{"title": "A Neural Model of Delusions and Hallucinations in Schizophrenia", "book": "Advances in Neural Information Processing Systems", "page_first": 149, "page_last": 156, "abstract": null, "full_text": "A  Neural Model of Delusions and \nHallucinations in  Schizophrenia \n\nEytan Ruppin and James  A.  Reggia \n\nDepartment of Computer Science \n\nUniversity of Maryland, College  Park,  MD  20742 \n\nruppin@cs.umd.edu \n\nreggia@cs.umd.edu \n\nDavid Horn \n\nSchool of Physics  and Astronomy, \n\nTel Aviv University, Tel  Aviv  69978, Israel \n\nhorn@vm.tau.ac.il \n\nAbstract \n\nWe  implement and study a computational model of Stevens'  [19921 \ntheory  of the  pathogenesis  of schizophrenia.  This  theory  hypoth(cid:173)\nesizes  that  the  onset  of schizophrenia  is  associated  with  reactive \nsynaptic regeneration occurring in brain regions receiving degener(cid:173)\nating temporal lobe  projections.  Concentrating on  one  such  area, \nthe  frontal  cortex,  we  model  a  frontal  module  as  an  associative \nmemory neural  network  whose  input synapses  represent  incoming \ntemporal  projections.  We  analyze  how,  in  the  face  of weakened \nexternal input projections, compensatory strengthening of internal \nsynaptic connections and increased  noise levels can maintain mem(cid:173)\nory  capacities  (which  are  generally  preserved  in  schizophrenia) . \nHowever,  These  compensatory  changes  adversely  lead  to  sponta(cid:173)\nneous,  biased  retrieval  of stored  memories,  which  corresponds  to \nthe occurrence  of schizophrenic  delusions  and hallucinations with(cid:173)\nout  any  apparent  external  trigger,  and  for  their  tendency  to  con(cid:173)\ncentrate on just few  central themes.  Our  results explain why  these \nsymptoms tend  to wane  as  schizophrenia progresses,  and why  de(cid:173)\nlayed  therapeutical intervention leads to a  much slower  response. \n\n\f150 \n\nEytan Ruppin,  James  A.  Reggia,  David Hom \n\n1 \n\nIntroduction \n\nThere has been a growing interest in recent years in the use of neural models to inves(cid:173)\ntigate  various  brain pathologies and their  cognitive  and  behavioral effects.  Recent \npublished  examples  of such  studies  include  models  of cortical  plasticity  following \nstroke,  Alzheimer's disease  and schizophrenia,  and cognitive  and behavioral explo(cid:173)\nrations  of aphasia,  acquired  dyslexia  and  affective  disorders  (reviewed  in  [1,  2]). \nContinuing this  line  of study,  we  present  a  computational account  linking specific \npathological  synaptic  changes  that  are  postulated  to  occur  in  schizophrenia,  and \nthe emergence of schizophrenic  delusions  and hallucinations.  The latter symptoms \ndenote  persistent,  unrealistic,  psychotic  thoughts  (delusions)  or  percepts  (halluci(cid:173)\nnations)  that may at times flood  the patient in  an overwhelming, stressful manner. \n\nThe wealth of data gathered  concerning  the pathophysiology of schizophrenia sup(cid:173)\nports  the  involvement  of  both  the  frontal  and  the  temporal  lobes.  On  the  one \nhand,  there  are  atrophic  changes  in  the  hippocampus  and  parahippocampal areas \nincluding neuronal loss and gliosis.  On the other hand, neurochemical and morpho(cid:173)\nmetric studies testify to an expansion of various receptor binding sites and increased \ndendritic branching in the frontal cortex of schizophrenics.  Stevens has recently pre(cid:173)\nsented a theory linking these  temporal and frontal findings,  claiming that the onset \nof schizophrenia  is  associated  with  reactive  anomalous sprouting  and  synaptic  re(cid:173)\norganization taking place  in the projection sites  of degenerating temporal neurons, \nincluding  (among various cortical  and subcortical  structures)  the frontal lobes  [3]. \n\nThis  paper  presents  a  computational study of Stevens'  theory.  Within the frame(cid:173)\nwork  of  a  memory  model  of  hippocampal-frontal  interaction,  we  show  that  the \nintroduction of the  'microscopic' synaptic  changes  that underlie  Stevens'  hypothe(cid:173)\nsis  can help  preserve  memory function  but results in specific  'pathological' changes \nin the  'macroscopic' behavior of the network.  A small subset of the patterns stored \nin the network are now spontaneously retrieved at times, without being cued by  any \nspecific  input pattern.  This emergent behavior shares some of the important char(cid:173)\nacteristics of schizophrenic delusions and hallucinations, which frequently  appear in \nthe  absence  of any  apparent external trigger,  and tend to concentrate  on  a  limited \nset of recurrent  themes  [4].  Memory  capacities  are fairly  preserved  in schizophren(cid:173)\nics,  until  late  stages  of the  disease  [5].  In  Section  2  we  present  our  model.  The \nanalytical  and  numerical  results  obtained  are  described  in  Section  3,  followed  by \nour conclusions in  Section 4. \n\n2  The Model \n\nAs  illustrated  in  Figure  1,  we  model  a  frontal  module  as  an  associative  memory \nattractor  neural  network,  receiving  its  input  memory  cues  from  decaying  exter(cid:173)\nnal  input  fibers  (representing  the  degenerating  temporal  projections).  The  net(cid:173)\nwork's internal connections,  which store the memorized patterns, undergo synaptic \nstrengthening  changes  that  model  the  reactive  synaptic  regeneration  within  the \nfrontal module.  The effect  of other  diffuse external projections is  modeled as back(cid:173)\nground  noise.  A  frontal  module  represents  a  macro-columnar unit  that  has  been \nsuggested  as a  basic functional building block of the neocortex  [6].  The assumption \nthat memory retrieval from  the  frontal  cortex is  invoked by  the  firing  of incoming \n\n\fA  Neural Model of Delusions and Hallucinations  in Schizophrenia \n\n151 \n\ntemporal projections  is  based  on the  notion  that  temporal structures  have  an  im(cid:173)\nportant role in establishing long-term memory in the neocortex and in the retrieval \nof facts  and events  (e.g.,  [7]). \n\n--(cid:173)\n\n,-, \n\nI \n\nmodules \n\nOther cortical \n\n\" \n: \n:  (Influence nxxJeled as,' \n! \n\nnoise) \n\n,\"  ......... .. \n\nI \nI \n\n.......  T \n.... .. \n\n\\ \n\n,I\"  .... .. \n\nFigure  1:  A  schematic  illustration  of the  model.  A  frontal  module  is  modeled \nas  an  attractor  neural  network  whose  neurons  receive  inputs  via  three  kinds  of \nconnections:  internal  connections  from other frontal  neurons,  external  connections \nfrom  temporal  lobe  neurons,  and  diffuse  external  connections  from  other  cortical \nmodules, modeled as  noise. \n\nThe attractor network we  use  is a biologically-motivated variant of Hopfield's ANN \nmodel,  proposed  by  Tsodyks  &  Feigel'man  [8].  Each  neuron  i  is  described  by  a \nbinary variable  Si  =  {1,0}  denoting  an active  (firing)  or passive  (quiescent)  state, \nrespectively.  M  =  aN distributed  memory patterns eIJ  are  stored  in  the  network. \nThe  elements  of each  memory  pattern  are  chosen  to be  1  (0)  with  probability p \n(1- p) respectively,  with p  ~ 1.  All N  neurons in the network have a fixed uniform \nthreshold O. \n\nIn its initial, undamaged state, the weights of the internal synaptic connections  are \n\nM \n\nWij  = N  L....t(eIJi - p)(e j  - p)  , \n\nCO\" \n\n(1) \n\nwhere  Co  =  1.  The post-synaptic potential (input field)  hi of neuron i  is  the sum of \ninternal contributions from other  neurons  and external projections  Fi e \n\nIJ=I \n\nhi(t) = L WijSj(t - 1) + Fie. \n\nj \n\nThe updating rule for  neuron  i  at time t  is  given by \n\nS .(t)  = {I,  with p~ob.  G(hi(t) - 0) \n\n0,  otherwIse \n\n1 \n\n(2) \n\n(3) \n\n\f152 \n\nEytan  Ruppin, James  A. Reggia, David Hom \n\nwhere G is the sigmoid function G(x) =  1/(1+exp( -x/T)), and T  denotes the noise \nlevel.  The activation level of the stored memories is  measured by  their  overlaps mil \nwith  the current state of the  network,  defined  by \n\nmll(t) = \n\n1 \n\n(1  _  )N L)er - p)Si(t)  . \n\nN \n\nP \n\nP \n\ni=l \n\n(4) \n\nStimulus-dependent  retrieval is  modeled  by  orienting the  field  Fe  with  one  of the \nmemorized patterns  (the  cued pattern, say e), such  that \n\nF/ = e . e\\  ,  (e  > 0)  . \n\n(5) \nFollowing the presentation of an external input cue, the network state evolves  until \nit  converges  to a  stable  state.  The  network  parameters are  tuned  such  that  in  its \ninitial, undamaged state it correctly retrieves the cued patterns (eo  =  0.035 , Co  =  1, \nT  =  0.005). \nWe also examine the network's behavior  in the  absence  of any specific  stimulus.  The \nnetwork  may  either  continue  to  wander  around  in  a  state of random low  baseline \nactivity,  or  it  may  converge  onto  a  stored  memory  state.  We  refer  to  the  latter \nprocess  as  spontaneous  retrieval. \nOur investigation of Stevens' work proceeds in two stages.  First we examine and an(cid:173)\nalyze the behavior of the network when it undergoes  uniform synaptic changes that \nrepresent  the  pathological  changes  occurring  in  accordance  with  Stevens'  theory. \nThese include the weakening of external input projections (e  !) and the increase  in \nthe  internal projections  (c  i) and  noise  levels  (T i).  In  the  second  stage,  we  add \nthe assumption that the internal synaptic compensatory changes have an additional \nHebbian  activity-dependent  component, and examine the effect  of the rule \n\nwhere Sk  is  1 (0)  only if neuron  k  has been  consecutively firing  (quiescent)  for  the \nlast  r  iterations,  and 'Y  is  a constant. \n\n(6) \n\n3  Results \n\nWe now show some simulation and analytic results, examining the effects of the 'mi(cid:173)\ncroscopic' pathological changes,  taking place in accordance  with Stevens' theory, on \nthe  'macroscopic'  behavior  of the  network.  The  analytical  results  presented  have \nbeen  derived  by  calculating the magnitude of randomly formed  initial 'biases',  and \ncomparing their effect  on the network 's dynamics versus the effect  of externally pre(cid:173)\nsented  input  cues.  This  comparison  is  performed  by  formulating  a  corresponding \noverlap  master  equation,  whose  fixed  point  dynamics  are  investigated  via  phase(cid:173)\nplane  analysis,  as  described  in  [9].  First,  we  study  whether  the  reactive  synaptic \nchanges  (occurring in  both internal and external,  diffuse  synapses)  are really  com(cid:173)\npensatory,  i.e., to what extent  can they help  maintaining memory capacities  in the \nface of degenerating external input synapses.  As illustrated in Figure 2, we find that \nincreased noise levels can (up to some degree)  preserve  memory retrieval in the face \nof decreased  external  input  strength.  Increased  synaptic  strengthening  preserves \n\n\fA  Neural Model of Delusions and Hallucinations in Schizophrenia \n\n153 \n\n, /  \n\ni' \n\n(a) \n\n1.0 \n\nO.B \n\n0.8 \n\n! . 6 \n\n0.( \n\n0.2 \n\n.  ~_.__  \u2022 \n\n:  - . _0.035 \n_  0.025 \nI \n- --- \u2022\u2022 0.015 \n--- \u2022\u2022 0.005 \n\n'- -- -\n\n0.0 \n\n0.000 \n\n- :...  ----\n\n0.010 \n\nT \n\n0.020 \n\n(b) \n\n1.0 \n\nO.B \n\n, \n\n- ._  ,! O.O35 \n------ \u2022\u2022  iO.025 \n---- \u2022\u2022  '0.015 \n- - \u2022 \u2022   10.014 \n-\nG---o \u2022\u2022 ; 0.013 \n<)--0._  10.012 \n--- \u2022\u2022 ; 0.005 \n\n0.6 \n\nt \n6 \n\n0.( \n\n0.2 \n\n0\u00b78. \n\n0.020 \n\nT \n\nFigure 2:  Stimulus-dependent retrieval performance, measured by the average final \noverlap  m,  as  a  function  of the  noise  level  T.  Each  curve  displays  this relation  at \na  different  magnitude of external  input  projections  e.  (a)  Simulation results.  (b) \nAnalytic approximation. \n\nmemory  retrieval  in  a  similar manner,  and  the  combined  effect  of these  synaptic \ncompensatory measures is synergistic. \n\nSecond,  although  the  compensatory  synaptic  changes  help  maintain  memory  re(cid:173)\ntrieval  capacities,  they  necessarily  have  adverse  effects,  leading  eventually  to  the \nemergence  of spontaneous  activation  of non-cued  memory  patterns;  the  network \nconverges  to some of its memory  patterns in  a  pathological, autonomous manner, \nin the absence  of any external input stimuli.  This emergence  of pathological spon(cid:173)\ntaneous  retrieval, when  either  the  noise  level  or  the  internal  synaptic strength  (or \nboth) are increased  beyond some point, is demonstrated in  Figure 3. \nThird,  when  the compensatory regeneration  of internal synapses  has an additional \nHebbian component (representing  a  period of increased  activity-dependent plastic(cid:173)\nity  due  to  the  regenerative  synaptic  changes),  a  biased spontaneous  retrieval  dis(cid:173)\ntribution is  obtained.  That is,  as  time evolves  (measured in time units of 'trials'), \nthe  distribution of patterns spontaneously  retrieved  by  the network  in  a  patholog(cid:173)\nical  manner  tends  to  concentrate  only  on one  or  two  of all  the  memory  patterns \nstored  in the network,  as  is  shown in Figure 4a.  This highly peaked  distribution is \nmaintained for  a few  hundred  additional trials until  memory retrieval  sharply  col(cid:173)\nlapses to zero  as  a  global mixed-state attractor is formed.  Such  a  mixed attractor \nstate  does not have  very  high overlap  with any  memorized pattern, and thus does \nnot  represent  any  well-defined  cognitive  or  perceptual  item.  It is  an  end  state  of \nthe Hebbian,  activity-dependent evolution of the network.  Yet,  even  after activity(cid:173)\ndependent  changes ensue, if spontaneous  activity does  not emerge  the distribution \nof retrieved  memories remains homogeneous  (see  Figure  4b).  Eventually,  a  global \n\n\f154 \n\n(a) \n\n1.0 \n\nO.B \n\n0.6 \n\n0.4 \n\n0.2 \n\nt \n! \n\n, , , \n\n0.0 \n\n0.005 \n\n, ,---.. \n\n-\n\n, \n, \n\n, \n\nAnalytIc  _Imollon \n\n-\n- - - - Slmulallon \n\n0.010 \n\nT \n\n0.015 \n\n0.020 \n\nEytan  Ruppin.  James A.  Reggia.  David Hom \n\n, ........... . \n\n~ .. .. .. ,1' \" \n\n, \n\n, \n\n(b) \n\n1.0 \n\nO.B \n\n0.6 \n\nf \n\n0.4 \n\n0.2 \n\n' \n\n0.0  ':--~-'-:-'-::--~~--~----:' \n4.0 \n\n2.0 \n\n3.0 \n\n2.5 \n\n3.5 \n\nFigure  3: \n(a)  Spontaneous  retrieval,  measured  as  the  highest  final  overlap  m \nachieved  with  any  of the  stored  memory  patterns,  displayed  as  a  function  of the \nnoise  level  T.  c = 1.  (b)  Spontaneous  retrieval  as  a  function  of internal synaptic \ncompensation factor  c.  T  = 0.009. \n\nmixed-state attractor is formed, and the network looses  its retrieval capacities,  but \nduring this process  no memory pattern gets to dominate the retrieval output.  Our \nresults  remain qualitatively similar even  when  bounds  are  placed  on  the  absolute \nmagnitude of the synaptic weights. \n\n4  Conclusions \n\nOur results suggest that the formation of biased spontaneous retrieval requires the \nconcomitant occurrence  of both degenerative  changes  in  the  external  input fibers, \nand regenerative Hebbian changes in the intra-modular synaptic connections.  They \nadd  support  to the plausibility of Stevens'  theory  by  showing that it may be  real(cid:173)\nized  within a  neural  model,  and account  for  a  few  characteristics  of schizophrenic \nsymptoms: \n\n\u2022  The emergence of spontaneous, non-homogeneous retrieval is a self-limiting \nphenomenon  (as  eventually  a  cognitively  meaningless  global  attractor  is \nformed) - this parallels the clinical finding that as schizophrenia progresses \nboth  delusions  and hallucinations tend  to wane,  while  negative  symptoms \nare  enhanced  [10]. \n\n\u2022  Once  converged to, the network has a much larger tendency to remain in a \nbiased  memory state than in  a  non biased one - this is  in  accordance  with \nthe  persistent  characteristic of schizophrenic florid  symptoms. \n\n\u2022  As  more  spontaneous  retrieval  trials  occur  the  frequency  of spontaneous \nretrieval increases - indeed, while early treatment in young psychotic adults \n\n\fA Neural  Model of Delusions and Hallucinations  in  Schizophrenia \n\n(a) \n\n155 \n\n(b) \n\n1.0 ~-~-~~-:----~-----. \n\n0.100 ~----\"-~~--~-----. \n\n0.8 \n\n0.080 \n\n~AftOf200trlals \n( \u00b7 Aft ... 400tr1a1a \n? \n\n~ 0.6 \n\nt \n\n1 \n.~ \n~ 0.4 \n\n0.2 \n\nG---f] Aft.r 200 tria_ \n{3 - -(:~Aft.r500trials \n~ - v Aft ... 800triala \n\nQ \n.,. \n\n>' ., \n~  ~ \n.,  . \n;  , \n~  1 \n\n~  0.000 \n\nt \n1 l  0.040 \n\n0.020 \n\nFigure 4:  (a) The distribution of memory patterns spontaneously retrieved.  The x(cid:173)\naxis enumerates the memories stored, and the y-axis denotes the retrieval frequency \nof each  memory.  'Y  = 0.0025.  (b)  The  distribution of stimulus-dependent retrieval \nof memories.  'Y  = 0.0025 . \n\nleads  to  early  response  within  days,  late,  delayed  intervention  leads  to  a \nmuch slower  response  during one or more months [11]. \n\nThe current  model generates some testable predictions: \n\n\u2022  On  the  neuroanatomical level,  the  model  can  be  tested  quantitatively by \n\nsearching  for  a  positive  correlation  between  a  recent  history  of florid  psy(cid:173)\nchotic  symptoms  and  postmortem  neuropathological  findings  of synaptic \ncompensation.  (For  example,  this  kind  of correlation,  between  indices  of \nsynaptic  area  and  cognitive  functioning  was  found  in  Alzheimer  patients \n[12]). \n\n\u2022  On  the  physiological level,  the increased  compensatory  noise should mani(cid:173)\n\nfest  itself in increased spontaneous neural  activity.  While this prediction is \nobviously difficult to examine directly,  EEG studies in schizophrenics show \nsignificant increase  in  slow-wave delta activity which  may reflect  increased \nspontaneous  activity  [13]. \n\n\u2022  On  the  clinical  level,  due  to  the  formation  of  a  large  and  deep  basin  of \nattraction around the memory pattern which is  at the focus of spontaneous \nretrieval,  the  proposed  model  predicts  that  its  retrieval  (and  the  elucida(cid:173)\ntion  of the  corresponding  delusions  or  hallucinations)  may  be  frequently \ntriggered  by  various  environmental  cues.  A  recent  study  points  in  this \ndirection  [14]. \n\n\f156 \n\nEytan  Ruppin,  James A.  Reggia,  David Hom \n\nAcknowledgements \nThis research  has  been  supported by  a  Rothschild Fellowship  to Dr.  Ruppin. \n\nReferences \n\n[1]  J.  Reggia,  R.  Berndt,  and  L.  D' Autrechy.  Connectionist models  in  neuropsy(cid:173)\n\nchology.  In  Handbook  of Neuropsychology,  volume 9.  1994, in  press. \n\n[2]  E.  Ruppin.  Neural  modeling of psychiatric  disorders.  Network:  Computation \n\nin  Neural  Systems,  1995.  Invited  review  paper,  to appear. \n\n[3]  J .R. Stevens.  Abnormal reinnervation as  a  basis for  schizophrenia:  A hypoth(cid:173)\n\nesis.  Arch.  Gen.  Psychiatry, 49 :238-243, 1992. \n\n[4]  S.K.  Chaturvedi  and  V.D.  Sinha.  Recurrence  of hallucinations in  consecutive \nepisodes of schizophrenia and affective disorder.  Schizophrenia  Research, 3: 103-\n106,  1990. \n\n[5]  M.  Marsel  Mesulam.  Schizophrenia  and  the  brain.  New  England  Journal  of \n\nMedicine,  322(12):842-845, 1990. \n\n[6]  P.S.  Goldman  and  W.J .H.  Nauta.  Columnar  distribution  of cortica-cortical \nfibers  in  the  frontal,  association,  limbic  and  motor  cortex  of the  developing \nrhesus  monkey.  Brain  Res.,  122:393-413,1977. \n\n[7]  L.  R. Squire.  Memory  and  the  hippocampus:  A  synthesis from findings  with \n\nrats,  monkeys, and humans.  Psychological  Review,  99 :195-231 , 1992. \n\n[8]  M.V. Tsodyks and M.V. Feigel'man. The enhanced storage capacity in  neural \n\nnetworks  with low  activity level.  Europhys.  Lett.,  6:101 - 105,  1988. \n\n[9]  D.  Horn  and E.  Ruppin.  Synaptic compensation in attractor neural  networks: \nModeling  neuropathological  findings  in  schizophrenia.  Neural  Computation, \npage To appear,  1994. \n\n[10]  W .T. Carpenter and R.W. Buchanan.  Schizophrenia.  New  England  Journal of \n\nMedicine,  330:10, 1994. \n\n[11]  P.  Seeman.  Schizophrenia  as  a  brain  disease:  The  dopamine  receptor  story. \n\nArch.  Neurol., 50:1093-1095, 1993. \n\n[12]  S.  T.  DeKosky  and  S.W.  Scheff.  Synapse  loss  in  frontal  cortex  biopsies \nin  alzheimer's  disease:  Correlation  with  cognitive  severity.  Ann.  Neurology, \n27(5):457-464,  1990. \n\n[13]  Y.  Jin, S.G. Potkin, D.  Rice,  and J. Sramek et. al.  Abnormal EEG responses to \nphotic stimulation in schizophrenic patients.  Schizophrenia  Bulletin, 16(4):627-\n634,  1990. \n\n[14]  R.E. Hoffman and J .A.  Rapaport. A psycholoinguistic study of auditory /verbal \nhallucinations:  Preliminary findings.  In  David A. and Cutting J. , editors,  The \nNeuropsychology  of Schizophrenia.  Erlbaum,  1993. \n\n\f", "award": [], "sourceid": 902, "authors": [{"given_name": "Eytan", "family_name": "Ruppin", "institution": null}, {"given_name": "James", "family_name": "Reggia", "institution": null}, {"given_name": "David", "family_name": "Horn", "institution": null}]}