{"title": "ICEG Morphology Classification using an Analogue VLSI Neural Network", "book": "Advances in Neural Information Processing Systems", "page_first": 731, "page_last": 738, "abstract": null, "full_text": "ICEG Morphology Classification using an \n\nAnalogue VLSI Neural Network \n\nRichard Coggins, Marwan Jabri, Barry Flower and Stephen Pickard \n\nSystems Engineering and Design Automation Laboratory \n\nDepartment of Electrical Engineering J03, \n\nUniversity of Sydney,  2006,  Australia. \n\nEmail:  richardc@sedal.su.oz.au \n\nAbstract \n\nAn  analogue  VLSI  neural  network  has  been  designed  and  tested \nto perform cardiac morphology classification tasks.  Analogue tech(cid:173)\nniques were chosen to meet the strict power and area requirements \nof an Implantable Cardioverter Defibrillator (ICD) system.  The ro(cid:173)\nbustness  of the neural network  architecture  reduces  the impact of \nnoise,  drift  and offsets  inherent  in  analogue approaches.  The net(cid:173)\nwork  is a  10:6:3 multi-layer percept ron  with on  chip  digital weight \nstorage,  a  bucket  brigade input  to feed  the  Intracardiac  Electro(cid:173)\ngram  (ICEG)  to  the  network  and  has  a  winner  take  all  circuit \nat  the  output.  The  network  was  trained  in  loop  and  included  a \ncommercial ICD in the signal processing path.  The system has suc(cid:173)\ncessfully distinguished arrhythmia for different patients with better \nthan 90%  true positive and true negative detections for  dangerous \nrhythms which cannot be detected  by present  ICDs.  The chip  was \nimplemented in 1.2um CMOS and consumes less than 200n W max(cid:173)\nimum average power in an area of 2.2  x 2.2mm2. \n\n1 \n\nINTRODUCTION \n\nTo  the  present  time,  most  ICDs  have  used  timing  information from  ventricular \nleads  only  to  classify  rhythms  which  has meant some  dangerous  rhythms  can  not \nbe  distinguished  from  safe  ones,  limiting  the  use  of the  device.  Even  two  lead \n\n\f732 \n\nRichard Coggins,  Marwan Jabri,  Barry Flower,  Stephen Pickard \n\n4.00 \n\nHO \n\n3.00 \n\n2.00 \nI.SO  ___  :::::::! \n\nQ \n\n1.00 \n\nO.SO \n\nFigure  1:  The Morphology of ST and VT retrograde  1:1. \n\natrial/ventricular systems fail  to  distinguish  some  rhythms  when  timing informa(cid:173)\ntion alone is  used  [Leong and Jabri,  1992].  A  case  in  point is the separation of Si(cid:173)\nnus Tachycardia (ST) from Ventricular Tachycardia with 1:1 retrograde conduction. \nST  is  a  safe  arrhythmia which  may occur  during  vigorous  exercise  and  is  charac(cid:173)\nterised  by a  heart rate of approximately 120 beats/minute.  VT retrograde  1:1  also \noccurs  at the same low  rate but can be a  potentially fatal condition.  False negative \ndetections  can cause  serious heart  muscle  injury while false  positive  detections  de(cid:173)\nplete  the  batteries,  cause  patient suffering  and may lead to costly  transplantation \nof the  device.  Figure 1 shows  however,  the way  in which  the morphology changes \non  the  ventricular  lead  for  these  rhythms.  Note,  that  the  morphology  change  is \npredominantly in the  \"QRS  complex\"  where  the  letters  QRS  are the  conventional \nlabels  for  the  different  points  in  the  conduction  cycle  during  which  the  heart  is \nactually pumping blood. \n\nFor a number of years,  researchers have studied template matching schemes in order \nto try and detect such morphology changes.  However,  techniques such as correlation \nwaveform analysis [Lin et.  al.,  1988],  though quite successful  are too computation(cid:173)\nally  intensive  to  meet  power  requirements.  In  this  paper,  we  demonstrate  that \nan  analogue  VLSI  neural  network  can  detect  such  morphology changes  while  still \nmeeting  the  strict  power  and  area  requirements  of an  implantable system.  The \nadvantages of an  analogue  approach  are  born out when  one  considers  that  an en(cid:173)\nergy  efficient  analogue  to  digital  converter  such  as  [Kusumoto et.  al.,  1993]  uses \n1.5nJ  per  conversion  implying 375n W  power  consumption for  analogue to  digital \nconversion of the ICEG alone.  Hence, the integration of a bucket brigade device and \nanalogue neural network provides a  very efficient way of interfacing to the analogue \ndomain.  Further,  since  the  network  is  trained  in  loop  with  the  ICD  in real  time, \nthe effects  of device  offsets,  noise,  QRS  detection jitter and signal distortion in the \nanalogue circuits  are largely  alleviated. \n\nThe next section discusses  the chip  circuit designs.  Section 3 describes  the method \n\n\fICEG  Morphology  Classification  Using  an  Analogue  VLSI Neural Network \n\n733 \n\nAowAcId . . .  \n\n1axl Syna .... AIRy \n\n\"(cid:173)\nColumn \nAoIcIr.-\n\nI \n\no.ta Reglsl... \n\nI \n\nIClkcMmux  10 DOD DO \n\nI WTAI \n\nBu1I ... \n\nFigure  2:  Floor  Plan and  Photomicrograph of the  chip \n\nused  to train the network for the morphology classification task.  Section 4 describes \nthe  classifier  performance  on  seven  patients  with  arrhythmia  which  can  not  be \ndistinguished using the heart rate only.  Section 5 summarises the results,  remaining \nproblems and future  directions for  the  work. \n\n2  ARCHITECTURE \n\nThe neural network  chip consists of a  10:6:3 multilayer perceptron,  an input bucket \nbrigade  device  (BBD)  and a  winner  take  all  (WTA)  circuit  at  the  output.  A floor \nplan  and  photomicrograph of the  chip  appears  in  figure  2.  The  BBD  samples the \nincoming  ICEG  at  a  rate of 250Hz.  For  three  class  problems,  the  winner  take  all \ncircuit  converts  the  winning  class  to  a  digital  signal.  For  the  two  class  problem \nconsidered  in  this  paper ,  a  simple  thresholding  function  suffices.  The  following \nsubsections briefly describe the functional elements of the chip .  The circuit diagrams \nfor  the  chip  building blocks  appear in  figure  3. \n\n2.1  BUCKET  BRIGADE DEVICE \n\nOne  stage  of the  bucket  brigade  circuit  is  shown  in  figure  3.  The  BBD  uses  a \ntwo  phase  clock  to  shift  charge  from  cell  to  cell  and  is  based  on  a  design  by \nLeong  [Leong,  1992] .  The  BBD  operates  by  transferring  charge  deficits  from  S \nto  D  in  each  of the  cells.  PHIl  and  PHI2  are  two  phase  non-overlapping  clocks. \nThe  cell  is  buffered  from  the  synapse  array  to  maintain high  charge  transfer  effi(cid:173)\nciency.  A sample and hold facility is  provided to store the input on  the gates of the \nsynapses.  The  BBD  clocks  are  generated  off  chip  and  are  controlled  by  the  QRS \ncomplex detector  in  the lCD. \n\n2.2  SYNAPSE \n\nThis  synapse  has  been  used  on  a  number  of  neural  network  chips  previously. \ne.g . [Coggins et.  al.,  1994] .  The synapse has five  bits plus sign weight storage which \n\n\f734 \n\nRichard Coggins,  Marwan  Jabri,  Barry Flower,  Stephen  Pickard \n\nNEURON \n\nBUJIOIII' \n\n00 \n\n\" \n\nBUCKET BRIGADE ClLL \n\n.---------------------------------------- --------------------, \n, \n, \n, \n, \n~ ! \n\nFigure 3:  Neuron,  Bucket  Brigade and Synapse Circuit  Diagrams. \n\nsets  the bias to a  differential  pair which performs the multiplication.  The bias ref(cid:173)\nerences  for  the weights are derived  from  a  weighted  current  source in the corner  of \nthe chip.  A four quadrant multiplication is achieved  by the four switches at the top \nof the differential pair. \n\n2.3  NEURON \n\nDue to the low  power  requirements,  the  bias currents  of the  synapse  arrays  are  of \nthe order  of hundreds  of nano  amps,  hence  the  neurons  must provide  an effective \nresistance  of many mega ohms to feed  the next synapse  layer  while  also  providing \ngain  control.  Without special  high  resistance  polysilicon,  simple resistive  neurons \nuse  prohibitive  area,  However,  for  larger  networks  with fan-in  much  greater  than \nten,  an  additional problem of common mode cancellation is  encountered,  That is, \nas  the fan-in  increases,  a  larger  common mode  range  is  required  or  a  cancellation \nscheme using  common mode feedback  is  needed. \nThe neuron of figure 3 implements such a cancellation scheme,  The mirrors MO/M2 \nand  Ml/M3  divide  the  input  current  and  facilitate  the  sum  at  the  drain  of M7. \nM7/M8  mirrors  the  sum  so  that  it  may  be  split  into  two  equal  currents  by  the \nmirrors  formed  by  M4,  M5  and  M6  which  are  then  subtracted  from  the  input \ncurrents.  Thus,  the  differential  voltage  vp  - Vm  is  a  function  of  the  transistor \ntransconductances,  the common mode input current  and  the feedback  factor ,  The \ngain  of the  neuron  can  be  controlled  by  varying  the  width  to  length  ratio  of the \nmirror transistors MO  and  Ml.  The implementation in this case  allows seven  gain \ncombinations, using  a  three  bit RAM  cell  to store  the gain, \n\n\fICEG  Morphology  Classification  Using  an  Analogue  VLSI Neural Network \n\n735 \n\nRunnngMUME \n\nImplantable \nC.cio\u00a5erlor \nDefibrillalOr \n\nNe ..... 1 \nNelwa'1< \nChip \n\nFigure 4:  Block  Diagram of the Training and Testing System. \n\nThe  importance  of a  common  mode  cancellation  scheme  for  large  networks  can \nbe  seen  when  compared  to  the  straight  forward  approach  of resistive  or  switched \ncapacitor  neurons.  This  may  be  illustrated  by  considering  the  energy  usage  of \nthe  two  approaches.  Firstly,  we  need  to  define  the  required  gain  of the  neuron \nas  a  function  of its  fan-in .  If we  assume  that  useful  inputs  to  the  network  are \nmostly sparse,  i.e.  with a small fraction of non-zero  values, then the gain is  largely \nindependent of the fan-in, yet the common mode signal increases linearly with fan(cid:173)\nin.  For  the  case  of a  neuron  which  does  not  cancel  the  common mode,  the  power \nsupply voltage  must be  increased  to  accommodate the common mode signal,  thus \nleading to a quadratic increase in energy use with fan-in.  A common mode cancelling \nneuron  on the  other  hand, suffers  only  a  linear increase  in  energy  use  with fan-in \nsince  extra voltage  range  is  not  required  and the  increased  energy  use  arises  only \ndue to the linear increase in common mode current. \n\n3  TRAINING SYSTEM \n\nThe system used  to train and test the neural network is shown in figure  4.  Control \nof training  and  testing  takes  place  on  the  PC.  The  PC  uses  a  PC-LAB  card  to \nprovide  analogue  and  digital  I/O . The  PC  plays  the  ICEG  signal  to the  input  of \nthe  commercial ICD  in  real  time.  Note,  that the  PC is  only required  for  initially \ntraining the network and in this case as a source of the heart signal.  The commercial \nICD  performs the function of QRS  complex detection using analogue circuits.  The \nQRS complex detection signal is then used to freeze  the  BBD  clocks of the chip,  so \nthat a  classification can take place. \nWhen training, a number of examples of the arrhythmia to be classified are selected \nfrom  a  single  patient  data base  recorded  during  an electrophysiological study and \npreviously classified by a cardiologist.  Since most of the morphological information \nis in the QRS complex, only these segments of the data are repeatedly presented to \n\n\f736 \n\nRichard Coggins.  Marwan  Jabri. Barry Flower.  Stephen  Pickard \n\nPatient  %  Training Attempts Converged \n\nRun 1 \n\nRun ~ \n\nH=3  H =  6  H=3  H=6 \n\n1 \n2 \n3 \n4 \n5 \n6 \n7 \n\n80 \n80 \n0 \n60 \n100 \n100 \n80 \n\n10 \n100 \n0 \n10 \n80 \n40 \n100 \n\nAverage \nIterations \n\n62 \n86 \n101 \n77 \n44 \n46 \n17 \n\n60 \n0 \n0 \n40 \n0 \n60 \n40 \n\n60 \n10 \n10 \n40 \n60 \n60 \n100 \n\nTable 1:  Training Performance of the system on seven  patients. \n\nthe network.  The weights are  adjusted according to the training algorithm running \non  the  PC  using  the  analogue  outputs of the  network  to reduce  the  output error . \nThe PC writes weights to the chip via the digital I/Os of the PC-LAB card and the \nserial weight  bus  of network.  The software  package  implementing the  training and \ntesting,  called  MUME  [Jabri et.  al., 1992],  provides  a  suite  of training  algorithms \nand control  options.  Online  training was  used  due  to its  success  in training small \nnetworks  and  because  the  presentation  of the  QRS  complexes  to the network  was \nthe slowest  part of the  training procedure.  The algorithm used for  weight  updates \nin this paper was  summed weight node perturbation  [Flower  and Jabri,  1993]. \n\nThe  system  was  trained  on  seven  different  patients  separately  all  of whom  had \nVT with  1:  1 retrograde  conduction.  Note,  that  patient  independent  training  has \nbeen  tried  but with mixed results  [Tinker,  1992] .  Table  1 summarises the training \nstatistics for the seven patients.  For each patient and each architecture, five  training \nruns  were  performed  starting  from  a  different  random  initial  weight  set.  Each \nof the  patients  was  trained  with  eight  of each  class  of arrhythmia.  The  network \narchitecture  used  was  10:H:1, where  H is  the  number  of hidden  layer  neurons  and \nthe unused neurons being disabled by setting their input weights to zero.  Two sets \nof data were  collected  denoted  Run  1 and  Run  2.  Run  1 corresponded  to output \ntarget  values  of \u00b10.6V within margin 0.45V and  Run 2 to output target  values  of \n\u00b10.2V within margin 0.05V.  A training attempt was  considered to have converged \nwhen the training set was correctly classified within two hundred training iterations. \nOnce  the  morphologies to  be  distinguished  have  been  learned for  a  given  patient, \nthe  remainder  of the  patient data base  is  played  back  in  a  continuous stream and \nthe outputs of the classifier  at each QRS  complex are logged and may be compared \nto the  classifications of a  cardiologist.  The  resulting  generalisation performance is \ndiscussed  in the next  section. \n\n4  MORPHOLOGY  CLASSIFIER GENERALISATION \n\nPERFORMANCE \n\nTable  2  summarises  the  generalisation  performance  of  the  system  on  the  seven \npatients for  the  training  attempts which  converged.  Most  of the  patients  show  a \ncorrect classification rate better than 90% for at least one architecture on one of the \n\n\fICEG  Morphology  Classification  Using  an  Analogue VLSI Neural  Network \n\n737 \n\nPatient \n\n1 \n2 \n3 \n4 \n5 \n6 \n7 \n\n1 \n2 \n3 \n4 \n5 \n6 \n7 \n\nNo.  of \n\nComplexes \nST  VT \n61 \n440 \n57 \n94 \n67 \n146 \n166 \n65 \n61 \n96 \n99 \n61 \n28 \n80 \n\n440 \n94 \n67 \n166 \n61 \n61 \n28 \n\n61 \n57 \n146 \n65 \n96 \n99 \n80 \n\n% Correct  Classifications Run  1 \n\nH  -\n\nST \n\n89\u00b110 \n99\u00b11 \n\n-\n\ni3 \nVT \n89\u00b13 \n99\u00b11 \n\n-\n\nH  =  6 \n\nST \n58\u00b10 \n100\u00b10 \n\n-\n\nVT \n99\u00b10 \n99\u00b11 \n\n-\n\n99\u00b11 \n94\u00b16 \n90\u00b15 \n99\u00b11 \n\n66\u00b144  76\u00b137 \n50\u00b13 \n75\u00b113 \n82\u00b11 \n89\u00b19 \n97\u00b11 \n84\u00b18 \n99\u00b11 \n97\u00b13 \n98\u00b15 \n99\u00b11 \n% Correct Classifications Run  2 \n99\u00b11 \n88\u00b12 \n94\u00b13 \n\n86\u00b114 \n94\u00b16 \n\n-\n\n99\u00b11 \n\n-\n\n99\u00b11 \n59\u00b12 \n49\u00b15 \n90\u00b110 \n99\u00b10 \n\n84\u00b12 \n76\u00b118 \n88\u00b12 \n92\u00b16 \n94\u00b13 \n\n-\n\n87\u00b17 \n84\u00b11 \n99\u00b11 \n94\u00b13 \n\n-\n\n100\u00b10 \n82\u00b15 \n99\u00b11 \n92\u00b13 \n\nTable  2:  Generalisation  Performance of the system on seven  patients. \n\nruns,  whereas,  a  timing based  classifier  can  not  separate  these  arrhythmia at  all. \nFor each  convergent  weight set the network  classified  the test  set  five  times.  Thus, \nthe \"%  Correct\"  columns denote  the mean and standard deviation  of the  classifier \nperformance with respect  to both training and testing  variations.  By  duty cycling \nthe bias to the  network  and buffers,  the  chip  dissipates less  than  200n W  power for \na  nominal heart rate of 120  beats/minute during generalisation. \n\n5  DISCUSSION \n\nReferring  to table  1 we  see  that the patient  3 data was  relatively difficult  to  train. \nHowever,  for  the one occasion when  training converged  generalisation performance \nwas  quite  acceptable.  Inspection  of this  patients  data showed  that  typically,  the \nmorphologies of the  two rhythms  were  very  similar.  The  choice  of output targets, \nmargins and  architecture  appear to  be  patient dependent  and possibly  interacting \nfactors.  Although larger margins make training easier for some patients they appear \nto also  introduce more variability in generalisation performance.  This may be due \nto  the  non-linearity  of the  neuron  circuit.  Further  experiments  are  required  to \noptimise  the  architecture  for  a  given  patient  and  to  clarify  the  effect  of varying \ntargets,  margins and  neuron gain.  Penalty terms could  also  be  added  to the  error \nfunction  to minimise the possibility of missed detections  of the dangerous  rhythm. \n\nThe  relatively slow  rate  of the heart  results  in  the  best  power  consumption  being \nobtained by  duty cycling  the  bias  currents  to the synapses  and the buffers.  Hence, \nthe bias settling time of the weighted current source  is the limiting factor for  reduc(cid:173)\ning power consumption further for this design.  By  modifying the  connection of the \ncurrent  source  to the synapses  using  a bypassing  technique  to reduce  transients in \n\n\f738 \n\nRiclulrd Coggins,  Marwan Jabri,  Barry Flower,  Stephen Pickard \n\nthe weighted currents,  still lower power  consumption could be  achieved. \n\n6  CONCLUSION \n\nThe  successful  classification  of  a  difficult  cardiac  arrhythmia  problem  has  been \ndemonstrated using. an analogue VLSI  neural network  approach.  Furthermore, the \nchip  developed  has  shown  very  low  power  consumption of less  than 200n W, meet(cid:173)\ning the requirements of an implantable system.  The chip  has performed well,  with \nover  90%  classification performance for  most patients studied and has proved  to be \nrobust when the real world influence of analogue QRS detection jitter is introduced \nby  a commercial implantable cardioverter  defibrillator placed in the signal path to \nthe classifier. \n\nAcknowledgements \n\nThe  authors  acknowledge  the funding  for  the  work  in  this  paper  provided  under \nAustralian Generic Technology Grant Agreement No.  16029  and thank Dr.  Phillip \nLeong  of the  University  of Sydney  and  Dr.  Peter  Nickolls  of Telectronics  Pacing \nSystems Ltd.,  Australia for  their helpful suggestions  and advice. \n\nReferences \n[Castro et.  al., 1993]  H.A.  Castro,  S.M.  Tam,  M.A.  Holler,  \"Implementation and \nPerformance of an analogue Nonvolatile Neural  Network,\"  Analogue  Integrated \nCircuits  and  Signal Processing,  vol.  4(2), pp.  97-113, September  1993. \n\n[Lin et.  al.,  1988]  D.  Lin,  L.A.  Dicarlo,  and  J .M.  Jenkins,  \"Identification  of Ven(cid:173)\ntricular Tachycardia using Intracavitary Electrograms:  analysis of time and fre(cid:173)\nquency  domain patterns,\"  Pacing  (3  Clinical Electrophysiology,  pp.  1592-1606, \nNovember  1988. \n\n[Leong,  1992]  P.H.W.  Leong,  Arrhythmia Classification  Using  Low  Power  VLSI, \n\nPhD Thesis,  University of Sydney,  Appendix B,  1992. \n\n[ Kusumoto et.  al., 1993]  K.  Kusumoto  et.  al.,  \"A  lObit  20Mhz  30mW  Pipelined \n\nInterpolating ADC,\"  ISSCC,  Digest  of Technical Papers,  pp.  62-63,  1993. \n\n[Leong and Jabri,  1992]  P.H.W.  Leong  and  M.  Jabri,  \"MATIC  - An  Intracar(cid:173)\n\ndiac  Tachycardia Classification  System\",  Pacing  (3  Clinical  Electrophysiology, \nSeptember  1992. \n\n[Coggins et.  al.,  1994]  R.J. Coggins and M.A. Jabri,  \"WATTLE: A Trainable Gain \nAnalogue VLSI Neural Network\",  NIPS6,  Morgan Kauffmann Publishers,  1994. \n\n[Jabri et.  al.,  1992]  M.A.  Jabri,  E.A.  Tinker  and  L.  Leerink,  \"MUME- A  Multi(cid:173)\nNet-Multi-Architecture Neural Simulation Environment\", Neural Network Sim(cid:173)\nulation Environments, Kluwer  Academic Publications, January,  1994. \n\n[Flower and Jabri, 1993]  B.  Flower  and  M.  Jabri,  \"Summed  Weight  Neuron  Per(cid:173)\n\nturbation:  an  O(N)  improvement over  Weight  Perturbation,\"  NIPS5,  Morgan \nKauffmann Publishers,  pp.  212-219,  1993. \n\n[Tinker,  1992]  E.A.  Tinker,  \"The  SPASM  Algorithm  for  Ventricular  Lead  Tim(cid:173)\n\ning and Morphology Classification,\"  SEDAL ICEG-RPT-016-92, Department of \nElectrical  Engineering,  University of Sydney,  1992. \n\n\f", "award": [], "sourceid": 1004, "authors": [{"given_name": "Richard", "family_name": "Coggins", "institution": null}, {"given_name": "Marwan", "family_name": "Jabri", "institution": null}, {"given_name": "Barry", "family_name": "Flower", "institution": null}, {"given_name": "Stephen", "family_name": "Pickard", "institution": null}]}