nincreases  with  normalization  to  about  28.8%,  26.4%,  39.4%,  20.5%,  17.8%  and \n22.5% over six channels respectively. \n\nIn addition, the magnitude and phase spectrums of the EEG signal before and after \nENT is provided in Figure 8. The ENT has no visible distortion to the signal in the \nlow  frequency  band  (0-4  Hz)  used  by  the  LF-ASD.  Therefore,  the  ENT  does  not \ndistort the features used by the LF-ASD. \n\n\f \n\n(a) \n\n(b) \n\nFigure 8:  Magnitude and phase spectrum of the EEG signal before and after ENT. \n\n4.3  Effect of ENT on the LF-ASD output  \n\nThe  two  major  benefits  of  the  ENT  to  the  low  frequency  EEG  data  result  in  the \nperformance  improvement  of  the  LF-ASD.    Subject  1\u2019s  ROC  Curves  with  and \nwithout  ENT  is  shown  in  Figure  9,  where  the  ROC-Curve  with  ENT  of  optimal \nparameter  value  is  above  the  ROC  Curve  without  ENT.  This  indicates  that  the \nimproved LF-ASD performs better. Table I compares the system performance with \nand  without  ENT  in  terms  of  TP  with  corresponding  FP  at  1%  across  all  the  5 \nsubjects. \n\nFigure 9:  The ROC Curves (in the section of interest) of Subject 1 with different \n\nWN values and the corresponding ROC Curve without ENT. \n\n \n\n \n \n \n \n \n \n \n \n\n\fTable I:  Performance of the LF-ASD with and without LF-ASD in terms of \n\n the True Positive rate with corresponding False Positive at 1%. \n\n \n\nTP without \n\nTP with \n\n \nSubject 1 \nSubject 2 \nSubject 3 \nSubject 4 \nSubject 5 \n\nENT \n66.1% \n82.7% \n79.7% \n79.3% \n90.5% \n\nENT \n85.0% \n90.4% \n88.0% \n87.8% \n88.7% \n\nPerformance \nImprovement \n\n18.9% \n7.7% \n8.3% \n8.5% \n-1.8% \n\n \nFor 4 out of 5 subjects, corresponding with the FP at 1%, the improved system with \nENT increased the TP value by 7.7%, 8.3%, 8.5% and 18.9% respectively. Thus, for \nthese subjects, the range of TP with FP at 1% was improved from 66.1%-82.7% to \n85.0%-90.4% with ENT. For the fifth subject, who had the highest non-normalized \naccuracy  of 90.5%,  the performance  remained  around 90% with  ENT.  In  addition, \nthis evaluation is conservative.  Since the codebook in the Feature Classifier and the \nparameters  in  the  Feature  Extractor  of  the  LF-ASD  were  derived  from  non-\nnormalized EEG, they work in favor of the non-normalized EEG. Therefore, if the \nparameters  and  the  codebook  of  the  modified  LF-ASD  are  generated  from  the \nnormalized EEG in the future, the modified LF-ASD may show better performance \nthan this evaluation. \n\n5  Conclusion \n\nThe evaluation with data from five able-bodied subjects indicates that the proposed \nsystem  with  Energy  Normalization  Transform  (ENT)  has  better  performance  than \nthe  original.  This  study  has  verified  the  original  hypotheses  that  the  improved \ndesign  with  ENT  might  have  two  major  benefits:  increased  the  class  separation \nbetween  active  and  idle  EEG  and  desensitized  the  system  performance  to  input \namplitude  variance.  As  a  side  benefit,  the  ENT  can  also  make  the  design  less \nsensitive to the mean input scale. \n\nIn  the  broad  band,  the  Energy  Normalization  Transform  is  a  non-linear  transform. \nHowever, it has no visible distortion to the signal in the 0-4 Hz band. Therefore, it \ndoes not distort the features used by the LF-ASD. \n\nFor  4  out  of  5  subjects,  with  the  corresponding  False  Positive  rate  at  1%,  the \nproposed  transform  increased  the  system  performance  by  7.7%,  8.3%,  8.5%  and \n18.9% respectively in terms of True Positive rate. Thus, the overall performance of \nthe  LF-ASD  for  these  subjects  was  improved  from  66.1%-82.7%  to  85.0%-90.4%. \nFor  the  fifth  subject,  who  had  the  highest  non-normalized  accuracy  of  90.5%,  the \nperformance  did  not  change  notably  with  normalization.  In  the  future  with  the \ncodebook  derived  from  the  normalized  data,  the  performance  could  be  further \nimproved.  \n\nReferences \n[1] Mason,  S.  G.  and  Birch,  G.  E.,  (2000)  A  Brain-Controlled  Switch  for  Asynchronous \nControl Applications. IEEE Trans Biomed Eng, 47(10):1297-1307. \n[2] Vaughan,  T.  M.,  Wolpaw,  J.  R.,  and  Donchin,  E.  (1996)  EEG-Based  Communication: \nProspects and Problems.  IEEE Trans Reh Eng, 4(4):425-430. \n\n\f \n\n[3] Jasper,  H.  and  Penfield,  W.  (1949)  Electrocortiograms  in  man:  Effect  of  voluntary \nmovement  upon  the  electrical  activity  of  the  precentral  gyrus.    Arch.Psychiat.Nervenkr., \n183:163-174. \n[4] Pfurtscheller,  G.,  Neuper,  C.,  and  Flotzinger,  D.  (1997)  EEG-based  discrimination \nbetween imagination of right and left hand movement.  Electroencephalography and Clinical \nNeurophysiology, 103:642-651. \n[5] Mason,  S.  G.  (1997)  Detection  of  single  trial  index  finger  flexions  from  continuous, \nspatiotemporal EEG.  PhD Thesis, UBC, January. \n[6] Green, D. M. and Swets, J. A. (1996) Signal Detection Theory and Psychophysics New York: \nJohn Wiley and Sons, Inc. \n\n\f", "award": [], "sourceid": 2422, "authors": [{"given_name": "Yu", "family_name": "Zhou", "institution": null}, {"given_name": "Steven", "family_name": "Mason", "institution": null}, {"given_name": "Gary", "family_name": "Birch", "institution": null}]}