{"title": "Using Feedforward Neural Networks to Monitor Alertness from Changes in EEG Correlation and Coherence", "book": "Advances in Neural Information Processing Systems", "page_first": 931, "page_last": 937, "abstract": null, "full_text": "Using Feedforward Neural Networks to \nMonitor  Alertness from  Changes in EEG \n\nCorrelation and  Coherence \n\nNaval  Health Research  Center,  P.O.  Box 85122 \n\nScott  Makeig \n\nSan Diego,  CA  92186-5122 \n\nTzyy-Ping Jung \n\nNaval  Health Research  Center  and \nComputational Neurobiology  Lab \nThe Salk Institute,  P.O.  Box 85800 \n\nSan Diego,  CA  92186-5800 \n\nHoward  Hughes  Medical  Institute and \n\nComputational Neurobiology Lab \nThe Salk Institute,  P.O.  Box 85800 \n\nTerrence J.  Sejnowski \n\nSan Diego,  CA  92186-5800 \n\nAbstract \n\nWe  report  here  that  changes  in  the  normalized  electroencephalo(cid:173)\ngraphic  (EEG)  cross-spectrum  can  be  used  in  conjunction  with \nfeedforward  neural networks to monitor changes in alertness of op(cid:173)\nerators  continuously  and  in  near-real  time.  Previously,  we  have \nshown  that EEG spectral  amplitudes covary with changes in alert(cid:173)\nness  as indexed by changes in behavioral error rate on an auditory \ndetection task [6,4].  Here, we report for the first time that increases \nin the frequency  of detection errors  in  this task are  also  accompa(cid:173)\nnied  by  patterns of increased  and  decreased  spectral  coherence  in \nseveral  frequency  bands  and  EEG  channel  pairs.  Relationships \nbetween  EEG  coherence  and  performance vary  between  subjects, \nbut within subjects,  their topographic and spectral profiles appear \nstable  from  session  to  session.  Changes  in  alertness  also  covary \nwith  changes  in  correlations  among  EEG  waveforms  recorded  at \ndifferent  scalp  sites,  and  neural  networks  can  also  estimate  alert(cid:173)\nness  from  correlation  changes  in  spontaneous  and  unobtrusively(cid:173)\nrecorded  EEG signals. \n\n1 \n\nIntroduction \n\nWhen humans become drowsy, EEG scalp recordings of potential oscillations change \ndramatically  in  frequency,  amplitude,  and  topographic  distribution  [3].  These \nchanges  are  complex  and  differ  between  subjects  [10].  Recently,  we  have  shown \n\n\f932 \n\nS. MAKEIG, T.-P.  JUNG, T.  J. SEJNOWSKl \n\nthat  using  principal  components  analysis  in  conjunction  with  feedforward  neural \nnetworks,  minute-scale  changes  in  performance  on  a  sustained  auditory  detection \ntask can  be estimated in near real-time from changes in the  EEG spectrum  at one \nor  more scalp  channels  [4,  6].  Here,  we  report,  first,  that loss  of alertness  during \nauditory detection task performance is  also accompanied by changes in spectral  co(cid:173)\nherence  of EEG  signals  recorded  at  different  scalp  sites.  The extent,  topography, \nand  frequency  content  of coherence  changes  linked  to  changes  in  alertness  differ \nbetween  subjects,  but  within  subjects  they  appear  stable  from  session  to  session. \nSecond,  since  most  coherence  changes  linked  to  alertness  are  not  associated  with \nsignificant  phase  differences,  moving correlation  measures  applied  to  wideband  or \nbandlimited  EEG  waveforms  also  covary  with  changes  in  alertness. \ning  coherence  and/or  correlation  information  into  neural  network  algorithms for \nestimating alertness from the EEG spectrum should enhance their accuracy  and ro(cid:173)\nbustness  and  contribute  to the  design  of practical neural  human-system interfaces \nperforming real-time monitoring of changes in operator  alertness. \n\nIncorporat(cid:173)\n\n2  Methods \n\nConcurrent EEG and behavioral data were collected for  the purpose of developing a \nmethod of objectively monitoring the alertness of operators of complex systems  [6] . \nTen adult volunteers  participated in  three or  more half-hour sessions  during which \nthey pushed one button whenever  they detected  an above-threshold auditory target \nstimulus (a brief increase in the level of the continuously-present background noise). \nTo maximize the chance of observing alertness decrements, sessions  were  conducted \nin a small, warm, and dimly-lit experimental chamber, and subjects were instructed \nto keep  their  eyes  closed. \n\nTargets were  350  ms increases  in the intensity of a  62  dB  white  noise  background, \n6 dB  above their threshold  of detectability, presented  at random time intervals at a \nmean rate of 10/min.  Short, and task-irrelevant probe tones oftwo frequencies  (568 \nand  1098  Hz)  were  interspersed  between  the  target  noise  bursts  at  2-4 s  intervals. \nEEG was collected from thirteen electrodes  located at sites of the Internation 10-20 \nSystem,  referred  to  the  right  mastoid,  at  a  sampling rate  of 312.5  Hz.  A  bipolar \ndiagonal electrooculogram (EOG) channel was also recorded for use in eye movement \nartifact correction and rejection.  Two sessions each from three of the subjects  were \nchosen for  analysis on  the basis of their  including more than 50  detection  lapses. \n\nA  continuous  performance  measure,  local error  rate, was  computed  by  convolving \nan  irregularly-spaced  performance  index  (hit=0/lapse=1)  with  a  95  s  smoothing \nwindow  advanced  through  the  performance data in  1.64 s  steps.  Target  hits  were \ndefined  as  targets  responded  to  within  a  100-3000  ms  window;  other  targets  were \ncalled  lapses.  After  eye  movement  artifacts  were  removed  from  the  data using  a \nselective  regression  procedure  [5],  and  data  containing  other  large  artifacts  were \nrejected  from  analysis,  complex EEG spectra  were  computed by  advancing  a  512-\npoint  (1.64  s)  data  window  through  the  data  in  0.41  s  steps,  multiplying  by  a \nHanning window, and converting to frequency  domain using an  FFT. \n\nComplex  coherence  was  then  computed  for  each  channel  pair  in  1.64  s  spectral \nepochs.  In  the  coherence  studies,  error  rate was  smoothed  with  a  bell-shaped  Pa(cid:173)\npoulis  window;  a  36  s  rectangular  window  was  used  to smooth the coherence  esti(cid:173)\nmates.  Finally, complex coherence was converted to coherence amplitude and phase \nand results  were  correlated with local error  rate.  A moving correlation measure be(cid:173)\ntween  (1-20  Hz)  bandlimited EEG waveforms was  computed for  each  channel  pair \nin  a  moving 1.64 s smoothing window,  and  then  smoothed using  a  causal  95-s  ex(cid:173)\nponential window.  The same window was used  to smooth the error rate time series \nfor  the correlation studies. \n\n\fUsing Feedforward Neural Networks to Monitor Alertness from  EEG \n\n933 \n\n3  Results \n\n1 \n\nu \n~ \n~ 0.8 \n<  0.6 \n~ \ne  0.4 \nl! \n0 u \n\n0.8 \n\n0.6 \n\nB :  0.4 \n~  0.2 \n\n0 \n\n0 \n\n\\9Chan\"I~~ \n\nFrequency: 9.1  Hz \n\n40 \n\n30 \n\n.. \nB 20 \n~  10 \n~  0 \n8 \ng \u00b710 \n~ \u00b720 \n\nTime on Task (min) \n\n(a) \n\n30 \n\nTime On Task (min) \n\n(b) \n\nFigure 1:  (a) Changes in coherence amplitude at 9.1 Hz  (upper traces) are correlated \nwith simultaneous changes in error rate during a half auditory detection task (lower \ntrace)  in  nine  indicated  central-frontal  channel  pairs.  (b)  Concurrent  changes  in \ncoherence  phase  at 15.25 Hz  (upper traces)  and local error rate (lowertrace) for  the \nsame session  and channel-pairs. \n\n3.1  Relation of Coherence Changes to Detection Performance. \n\nDuring the first  2-3  minutes of the session  shown  in  Fig.  la,  the  subject  detected \nall  targets  presented,  and  coherence  amplitudes  remained  high  (0.9).  In  minutes \n8-10,  however,  when  the  subject  failed  to  make  a  single  detection  response(lower \ntrace), coherence  amplitude fell to as low as 0.6.  Overall correlations for  this session \nbetween the coherence  and error rate time series  in these channel pairs ranged from \n-0.590 to -0.776. \n\nIn  the  same  session,  coherence  phase  at  15  Hz  also  covaried  with  performance \n(Fig.  1 b).  During  low-error  portions  of the  session,  there  was  no  detectable  co(cid:173)\nherence  phase  lag  at  15  Hz  within  the same  nine  channel  pairs,  whereas  while  the \nsubject  performed  poorly,  a  20  degree  phase  lag  appeared  during which  15  Hz  ac(cid:173)\ntivity at frontal sites  lead activity at frontal sites by  3 ms.  Overall  correlations for \nthis session  between  coherence  phase and error  rate for  these  channel pairs ranged \nfrom  0.416  to  0.689.  Correlations  between  coherence  amplitude  and  error  rate  at \n80 EEG frequencies  (Fig.  2a,  upper traces)  included two broad bands of strong neg(cid:173)\native  correlations  (3-12  Hz  and  15-20  Hz),  while  appreciable  correlations  between \ncoherence  phase and performance were  confined  to much narrower frequency  bands \n(lower  traces). \n\nTo estimate the significance ofthese coherence correlations, surrogate moving coher(cid:173)\nence  records  were  collected  10  times using  randomly-selected,  asynchronous blocks \nof contiguous EEG data for  each channeL  Correlations between the resulting surro(cid:173)\ngate moving coherence  time series  and error  rate were  computed, and the 99.936th \npercentile  of the  distribution  of (absolute)  correlations  was  determined.  For  the \nsubject  whose  data  is  shown  here,  this  value  was  0.485 .  Under  conservative  as(cid:173)\nsumptions  of complete  independence  of adjacent  frequencies,  this  should  give  the \n(p=0.05)  significance  level  for  the  maximum  absolute  correlation  in  each  80-bin \ncorrelation  spectrum.  (The  heuristic  estimate  of this  significance  level  from  the \nsurrogate  data  was  0.435).  In  the  two  sessions  from  this  subject,  however,  more \nthan  20%  of all  the  78  channel-pair  coherence  correlations  were  larger  in  absolute \nvalue  than  0.485,  implying  that  coherence  amplitude changes  at  many scalp  sites \nand frequencies  are  significantly related  to changes in  alertness  in this subject. \n\n\f934 \n\nS. MAKEIG, T.-P. JUNG, T.  J. SEJNOWSKI \n\n3.2  Spectral and Topographic Stability \n\nc \n0 \n\n.'\" \ni  ~ 9 C:annel pairs \n(30.5 \nc. \n. . .... .. ..\ne  0 \n< \n]-0.5 \n0 \n(,) \n\n..... .. .\u2022\u2022. ... . . \n\n.. . .. \n\n(,) \n\nc \n0 'i  o. \n1l \n0  o  . \n~ -0. \n...: \n1! \n0 \n(,) \n\n0  . \"  ..... \n\n-S \n.~ \n\n~  t \n] \nt \n~ \n~ \n\" \n\u00b7i \nil 0 \n\n\u00a7 \n\nSubject p37 \n25essions \n\n.... 1-~ir@t;B \n\u00ae2 \n~: \n~: \n\n(,) \n\nFrequency (Hz) \n\n(a) \n\n0 \n\n5 \n\n10 \n\n15 \n\n20 \n\n25 \n\n30 \n\nFrequency (Hz) \n\n(b) \n\nFigure 2:  (a)  Correlation spectra showing correlations  between  moving-average co(cid:173)\nherence  and error  rate for  the same session  and channel-pairs.  Small letters  'a,b,c' \nindicate the frequencies  analyzed  in  Figs.  1 and  3.  (b)  Cluster  analysis of correla(cid:173)\ntions  between  coherence  amplitude and  error  rate  at  41  frequencies  (0.6  Hz  to  25 \nHz).  Means  of six sets  of channel  pairs  derived  from  cluster  analysis  of 78  similar \ncoherence  correlation spectra from  all  pairs of 13  scalp  channels;  superimposed  on \nthe same means for  a  second  session from  the same subject. \n\nThe sign,  size,  and spectral  and  topographic structure  of correlations  between  co(cid:173)\nherence  amplitude and error  rate at each  frequency  were  stable across  two sessions \nfor  most channel  pairs and frequency  bands.  Fig.  2b shows  mean spectral  correla(cid:173)\ntions in both sessions from  the same subject for  six clusters  of similar channel-pair \ncorrelation spectra identified  by  cluster  analysis on  results  of the first  session.  Ex(cid:173)\ncept  near  5  Hz,  the  size  and  structure  of the  correlation  spectra  for  the  second \nsession  replicate  results  of  the  first  session.  The  spectral  stability  of monotonic \nrelationships between  EEG coherence  and auditory detection performance suggests \nthat  coherence  may be  used  to  predict  changes  in  performance level  from  sponta(cid:173)\nneous  EEG  data collected  continuously  and  unobtrusively  from  two  or  more scalp \nchannels. \n\n3.3  EEG  Waveform  Correlations and Performance \n\nIn  most  cases,  coherence  phase  lags  in  these  data  are  small,  and  correlations  be(cid:173)\ntween  changes  in  phase  lag  and  performance  were  insignificant.  We  therefore  in(cid:173)\nvestigated  whether  moving-average correlations between band-limited EEG signals \nin  different  scalp  channels  might also  be used  to predict  changes  in alertness,  pos(cid:173)\nsibly  at  a  lower  computational  cost,  by  studying  the  relationship  between  error \nrate  and  changes  in  moving-average  correlations  of time-domain  EEG  waveforms \n(1-20  Hz  bandpass)  in  the  same  6  sessions.  Again,  we  found  that  the  strength \nand  topographic  structure  of significant  relationships  between  moving-correlation \nand  performance  measures  are  stable  within,  and  variable between  subjects.  For \neach subject, we selected  8 EEG channel pairs whose moving-correlation time series \ncorrelated most highly with error rate,  and used  these  to train a  multilinear regres(cid:173)\nsion  network  and  three  feedforward  three-layer  perceptrons  to estimate error  rate \nfrom  moving-average correlations.  The feedforward  neural  networks  had  3,  4,  and \n5 hidden units, respectively.  Weights and biases of the network were  adjusted using \nthe  error  backpropagation  algorithm  [9].  Conjugate gradient  descent  was  used  to \nminimize the mean-squared error between network output and the actual error rate \n\n\fUsing Feedforward Neural Networks to Monitor Alertness from EEG \n\n935 \n\ntime series.  Cross-validation  [7]  was  used  to  prevent  the network  from  overfitting \nthe training data.  For  each  of the  6  training-testing session  pairs and each  neural \nnetwork  architecture,  the  time course  of error  rate was  estimated five  times  using \ndifferent  random initial weights  between -0.3 and 0.3.  We tested  the generalization \nability  of the  models  on  second  sessions  from  the  same  subjects.  The  procedure \nsimulated potential real-world alertness monitoring applications in which pilot data \nfor  each operator would be used  to  train a  network to estimate his or her  alertness \nin subsequent  sessions from unobtrusively-recorded  EEG  data. \n\nAccuracy of error rate estimation in the test sessions was almost identical for neural \nnetworks with 3,  4,  and 5 hidden  units.  Each  was more accurate  than multivariate \nlinear regression.  Figure 3 shows the time courses of actual and estimated error rate \nin one pair of training (top  paneQ  and test sessions.  Results for  two  other subjects \nwere  equivalent.  Table  1  shows  the  average  correlations  and  root-mean-squared \nestimation error  between  actual  and estimated error  rate time series for  6 sessions, \n2 each on 3 subjects using a feedforward neural network with 3 hidden units.  Results \nusing  4  or  5 hidden  units  are  equivalent.  Diagonal cells  show  results  for  training \nsessions,  off-diagonal  cells  for  test  sessions.  The  nonlinear  adaptability  of three(cid:173)\nlayer  perceptrons  give  improved  estimation  performance  over  multivariate linear \nregression,  reducing  the  RMS  estimation  error  in  the  test  sessions  from  0.255  to \n0.225  (F(l, 5) =  1234.29;p ~ 0.0001), and increasing the mean correlation between \nactual and estimated error  rate time series  from  0.63  to  0.67  (F(l, 5)  =  549.5;p ~ \n0.0001). \n\n4  Discussion \n\nSpectral  coherence  of EEG  waveforms  at  different  scalp  sites  has  been  measured \nfor  nearly 30  years  [11].  and is  the subject  of a  steadily  increasing  number of clin(cid:173)\nical,  behavioral,  and  developmental  EEG  studies.  Coherence  values  are  known  to \nbe higher  in  sleep  than  in  waking  [8],  and  wake-sleep  transitions  have been  noted \nto be  preceded  by  increased  coherence  at  some  frequencies  [2].  Our  results,  from \ndata on  three  subjects  performing  a  sustained  auditory  detection  task  under  so(cid:173)\nporific  conditions,  suggest  that  during  drowsiness,  coherence  may  either  increase \nor  decrease,  depending  on  the  subject,  analysis  frequency,  and  electrode  sites  an(cid:173)\nalyzed.  However,  in  individual subjects  the spectral  and  topographic  structure  of \nalertness-related  coherence  changes  appears stable from session to session. \n\nEEG  correlation  and  coherence  are intimately related:  changes  in  moving-average \ncorrelations  of EEG  waveforms  reflect  changes  in  broad-band,  zero-lag  coherence \nof activity  at  the  same sites.  The possibility  of using  moving-average  correlation \nmeasures  of electrophysiological  activity  to  monitor state  changes  in  animals was \ndiscussed  by  Arduini  [1],  but  to  our  knowledge  this  approach  has  not  previously \nbeen  applied to human EEG. \n\nThe origin  and  function  of nonstationarity  in  EEG  synchrony  are  not  yet  under(cid:173)\nstood.  Decreased  EEG  coherence  during drowsiness  might result from  inactivation \nof subcortical  brain  systems  coordinating  activity  in  separate  cortical  EEG  gen(cid:173)\nerators  during wakefulness,  or from emergence  of drowsiness-related  EEG  activity \nprojecting  preferentially  to  one  part  of the  scalp  surface.  Similarly,  increases  in \ncoherence  in  drowsiness  might either result from  increased  synchrony between  cor(cid:173)\ntical  generators,  or  from  volume  conduction  of enhanced  activity  generated  at  a \nsingle cortical or subcortical site.  Measuring changes in  EEG  coherence  and corre(cid:173)\nlation during  other  cognitive  tasks  give  clues  to the  possible  role  of variable  EEG \nsynchrony in  brain and cognitive dynamics. \n\nWe are now investigating to what extent moving EEG coherence  and/or correlation \n\n\f936 \n\nS. MAKEIG, T.-P. JUNG, T. J. SEJNOWSKl \n\nmeasures,  in  combination with spectral  amplitude measures  [4],  will  allow  practi(cid:173)\ncal,  robust,  continuous,  and near-real time estimation of alertness  level in  auditory \ndetection  and other task environments. \n\nEstimated and Actual Error Rates \n\nTralnln~ : 3674 \nTestse : 3674 \nRMS : 0.1706 \nCorr : 0.8246 \n\n,\" .. . :..,.,,.,,,,. \n\nI \n\n.-.-.-.~.~.-\n\n5 \n\n10 \n\n15 \n\nTime on Task (min) \n\n20 \n\n25 \n\n30 \n\nTralnl~: 3674 \nTest se : 3648 \nRMS : 0.2418 \n\u2022 .rCorr : 0.7159 \np  , \n\n5 \n\n10 \n\n15 \n\nTime on Task (min) \n\n20 \n\n25 \n\n30 \n\n?i \n8  0.8 \nE  0.6 \nG> \nOJ  0.4 \na: \ng 0.2 \n0 \n\nW \n\n-0.2 \n0 \n\n~ \n8  0.8 \n~ \n0.6 \n~  0.4 \na: \ng 0.2 \nw \n0 \n-0.2 \n0 \n\nFigure 3:  Changes  in  detection  rate  (95-s  exponential window)  and  their  estimate \nusing  a  feedforward  three-layer  perceptron  on  moving correlations  between  (1-20 \nHz)  band passed EEG signals for  8 selected pairs of 7 scalp channels.  The top panel \nshows  the  training session,  the bottom  panel  the  testing session.  Solid  lines  show \nthe actual error  rate time course; dashed lines,  the estimate.  Correlation and RMS \nerror  between  the two  are indicated. \n\nTable 1:  The results of alertness monitoring using moving EEG pairwise correlation. \n\nTest \nset \n3654 \n\n3656 \n\nSubject  B \n\nTraining  Set \n\n3654 \n\nrms:  0.17 \ncorr  :  0.83 \nrms:  0.25 \ncorr  :  0.54 \n\n3656 \n\nrms  :  0.22 \ncorr  :  0.73 \nrms :  0.14 \ncorr:  0.76 \n\nTest \nset \n3648 \n\n3674 \n\nTest \nset \n3665 \n\n3673 \n\nSubject  A \n\nTraining  Set \n\n3648 \n\nrms:  0.17 \ncorr  :  0.87 \nrms:  0.21 \ncorr  :  0.73 \n\n3674 \n\nrms :  0.26 \ncorr  :  0.68 \nrms:  0.17 \ncorr:  0.83 \n\nSubject  C \n\nTraining  Set \n\n3665 \n\nrms  :  0.19 \ncorr  :  0.76 \nrms :  0.18 \ncorr  :  0.67 \n\n3673 \n\nrms:  0.23 \ncorr  :  0.65 \nrms:  0.17 \ncorr:  0.70 \n\n\fUsing Feedforward Neural Networks to Monitor Alertness from  EEG \n\n937 \n\nAcknowledgments \n\nThis work was supported by a grant (ONR.Reimb.30020.6429) to  the Naval Health \nResearch  Center  by  the  Office  of  Naval  Research.  The  views  expressed  in  this \narticle  are  those  of the  authors  and  do  not  reflect  the  official  policy  or position of \nthe Department of the  Navy,  Department of Defense,  or  the  U.S.  Government.  We \nacknowledge  the  contributions of Keith  Jolley,  F.scot  Elliott,  and  Mark  Postal in \ncollecting and processing  the  data,  and  thank Tony  Bell  for  suggestions. \n\nReferences \n\n[1]  Arduini A ..  1979. In-phase brain activity and sleep.  Electroencephalog clin Neu(cid:173)\n\nrophysioI47,441-9 \n\n[2]  Borodkin SM, GrindeI' OM, Boldyreva GN, Zaitsev VA  & Luk'ianov V.1.1987. \nDynamics of the spectral-coherent characteristics of the human EEG in healthy \nsubjects  and brain pathology. Zh  Vyssh  Nerv  Deiat 37,  22-30 \n\n[3]  Davis  H.,  Davis  P.A.,  Loomis A.L.,  Harvey  E.N.,  & Hobart  G.  1938.  Human \n\nbrain potentials during the onset of sleep.  J  Neurophysiol1,  24-38 \n\n[4]  Jung T-P,  Makeig S.,  Stensmo  M.,  &  Sejnowski  T.  Estimating alertness  from \n\nthe  EEG power spectrum,  submitted for  publication. \n\n[5]  Kenemans  J.L .,  Molenaar  P.C.M. ,  Verbaten  M.N.  & Slangen  J.L.  1991.  Re(cid:173)\nmoval of the ocular artifact from the EEG: a comparison of time and frequency \ndomain methods with simulated and real  data.  Psychophysiolog 28,  114-121 \n\n[6]  Makeig S.  & Inlow  M.  1993.  Lapses  in  alertness:  Coherence  of fluctuations  in \nperformance and EEG spectrum.  Electroencephalog  clin Neurophysiol86, 23-35 \n[7]  Morgan  N.,  & Bourlard  H.  1990.  Generalization and  parameter estimation in \nfeedforward  nets:  some experiments.  Neural  Information  Processing  Systems, \n2,  630-637. \n\n[8]  Nielsen T, Abel A,  Lorrain D, & Montplaisir J . 1990. Interhemispheric EEG co(cid:173)\n\nherence  during sleep  and wakefulness in left- and right-handed subjects. Brain \nand  Cognition  14,  113-25 \n\n[9]  Rumelhart D,  Hinton G, & Williams R,  1986. Learning internal representation \n\nby  error propagation,  Parallel  distributed processing,  Chap.  8. \n\n[10]  Santamaria J. & Chiappa K.H.  1987. The EEG of drowsiness in normal adults. \n\nJ  Clin  Neurophysiol4,  327-82 \n\n[11]  Walter  D.O.  1968.  Coherence  as  a  measure  of  relationship  between  EEG \n\nrecords.  Electroencephalog  clin  N europhysiol 24,  282 \n\n\f", "award": [], "sourceid": 1087, "authors": [{"given_name": "Scott", "family_name": "Makeig", "institution": null}, {"given_name": "Tzyy-Ping", "family_name": "Jung", "institution": null}, {"given_name": "Terrence", "family_name": "Sejnowski", "institution": null}]}