<%BANNER%>

A Low Bandwidth Pulse-Based Neural Recording System

Permanent Link: http://ufdc.ufl.edu/UFE0042419/00001

Material Information

Title: A Low Bandwidth Pulse-Based Neural Recording System
Physical Description: 1 online resource (132 p.)
Language: english
Creator: YEN,SHENG-FENG
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2011

Subjects

Subjects / Keywords: ADAPTIVE -- ADC -- AMPLIFIER -- BANDWIDTH -- COMPARATOR -- GM -- IF -- IMPLANTED -- INVITRO -- INVIVO -- NEURON -- POWER -- PULSE
Electrical and Computer Engineering -- Dissertations, Academic -- UF
Genre: Electrical and Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: This research tests for the first time in-vivo a data reduction scheme based on a modified integrate-and-fire pulse encoding for an implanted neural recording system in wireless transmission applications. Wireless transmission from implanted multi-channel recordings imposes many constraints on the system but the major constraint is bandwidth. Other constraints such as large dynamic range, low power consumption, small device size and noise robustness, are serious but can be more easily met. This neural recording system consists of a front-end hardware recording part and a back-end signal processing part. The integrate-and-fire (IF) mechanism is adopted in the analog front-end circuit design to achieve low bandwidth for data compression. The neural signal is encoded and transformed into a pulse representation. The encoded pulse representation is inherently noise robust and beneficial in wireless transmission for the signal processing in the digital back-end system. As a result, a traditional analog-to-digital converter (ADC), is not required in this neural recording application. In the digital back-end part, the system can either reconstruct the recorded pulses back to a traditional sampled continuous-time signal, and then sort neural spikes upon the reconstructed signals, or directly execute the pulse-based spike sorting algorithm in the pulse domain, even when the maximum inter-pulse interval (IPI) of the encoded pulses is in sub-Nyquist regions. In this research, we first successfully record action potentials via the UF system adopting the IF neuron circuit in the in-vivo recording. To conduct an in-vivo recording, the implanted electrode must be well placed to detect available neural signals, and the analog and digital parts of the UF system need parameter optimization and calibration. In addition, the dual system experiment, comprising the UF system and the TDT system, verifies that the UF recording system extracts quality in-vivo signals. In an in-vivo recording, the spike sorting results for these recording systems classify the same neural signals. The UF system can record 1000 muVpp high action potential signals but induces about 3.6 dB higher noise levels than the TDT recording system does. The SNR of the UF system is about 11.43 dB with a pulse rate less than 30 Kpulses/sec while the SNR of the TDT system ranges is about 15.03 dB with a bandwidth of 400 Kbits/sec. The trade-off of SNR and recording bandwidth is observed. Although the sorted spikes in the reconstructed signal are distorted, the distortion is constant throughout the recording and the error does not influence the neural signal classification. This experiment shows that the decrease of the SNR does not influence the spike sorting result. The modified UF system can reduce the wireless transmission bandwidth via three versatile neuron circuit strategies: the adaptive, leaky and refractory components of the neuron circuit form the adaptive leaky refractory integrate-and-fire (ALRIF) neuron circuit. MATLAB simulation results for all these neuron circuit models show a proof of concept. The refractory neuron circuit limits the maximum peak data bandwidth. The leaky neuron circuit filters out high-frequency noise, which further reduces bandwidth. The adaptive neuron circuit achieves more than 40% data compression compared to the simple IF neuron circuit in simulation. The idea of the adaptive neuron circuit is novel in the integration with a simple IF neuron circuit and unique for reconstruction purposes. The design, fabrication and test, of the adaptive neuron circuit are presented. Simulation of the complete ALRIF neuron circuit illustrates its performance by showing three distinctly sorted spikes in a neural simulator test.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by SHENG-FENG YEN.
Thesis: Thesis (Ph.D.)--University of Florida, 2011.
Local: Adviser: Harris, John G.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2011
System ID: UFE0042419:00001

Permanent Link: http://ufdc.ufl.edu/UFE0042419/00001

Material Information

Title: A Low Bandwidth Pulse-Based Neural Recording System
Physical Description: 1 online resource (132 p.)
Language: english
Creator: YEN,SHENG-FENG
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2011

Subjects

Subjects / Keywords: ADAPTIVE -- ADC -- AMPLIFIER -- BANDWIDTH -- COMPARATOR -- GM -- IF -- IMPLANTED -- INVITRO -- INVIVO -- NEURON -- POWER -- PULSE
Electrical and Computer Engineering -- Dissertations, Academic -- UF
Genre: Electrical and Computer Engineering thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: This research tests for the first time in-vivo a data reduction scheme based on a modified integrate-and-fire pulse encoding for an implanted neural recording system in wireless transmission applications. Wireless transmission from implanted multi-channel recordings imposes many constraints on the system but the major constraint is bandwidth. Other constraints such as large dynamic range, low power consumption, small device size and noise robustness, are serious but can be more easily met. This neural recording system consists of a front-end hardware recording part and a back-end signal processing part. The integrate-and-fire (IF) mechanism is adopted in the analog front-end circuit design to achieve low bandwidth for data compression. The neural signal is encoded and transformed into a pulse representation. The encoded pulse representation is inherently noise robust and beneficial in wireless transmission for the signal processing in the digital back-end system. As a result, a traditional analog-to-digital converter (ADC), is not required in this neural recording application. In the digital back-end part, the system can either reconstruct the recorded pulses back to a traditional sampled continuous-time signal, and then sort neural spikes upon the reconstructed signals, or directly execute the pulse-based spike sorting algorithm in the pulse domain, even when the maximum inter-pulse interval (IPI) of the encoded pulses is in sub-Nyquist regions. In this research, we first successfully record action potentials via the UF system adopting the IF neuron circuit in the in-vivo recording. To conduct an in-vivo recording, the implanted electrode must be well placed to detect available neural signals, and the analog and digital parts of the UF system need parameter optimization and calibration. In addition, the dual system experiment, comprising the UF system and the TDT system, verifies that the UF recording system extracts quality in-vivo signals. In an in-vivo recording, the spike sorting results for these recording systems classify the same neural signals. The UF system can record 1000 muVpp high action potential signals but induces about 3.6 dB higher noise levels than the TDT recording system does. The SNR of the UF system is about 11.43 dB with a pulse rate less than 30 Kpulses/sec while the SNR of the TDT system ranges is about 15.03 dB with a bandwidth of 400 Kbits/sec. The trade-off of SNR and recording bandwidth is observed. Although the sorted spikes in the reconstructed signal are distorted, the distortion is constant throughout the recording and the error does not influence the neural signal classification. This experiment shows that the decrease of the SNR does not influence the spike sorting result. The modified UF system can reduce the wireless transmission bandwidth via three versatile neuron circuit strategies: the adaptive, leaky and refractory components of the neuron circuit form the adaptive leaky refractory integrate-and-fire (ALRIF) neuron circuit. MATLAB simulation results for all these neuron circuit models show a proof of concept. The refractory neuron circuit limits the maximum peak data bandwidth. The leaky neuron circuit filters out high-frequency noise, which further reduces bandwidth. The adaptive neuron circuit achieves more than 40% data compression compared to the simple IF neuron circuit in simulation. The idea of the adaptive neuron circuit is novel in the integration with a simple IF neuron circuit and unique for reconstruction purposes. The design, fabrication and test, of the adaptive neuron circuit are presented. Simulation of the complete ALRIF neuron circuit illustrates its performance by showing three distinctly sorted spikes in a neural simulator test.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by SHENG-FENG YEN.
Thesis: Thesis (Ph.D.)--University of Florida, 2011.
Local: Adviser: Harris, John G.

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2011
System ID: UFE0042419:00001


This item has the following downloads:


Full Text

PAGE 1

ALOWBANDWIDTHPULSE-BASEDNEURALRECORDINGSYSTEMBySHENG-FENG(STEVE)YENADISSERTATIONPRESENTEDTOTHEGRADUATESCHOOLOFTHEUNIVERSITYOFFLORIDAINPARTIALFULFILLMENTOFTHEREQUIREMENTSFORTHEDEGREEOFDOCTOROFPHILOSOPYUNIVERSITYOFFLORIDA2011

PAGE 2

c2011Sheng-Feng(Steve)Yen 2

PAGE 3

I dedicatethisdissertationtomylovingparentsforalloftheirlovingsupportduringthe researchprocess.Mymother'sselesssacrices,andgenerousnancialsupport,kept mefromjeopardizingmyacademicdegree.Specialthankstomyadvisor,Dr.Harris, whoofferedmearesearchopportunityandhelpedmerevampthisdissertation. 3

PAGE 4

ACKNOWLEDGMENTS Iwouldliketoexpressmygratitudetoallthepeoplewhomademyresearchpossible.AspecialthankyoutoProfessorJohnHarris,forallowingmetoworkonafascinatingprojectoflowbandwidthpulse-basedneuralrecordingsystem.IamespeciallygratefulforhisencouragementwhileIwasateachingassistant.Inaddition,IappreciatethefreedomandsupportIhadduringmyresearch.Dr.Harris'experienceandexpertisefacilitatedmystudiesduringmydoctoralprogram.Allmycommitteemembers,Dr.Harris,Dr.Principe,Dr.CarneyandDr.Fox,providedinvaluablefeedbackthroughoutmyresearch.Theirinsights,comments,questionsandexpertise,helpedshapedissertation.IwanttothankDr.Zhou,apostdoctoralresearcherinDr.Carney'sgroup,forallofhishardworkinobtainingtheneuralrecordingsfromtrainedrats,aswellasperformingtheimplantationsurgery,whichenabledmetosetuptheentireexperimentalapparatus.Dr.Zhoualsohelpedmeconductthein-vivorecordingexperimentandtestedmyrecordingsystemhardwarealongwiththeTucker-DavisTechnologies(TDT)recordingsystem.Manylabmateshaveparticipatedindiscussionsaboutmyresearchandprovidedhelpfulinsightand/orquestions.SpecialthanksgotoJieJessieXu,oneofthepioneersoflowbandwidthpulse-basedrecordingsystems,whoprovidedmewithmanyinsights.Lastbutnotleast,Iwouldliketoacknowledgemyfundingsources.MyresearchwassupportedthroughteachingandresearchassistantpositionsintheDepartmentofElectricalandComputerEngineeringwithadditionalfundingfromaNIHgrant.IwouldalsoliketoacknowledgemycollaboratesandsupporterswhileearningmyM.S.degreeatU.C.L.A.TheyhelpedmemastertheoutstandingchallengesIfaced.IwouldliketoexpressmygratitudetoDr.Seong-JinKim(Post-Doc),Werayut(Yut)Srituravanich,Shih-Kang(Scott)Fan,Yen-WenLu,Chen-Kang(Clint)Yangaswellasthekindestcouple,Ya-JiTasiandTao-YiWang.Thankstotheseextraordinarypeople,IwasabletoovercomethehopelessnessIoftenfacedduringmyexperienceatU.C.L.A. 4

PAGE 5

WhileatU.C.L.A,Iexperiencedsomeofthedeepestfrustrationsinmylife.Theaforementionedfriendsservedasatoweroflightinatimeofdarkness.Therefore,Iamgratefultothemforsupportingmeduringthosetwoyears(20012003)atU.C.L.A.Inconclusion,thecompletionofmyPh.D.wasasolemn,stirring,andheroicjourney. 5

PAGE 6

TABLEOFCONTENTS page ACKNOWLEDGMENTS .................................. 4 LISTOFTABLES ...................................... 8 LISTOFFIGURES ..................................... 9 ABSTRACT ......................................... 17 CHAPTER 1INTRODUCTION ................................... 19 1.1WiredNeuralRecordingSystem ....................... 20 1.2ConventionalWirelessNeuralRecording ................... 20 1.3BiphasicIFEncodingNeuronCircuit ..................... 23 1.4FeatureExtractionMethod .......................... 26 1.5ResearchGoal ................................. 27 1.6DissertationOverview ............................. 28 2FRONT-ENDHARDWAREDESCRIPTIONANDINTEGRATIONTOTHEEXISTINGSYSTEM ....................................... 30 2.1UFSystemOverview .............................. 30 2.2Bio-amplier .................................. 32 2.3V-IConverter .................................. 34 2.4BiphasicIFCircuit ............................... 38 2.5USBRecordingBoard ............................. 39 2.6BenchTopTest ................................. 40 2.6.1ParallelRecordingPlatformDesign .................. 40 2.6.2PlatformCongurationandSetup ................... 42 2.6.3UFSystemNoise ............................ 44 2.7AnalogFront-endSystemIntegration ..................... 47 3BACK-ENDSOFTWARETESTANDCHARACTERIZATION .......... 51 3.1Reconstruction ................................. 51 3.2ConventionalSpikeSorting .......................... 53 3.3Pulse-basedSpikeSorting .......................... 56 3.3.1SystematicTrainingforTemplateGeneration ............. 59 3.3.2NeuralSimulatorTest ......................... 62 3.4FullSystemConguration ........................... 66 4FULLNEURALRECORDINGSYSTEMOPERATION .............. 68 4.1AcuteIn-vivoExperiment ........................... 68 4.2DataCollectionPriortoCalibration ...................... 73 6

PAGE 7

4.3UFSystemCalibration ............................. 75 4.4FurtherPossibleImprovements ........................ 85 5NEURONCIRCUITDESIGNFORFURTHERBANDWIDTHREDUCTION ... 89 5.1Background ................................... 89 5.2RefractoryIntegrate-and-Fire(RIF)Neuron ................. 90 5.3LeakyIntegrate-and-Fire(LIF)Neuron .................... 92 5.3.1LeakyComponentImplementation .................. 92 5.3.2LeakyNeuronModelSimulation .................... 95 5.4AdaptiveIntegrate-and-Fire(AIF)Neuron .................. 95 5.4.1PerformancewithDifferentSNR .................... 101 5.4.2PerformancewithSlowDCDrift .................... 103 5.4.3PerformancewithLargeDynamicRange ............... 106 5.5AdaptiveNeuronCircuit ............................ 106 5.5.1AdaptiveCircuitDesign ........................ 107 5.5.2ImprovedAdaptationMechanism ................... 109 5.6AdaptiveLeakyRefractoryIntegrate-and-Fire(ALRIF)Neuron ...... 113 5.7ALRIFNeuronBenchTopTest ........................ 113 5.8Discussion ................................... 117 5.8.1WideDynamicRangeAdaptiveComponentDesign ......... 117 5.8.2ALRIFNeuronSystemFundamentalLimitation ........... 120 5.8.3Calibration ................................ 122 5.8.4PulseDensityAnalysis ......................... 123 6CONCLUSIONS ................................... 125 REFERENCES ....................................... 128 BIOGRAPHICALSKETCH ................................ 132 7

PAGE 8

LISTOFTABLES Table page 1-1DataCompressionTechnologies .......................... 27 2-1RecordingresultsofSNRfordifferenttestingsetup ................ 47 3-1DetectionresultsofrecordingfromFigure 2-13 setting .............. 56 3-2PatternrecognitionresultsfromFigure 2-13 setting ................ 56 4-1Theperformanceofthein-vivorecordingwiththeTDTsystemduringtheacuteexperiment ...................................... 83 4-2Theperformanceofthein-vivorecordingwiththeUFsystemduringtheacuteexperiment ...................................... 84 4-3Thedataanalysisoftheacutein-vivoexperimentfortheTDTandtheUFrecordingsystems ........................................ 84 8

PAGE 9

LISTOFFIGURES Figure page 1-1Characteristicsofextracellularneuralsignals. .................. 20 1-2Currentpopularneuralrecordingsetup:theanimalundertestingstaysinacagewithalongwireattachedtotheelectrodes,whichisimplantedintoitsbrain,andallthewaytodistantrecordinginstruments(left).Azoom-inviewrecordselectrodesimplantedintotheanimal'sbrainthroughsurgery(right). 21 1-3WirelessNeuralRecordingScheme.Theneuroncircuit(circledinred)isaneuralinspiredanalogcircuitwhichcanencodedatainordertoreducebandwidthwithasimplerdesignthananADC. ................. 22 1-4UFoverallneuraldatareductionapproaches. .................. 23 1-5Anexampleofaninputsignalandit'sbiphasicrepresentation. ........ 24 1-6Blockdiagramofthebiphasicencodingwithintegrate-and-re(IF)neuron. .. 25 1-7Blockdiagramofbiphasicencodingwithleakyintegrate-and-re(LIF)circuit. 26 2-1Theexistingneuralrecordingsystem:bio-amplier,V-Iconverter,andbiphasicintegrate-and-re(IF)recordingsystemwiththerefractorycomponentandtheleakycomponent. ............................... 31 2-2Schematicofabio-amplier. ........................... 32 2-3Transistorleveldesignofoperationtransconductanceamplierusedinthebio-amplier. .................................... 33 2-4SchematicofaV-Iconverter. ........................... 34 2-5Transistorleveldesignofoperationtransconductanceamplier(OTA2). ... 35 2-6TheschematicofthenewV-Iconverterandintegrator. ............. 37 2-7SchematicofthecomparatorcircuitusedinthebiphasicIFneuron. ...... 38 2-8TheParallelRecordingConcept. ......................... 40 2-9Abenchtoptestingsetupinparallelrecordingwithaneuralsimulatorasthesignalsource. .................................... 42 2-10A20secsegmentoftheTDTsystemrecordingrepresentingatypicalqualityneuralsignalrecording.TheSNRisabout23.5dB. ............... 42 2-11Aphotooftheparallelrecordingexperimentsetupforabench-toptestwithaneuralsimulator. ................................. 43 9

PAGE 10

2-12A30secrecordingbeforethetestingimprovementisinvolved.(A)ThedirectrecordingisfromtheTDTrecordingsystem.ThelowSNRisduetothenoisecoupledfromtheUFrecordingsystem(B)Testpointismeasuredattheinputofthebio-amplier.ThelowSNRisduetothegroundloopandradiationnoiseintheUFsystem.(C)Testpointismeasuredattheoutputofthebio-amplierintheUFrecordingsystem.SNRisbetterthan(B)becausethebuilt-inlterfunctioninthebio-amplier. ............................ 44 2-13Aphotooftheimprovedparallelrecordingexperimentsetupforabenchtoptestwithaneuralsimulator. ............................ 45 2-14Aphotoofacloseviewoftheimprovedparallelrecordingexperimentsetupforabenchtoptestwithaneuralsimulator. ................... 46 2-15The20secrecordingfromtheimprovedparallelrecordingsetupdirectlyrecordedfromtheTDTrecordingpath(top),andreconstructedsignalsrecordedfromtheUFrecordingsystem(bottom). ..................... 46 2-16Thesystemcongurationofanadaptiveleakyrefractoryintegrate-and-re(ALRIF)neuroncircuit. ............................... 47 2-17ThePCBdesignoftheanalogfront-endrecordingsystem:(a)thepowersupplyboardprovidesdifferentDCvoltagesforthewholesystem,(b)themainboardofencodingcircuitsincluding(1)theamplierchipsand(2)theV-Iconverterandtherefractoryleakyintegrate-and-reneuroncircuitchip,(c)theadaptivecomponentcircuitboardand(d)theUSBinterfaceboard. .. 48 2-18AphotographoftheUFanalogfront-endrecordingsystem. .......... 50 3-1Alignmentofthreesignals:(top)biphasicpulsetrainsrecordedthroughtheUSBboard,(middle)theoriginalsignalsrecordedthroughtheTDTsystem,and(bottom)reconstructedsignalsattheback-endoftheUFrecordingsystemtestedwiththesetting 2-13 ........................... 53 3-2ThespikesortingresultforthereconstructedsignalsbasedontemplatematchingandaPCAclustercuttingalgorithm. ....................... 54 3-3AcomparisonofspikesortingresultsfromtheneuralsimulatorbetweenTDTdirectrecordingandreconstructedsignals. ................... 57 3-4Asegmentofrecordingdatafromtheneuralsimulator. ............. 60 3-5PulsesgeneratedfromasimpleIFencodermodelandpotentialpulsesignaturesaremarkedinred. ................................. 60 3-6Truepositiverate(TPR)v.s.differentconvolvingwithdifferentneuronsforthecaseofasimpleIFencodingfeatureextractor. ............... 61 10

PAGE 11

3-7AveragedtemplatesforeachneuronforthecaseofasimpleIFencodingfeatureextractor. .................................. 61 3-8PotentialpulsesignaturesaremarkedinredinthetestingphaseforthecaseofasimpleIFencodingfeatureextractor. .................... 62 3-9PulsesgeneratedfromaLIFencodermodelandpotentialpulsesignaturesaremarkedinred. ................................. 64 3-10Truepositiverate(TPR)v.s.differentconvolvingwithdifferentneuronsforthecaseofaLIFencodingfeatureextractor. ................... 64 3-11AveragedtemplatesforeachneuronforthecaseofaLIFencodingfeatureextractor. ...................................... 65 3-12PotentialpulsesignaturesaremarkedinredinthetestingphaseforeachneuronforthecaseofaLIFencodingfeatureextractor. ............ 65 3-13ThecongurationoftheUFneuralrecordingsystem. .............. 66 4-1Thein-vivorecordingpreparation:(a)theUFanalogfront-endrecordingboardconnectstheDB-25adapterfromtheelectrodebufferandtheUSBporttothePCand(b)theratrsthasanelectrodearrayimplantedintoitsbrainfortheneuralsignaldetectionandabufferhookedonthetopholdssignalstotheelectronicsintheback. ....................... 68 4-2Thetestingconceptforthein-vivorecordingintheacuteexperiment. ..... 69 4-3Thewholeacutein-vivoexperimentsetup:theratundertest(top)isconnectedtoeithertheUFrecordingsystem(center)ortheTDTrecordingsystem(bottom).ThePC(left)hastheMATLABprogramtocontrolbothsystemstorecordanddisplayrecordings. .............................. 69 4-4Aphotoshowstherathasbeenimplantedwithanelectrodearraydowntothehippocampusareafortheneuralsignalrecordingintheacuteexperiment. 70 4-5Aphotoshowstheratisanesthetized.Itwouldbekeptasleepbythe1.5%isouranceandplacedonahotplatethroughouttheacuteexperiment. ... 70 4-6TheMATLABinterfaceprogrammonitorsthetotal16channelsfromtheimplantedelectrodearrayfortheneuronringactivityineachchannel. .......... 71 4-7TheUFanalogfront-endsystemboardextractstheneuralsignalsfromtheheadstagethroughaDB-25adapter(leftside)andpassestheoutputpulsestotheback-endthroughaUSBcableattached(rightside). ........... 72 4-8Thein-vivorecordingsetup:theratwiththeelectrodeimplantedinitsbrainisanesthetizedforconvenienceandhookedtotherecordingsystem. ..... 73 11

PAGE 12

4-9Thein-vivorecording:eithertheTDTortheUFrecordingsystemconnectstotheelectrodesimplantedintotherattorecordactionpotentials. ...... 73 4-10A33secondlongin-vivorecordingviatheUFrecordingsystem. ....... 74 4-11A30secondlongin-vivorecordingviatheTDTrecordingsystem. ...... 74 4-12AcloserviewintoFigure 4-10 .Twodistinctactionpotentialsarerevealed. .. 76 4-13AcloserviewintoFigure 4-11 .Twodistinctactionpotentialsarerevealed. .. 76 4-14AcloserviewofthespikesortingresultfortheTDTsystemrecording.Twoclassesofactionpotentialsaresortedindifferentcolors. ............ 77 4-15AcloserviewofthespikesortingresultfortheUFsystemrecording.Twoclassesofactionpotentialsaresortedindifferentcolors. ............ 77 4-16Asummaryofpile-upclassiedsortsofactionpotentialsforbothrecordingsfromtheUFsystemandtheTDTsystem. .................... 78 4-17TheUFrecordingsystemiscalibratedthroughtheneuralsimulatortest. ... 78 4-18Aphotoofthein-vivorecordingconductedwiththeUFsystem. ........ 79 4-19Aphotoofthein-vivorecordingconductedwiththeTDTsystem. ....... 79 4-20A20secondslongoftherstin-vivorecordingwiththeTDTsystem. .... 80 4-21A27secondslongofthein-vivorecordingthroughtheTDTsystem5minutesaftertherecordinginFigure 4-20 ......................... 80 4-22At32secondslongofthein-vivorecording,theUFsystemrecorded5minutesaftertherecordinginFigure 4-21 ......................... 81 4-23At32secondslongofthein-vivorecording,theUFsystem5minutesaftertherecordinginFigure 4-22 ........................... 81 4-24ApartoftheTDTdirectrecordingsignals(bottomtrace)showsthesortedspikesaremarkedanddisplayedseparately(toptrace). ............ 82 4-25ApartofthereconstructedsignalsfromtheUFsystemback-end(bottomtrace)showsthesortedspikesaremarkedanddisplayedseparately(toptrace). ........................................ 82 4-26ThespikesortingresultsforboththedirectrecordingfromtheTDTsystem(left)andthereconstructedsignalsfromtheUFsystem(right).Bothresultsshowthesimilarsinglesort. ............................ 83 12

PAGE 13

4-27A30mslongtransientanalysisfor1kHzsinewaveinputsignaltotheexistingsysteminFigure 2-1 .(top)vout-representsthenegativepulsetrain,(middle)vout+representsthepositivepulsetrainand(bottom)vout2representstheintegrate-and-reresponseonthemembranevoltage. ............. 85 4-28Proposedfullydifferentialleakyrefractoryintegrate-and-reencoding(FD-LRIF)analogfront-endcircuitsystem. .......................... 86 4-29A30mslongtransientanalysisfor1kHzsinewaveinputsignaltotheexistingsysteminFigure 2-1 .(top)vout gmrepresentstheintegrate-and-reresponseonthemembranevoltage,(middle)vout+representsthepositivepulsetrainand(bottom)vout-representsthenegativepulse. ................ 87 4-30Thetransistorleveldesignforthefullydifferentialamplierdesignwithacommon-modefeedbackcircuit. ......................... 88 5-1Asimulationofpulseratev.s.inputsinewaveamplitudefordifferentrefractoryperiodsettingsofarefractoryneuron. ...................... 90 5-2Transistorleveldesignofrefractorybufferusedintheintegrated-and-recircuit. ........................................ 91 5-3Schematicofaleakyintegrate-and-re(LIF)circuit. ............... 92 5-4Aschematicoftheleakycomponentcircuitdesign. ............... 93 5-5Afamilyofcurrent-to-voltagecurvesfordifferentlinearinputvoltagerangesinthesub-thresholdoperationmodefortheleakycomponent. ......... 94 5-6Asimulationofpulseratesv.s.inputsinewaveamplitudesfordifferentbiascurrentdesignsettingsinaleakycomponentwithalinearinputvoltagerangeof0.5V. ....................................... 95 5-7Asimulationdemonstratestheoperationofanadaptiveneuronmodel. .... 96 5-8Pulseratesv.s.inputsignalamplitudeswithan1KHzinputsinwaveanddifferentincrementalvoltagethresholdadaptation(Vthstep)foreachpulseintheproposedadaptiveneuronmodel. ..................... 97 5-9Acomparisonofpulseratesgeneratedfromdifferentthresholdvoltagesbetweenanadaptiveneuronandasimpleneuronforasinewaveinputwithanamplitudeof500V. ...................................... 98 5-10Asegmentofrecordedactionpotentialsignalsusedfortestingdifferentneuronmodelsinthischapter.Twoactionpotentialsaremarkedinred. ........ 99 5-11AcomparisonofSERv.s.pulseratesbetweenanadaptiveneuron(blue)andasimpleneuron(red)forapplyingdifferentthresholdvoltages. ...... 99 13

PAGE 14

5-12AcomparisonofSERv.s.pulseratesbetweenanadaptiveneuron(blue)andasimpleneuron(red)forapplyingdifferentscalingfactorstotheactionpotentials. ...................................... 100 5-13Anoverlapplottingofrecordedandreconstructedsignals(top)andpulsesgeneratedbyanadaptiveneuron(bottom). ................... 100 5-14Anoverlapplottingofrecordedandreconstructedsignals(top)andpulsesgeneratedbyasimpleneuron(bottom). ..................... 101 5-15Anoverlapplottingofrecordedandreconstructedsignalswithactionpotentialsmarkedinred(top)andpulsesgeneratedbyanadaptiveneuron(bottom). .. 102 5-16Anoverlapplottingofrecordedandreconstructedsignalswithactionpotentialsmarkedinred(top)andpulsesgeneratedbyasimpleneuron(bottom). ... 102 5-17AcomparisonofSERv.s.pulseratesbetweenanadaptiveneuron(blue)andasimpleneuron(red)forapplyingdifferentthresholdvoltages. ...... 103 5-18Acomparisonofpulseratesgeneratedwithdifferentthresholdvoltagesbetweenanadaptiveneuronandasimpleneuronfortherecordedactionpotentialsassociatedwitha90Hz,100Vsinewave. .................... 103 5-19Anoverlapplottingofrecorded(blue)andreconstructed(red)signalswheretherstactionpotentialisscaledbythreetimes(top)andpulsesgeneratedbyasimpleneuron(bottom). ........................... 104 5-20Anoverlapplottingofrecorded(blue)andreconstructed(red)signalswheretherstactionpotentialisscaledbythreetimes(top)andpulsesgeneratedbyanadaptiveneuron(bottom). ......................... 104 5-21AcomparisonofSERv.s.pulseratesbetweenanadaptiveneuron(blue)andasimpleneuron(red)forapplyingdifferentinitialthresholdvoltageswiththeinputtestingdatawheretherstactionpotentialisscaledbythreetimesforthelargedynamicrangetest. ......................... 105 5-22Acomparisonofpulseratesgeneratedwithdifferentinitialthresholdvoltagesbetweenanadaptiveneuronandasimpleneuronfortheinputtestingdatawheretherstactionpotentialisscaledbythreetimes. ............ 105 5-23Thecircuitcongurationofanadaptivecomponentcombinedtothedevelopedneuralrecordingsystem. .............................. 106 5-24Theconceptualcongurationofthecircuitdesignforanadaptivecomponent. 107 5-25Theschematicdesignofthebiphasicneuroncircuitwithanadaptivecomponentandarefractorycomponent. ............................ 108 14

PAGE 15

5-26Asimulationresultofthebiphasicneuroncircuitwithanadaptivecomponentandarefractorycomponent:(A)asingletonesinewaveinput,(B)theoutputofthebio-amplier,(D)apulsetraingeneratedinthepositivechannel,(E)apulsetraingeneratedinthenegativechannel,(F)thedynamicpositivethresholdvoltagerespondingtotheinputpositivechannelpulsetrain,and(G)thedynamicnegativethresholdvoltagerespondingtotheinputnegativechannelpulsetrain. ................................. 109 5-27Circuitdiagramoftheadaptivecomponents. ................... 110 5-28Thediagramshowstheconceptoftheimprovedadaptivecomponentmechanism.Theplotinredrepresentstheadaptivethresholdvoltageandtheplotinbluerepresentstheintegrationprocessonthecapacitor.Pulsesaregeneratedwhentheintegratedvoltagereachestheinstantthresholdvoltage. ...... 111 5-29TheSPICEsimulationoftheproposedlow-bandwidthadaptivebiphasicneuronusingtheadaptivecomponent 5-27 (b). ...................... 111 5-30AcomparisonofSERv.s.pulseratesbetweenanadaptiveleakyneuron(blue)andasimpleneuronwithaleakycomponent(red)fordifferentbiascurrentsfrom10nA100nAandtheleakycomponenthasalinearinputvoltagerangeof0.5V. .................................... 112 5-31Anoverlapplottingofrecorded(blue)andreconstructed(red)signals.Signalscannotbereconstructed(red)perfectlywhenpulsesarenotsufcient(top)andpulsesgeneratedbyaleakysimpleneuron(bottom). ........... 114 5-32Anoverlapplottingofrecorded(blue)andreconstructed(red)signals.Signalscannotbereconstructed(red)perfectlywhenpulsesarenotsufcient(top)andpulsesgeneratedbyanadaptiveleakyneuron(bottom). ......... 114 5-33Abench-toptestingsetup:anindependentsignalsourcepumpsanalogsignalsintothecircuitboardundertesting,ampliedoutputcouldbemonitoredintheoscilloscopeandencodedpulsesarerecordedthroughtheUSBinterfaceintothecomputerfortheback-endsignalreconstruction. ............ 115 5-34TheimprovedadaptivemechanismemployedintoanALRIFneuroncircuitmanifestsontoaoscilloscopedisplayforabenchtoptestwithasinewaveinput:topandbottomtracesrepresentboththepositiveandnegativeadaptivethresholdvoltagesrespectively.Themiddletracerepresentsthemembranevoltagetoshowtheintegrate-and-reprocessundertheadaptationcondition. 116 5-35Anoscilloscopedisplayshowsthethresholdvoltageadaptationcorrespondingtoeachintegrate-and-recycleonthemembranevoltagefortheALRIFneuroncircuit. ........................................ 116 15

PAGE 16

5-36Top:recordedbiphasicpulsesfortheALRIFanalogfront-endrecordingsystemwiththeneuralsimulatorinput.Bottom:reconstructedsignalsintheback-end.Threeseparateandclearactionpotentialsarerevealed. ............ 117 5-37Theback-endspikesortingprocessisexecutedinSpike2TM.Threekindsofsortedspikesaremarkedindifferentcolors. ................... 118 5-38Threedistinctpile-upwaveformsofsortedspikesfromthereconstructedsignalsrecordedfromtheALRIFanalogfront-endrecordingsystem. ..... 118 5-39Theschematicofthewidedynamicrangeadaptivecomponentdesign. .... 119 5-40TransistorleveldesignforthelargerlinearityOTAdesign. ........... 120 5-41ThesimulationofthetransientresponseoftheALRIFneuroncircuitwithawiderdynamicrangeadaptivecomponentdesign:(top)vout)]TJ /F1 11.955 Tf 7.08 1.8 Td[(gmrepresentsthemembranevoltageforadaptedintegrate-and-reresponse,(middle)vref-representstheadaptivethresholdvoltageforthenegativechanneland(bottom)vref+representstheadaptivethresholdvoltageforthepositivechannel. ... 121 16

PAGE 17

AbstractofDissertationPresentedtotheGraduateSchooloftheUniversityofFloridainPartialFulllmentoftheRequirementsfortheDegreeofDoctorofPhilosopyALOWBANDWIDTHPULSE-BASEDNEURALRECORDINGSYSTEMBySheng-Feng(Steve)YenMay2011Chair:JohnG.HarrisMajor:ElectricalandComputerEngineering Thisresearchtestsforthersttimein-vivoadatareductionschemebasedonamodiedintegrate-and-repulseencodingforanimplantedneuralrecordingsysteminwirelesstransmissionapplications.Wirelesstransmissionfromimplantedmulti-channelrecordingsimposesmanyconstraintsonthesystembutthemajorconstraintisbandwidth.Otherconstraintssuchaslargedynamicrange,lowpowerconsumption,smalldevicesizeandnoiserobustness,areseriousbutcanbemoreeasilymet. Thisneuralrecordingsystemconsistsofafront-endhardwarerecordingpartandaback-endsignalprocessingpart.Theintegrate-and-re(IF)mechanismisadoptedintheanalogfront-endcircuitdesigntoachievelowbandwidthfordatacompression.Theneuralsignalisencodedandtransformedintoapulserepresentation.Theencodedpulserepresentationisinherentlynoiserobustandbenecialinwirelesstransmissionforthesignalprocessinginthedigitalback-endsystem.Asaresult,atraditionalanalog-to-digitalconverter(ADC),isnotrequiredinthisneuralrecordingapplication.Inthedigitalback-endpart,thesystemcaneitherreconstructtherecordedpulsesbacktoatraditionalsampledcontinuous-timesignal,andthensortneuralspikesuponthereconstructedsignals,ordirectlyexecutethepulse-basedspikesortingalgorithminthepulsedomain,evenwhenthemaximuminter-pulseinterval(IPI)oftheencodedpulsesisinsub-Nyquistregions. 17

PAGE 18

Inthisresearch,werstsuccessfullyrecordactionpotentialsviatheUFsystemadoptingtheIFneuroncircuitinthein-vivorecording.Toconductanin-vivorecording,theimplantedelectrodemustbewellplacedtodetectavailableneuralsignals,andtheanaloganddigitalpartsoftheUFsystemneedparameteroptimizationandcalibration.Inaddition,thedualsystemexperiment,comprisingtheUFsystemandtheTDTsystem,veriesthattheUFrecordingsystemextractsqualityin-vivosignals.Inanin-vivorecording,thespikesortingresultsfortheserecordingsystemsclassifythesameneuralsignals.TheUFsystemcanrecord1000Vpphighactionpotentialsignalsbutinducesabout3.6dBhighernoiselevelsthantheTDTrecordingsystemdoes.TheSNRoftheUFsystemisabout11.43dBwithapulseratelessthan30Kpulses/secwhiletheSNRoftheTDTsystemrangesisabout15.03dBwithabandwidthof400Kbits/sec.Thetrade-offofSNRandrecordingbandwidthisobserved.Althoughthesortedspikesinthereconstructedsignalaredistorted,thedistortionisconstantthroughouttherecordingandtheerrordoesnotinuencetheneuralsignalclassication.ThisexperimentshowsthatthedecreaseoftheSNRdoesnotinuencethespikesortingresult. ThemodiedUFsystemcanreducethewirelesstransmissionbandwidthviathreeversatileneuroncircuitstrategies:theadaptive,leakyandrefractorycomponentsoftheneuroncircuitformtheadaptiveleakyrefractoryintegrate-and-re(ALRIF)neuroncircuit.MATLABsimulationresultsforalltheseneuroncircuitmodelsshowaproofofconcept.Therefractoryneuroncircuitlimitsthemaximumpeakdatabandwidth.Theleakyneuroncircuitltersouthigh-frequencynoise,whichfurtherreducesbandwidth.Theadaptiveneuroncircuitachievesmorethan40%datacompressioncomparedtothesimpleIFneuroncircuitinsimulation.TheideaoftheadaptiveneuroncircuitisnovelintheintegrationwithasimpleIFneuroncircuitanduniqueforreconstructionpurposes.Thedesign,fabricationandtest,oftheadaptiveneuroncircuitarepresented.SimulationofthecompleteALRIFneuroncircuitillustratesitsperformancebyshowingthreedistinctlysortedspikesinaneuralsimulatortest. 18

PAGE 19

CHAPTER1INTRODUCTION Neuralrecordingisfundamentaltocomputationalneuroscienceresearch.Thebasicinformationprocessingunitinthebrainistheneuron.Neuralsignalspropagateintheformofelectricalpulses,calledactionpotentialsorspikes,totransmitinformation.TheeldofneurophysiologyoriginatedfromtheworkofluminariessuchasHodgkinandHuxley,whodevelopedthevoltageclamp,andcreatedtherstmathematicalmodeloftheactionpotential.Neurophysiologyemphasizesdescriptionsoffunctionalandbiologicallyrealisticneuronsandtheirphysiologyanddynamics. Forthepurposeofstudyingneuralinformationprocessing,simultaneousrecordingfromalargegroupofneuronsisrequiredfromlivingsubjects.Inmoderntechnologies,microelectrodesimplantedintothebrain,areusedtorecordtheseextracellularelectricalspikes[ 1 9 ].Theserecordingsarecalledneuralrecordings,becausetheelectricalvoltagesgeneratedbyneuronsaremeasured. Itisimportanttoconsidertheprimaryrequirementsofqualityneuralrecording:(1)long-termrecordingsfromanimplantforweeks,monthsandevenyears.(2)uptohundredsofchannelsofelectrodesand(3)highsignal-to-noiseratio(SNR)signals.Currentneuralrecordingsystems,stillhavealotofroomtoimprovetheirrecordingcapacityandsignalquality. Inthiswork,extracellularneuralsignals,alsocalledactionpotentials,arenecessarytorecordthisinformation.Theseneuralsignalsarecharacterizedbyamplitudesrangingfrom50Vto500V,withnoiselevelsofabout20Vrms,andlastingabout1to2ms,asillustratedinFigure 1-1 .Inordertoinferthefunctionofneuralsystem,thenecessarystepistoperformspikesorting.Spikesortingisrequiredwhenrecordingmultipleneuralsignals,anditiscriticaltoclassifyingactionpotentialsredfromdifferentneuroncells.Theuniquefeatureofneuralsignals,isthatneuronsareinactivemostofthetime,and 19

PAGE 20

Figure1-1. Characteristicsofextracellularneuralsignals. sparselygenerateactionpotentials,whicharethekeyforincorporatingcompressionintotheneuralsignalrecordingsystem. 1.1WiredNeuralRecordingSystem Thecurrentwiredneuralrecordingsystems,recorddatafromimplantedmicroelectrodesusingbundlesofwires[ 10 12 ].Becauseimplantedelectrodesareconnectedtoexternalelectronicsthroughwires,subjectmovementisseriouslyrestricted,asshowninFigure 1-2 .Thewiredneuralrecordingsystemshavemanyrisksinlong-term,multi-channelrecording.Forinstance,therstrisk,isaninfectioninducedfromwirespassingthroughtheanimalskin.Thesecondriskisthelimitednumberofwiresavailablefromcommutators,whichcanreducetanglingandtorqueappliedtowires.Thethirdriskistheexternalnoiseandinterferingsignals,whicheasilycoupletowirestocorruptweakneuralsignals(<500V).Inpractice,awiredrecordingsystemisnotpracticalinscalinghundredsofchannelswithoutcausingtoomuchmechanicalconstrainttothesubjectundertesting.Theseissueshaveforcedresearcherstoconsiderwirelesstransmission,ratherthanwiredtransmissioninneuralrecordingsystems. 1.2ConventionalWirelessNeuralRecording Wirelessneuralrecordingsystemshavebeenbuiltfromdiscretemodulessincethe1990s.Morerecentwirelesssystems,haveutilizedintegratedcircuits(ICs)for 20

PAGE 21

Figure1-2. Currentpopularneuralrecordingsetup:theanimalundertestingstaysinacagewithalongwireattachedtotheelectrodes,whichisimplantedintoitsbrain,andallthewaytodistantrecordinginstruments(left).Azoom-inviewrecordselectrodesimplantedintotheanimal'sbrainthroughsurgery(right). amplicationoftheneuralsignals[ 13 15 ].Recently,digitalmodulationhasbecomepopularinwirelessneuralrecording[ 16 ][ 17 ].Abattery-poweredneuralrecordingsystemtransmittedneuralsignalsfromananimalusinganalogFMmodulation,withanoff-chipinductor[ 18 ].Analog-to-digitalconvertors(ADCs),areneededtodigitizeampliedneuralactionpotentials,beforemodulationandtransmission,inconventionalwirelessneuralrecording.Harrison[ 19 ]developedawirelessneuralrecordingsystemthatinterfacedtoa100-channelmicroelectrodearray.Thesesystemsplacestrongconstraintsonwirelesstransmission,duetotheneedtoscaleuptohundredsofchannelswithfullneuralsignals,whichcanbespikesortedintheback-end.Forexample,whentransmittingrawdatafrom100channelsata25KHzsamplingratewith8-bitofresolution,theseconventionalneuralrecordingsystemswillneedmorethan20Mbpsbandwidth.Thisbandwidthexceedsthelimitof500Kbpsavailabletotransmitthroughskin.Thisbandwidthcanbeexceeded,butonlyatagreatpowercost,sinceexcesspowerdissipationcausestissuedamageandshorterbatterylife.Therefore, 21

PAGE 22

Figure1-3. WirelessNeuralRecordingScheme.Theneuroncircuit(circledinred)isaneuralinspiredanalogcircuitwhichcanencodedatainordertoreducebandwidthwithasimplerdesignthananADC. datareductionisrequiredandindeed,thereisatradeoffbetweendatareductionandtheamountofinformationretained.Inaddition,differentdatareductionstrategiesaffecttheback-endsignalprocessingtasks,suchasreconstructionandspikesorting,asillustratedinFig 1-3 .Thetraditionalscheme,istohaveanamplierandanADCfortheanalogfront-end.TheproposedideaistocreateasimplerencodingcircuittoreplacetheADCanddecodetheinformationintheback-end.Theencodingcircuitdevelopedisinspiredfromarealneuroncellactionmodel.Therefore,itiscalledaneuroncircuit,whichgeneratespulsesjustasneuroncellsreactionpotentials. Becausebandwidthisthefundamentalroadblockinwirelessneuralrecording,wecantakeadvantageofthepropertythatspikesaresparseinneuralrecording.TheUniversityofFloridaproposedstrategiesthatfocusoninformationforspikeswaveforms,ratherthanforincreasednoise.Thisprovidessignicantdatareduction,comparedtoconventionallysamplingandquantizingtheentirerecording.TheUniversityofFloridahasdevelopeddatareductionstrategiesfordifferentapplications,allofwhichcanbeimplementedinlow-poweranalogVLSI.Acomparisonofdatareductionstrategies,islistedinorderofbandwidthreductioninFigure 1-4 22

PAGE 23

Figure1-4. UFoverallneuraldatareductionapproaches. 1.3BiphasicIFEncodingNeuronCircuit Theideaoftheencodingcircuitdesignoriginatedfromneurobiology.Biologicalsystemsrepresentsensoryinformation,usingthetimingofall-or-nothingactionpotentials[ 20 ].Inspiredbysuchresearchresultsfromneuroscience,alow-powerpulsesignalrepresentationcircuithasbeenproposedforneuralrecordingapplications[ 21 ]. TherststrategytoreducebandwidthwasdevelopedbyChen[ 22 ][ 23 ].Thisstrategyencodesthedatatoreducebandwidthwithanintegrate-and-re(IF)circuit,sothatintheorythesignalcanbeperfectlyreconstructedontheback-end,andusetraditionalspikesortingtechniques.ThisIFcircuitintegratesthecurrentconvertedfromtheampliedsignalvoltage.Theamplitudeinformationisencodedintoanasynchronousdigitalpulsetrain.TheIFcircuitcanefcientlyencodetheanaloginformationwith 23

PAGE 24

Figure1-5. Anexampleofaninputsignalandit'sbiphasicrepresentation.a)showstheinputsignalandb)showsthebiphasicrepresentation. respecttoconventionADC.Theprincipleistoencodethesignalamplitudeintothetimedifferencesbetweenpulses.Theresultingencodedpulsetrain,hasbetternoiseimmunitythananalogsignalsinwirelesstransmission. TheIFcircuitconvertsananalogvoltagewaveformtoapulsetrain,sothattheoriginalanalogsignalcanbereconstructedinexactlythesameway,throughadigitalalgorithmundercertainassumptions.Theoutputsignal(V)oftherststagebio-amplierwasconvertedtothecurrent(I).Theexistingpulseoutputcircuitintegratedthiscurrent,andshiftedtoguaranteepositiveonlyoutputs[ 23 ].ThepositivecurrentwasthenfedintotheIFcircuitstogenerateapulsetrain. However,ifonlypositivecurrentexists,theoverallringrate,thepowerconsumption,andtherequiredcommunicationbandwidthwouldbegreatlyincreased.Tosolvethisproblem,abiphasicIFcircuitdesignwasrstproposedbyChen[ 22 ].Ratherthan 24

PAGE 25

Figure1-6. Blockdiagramofthebiphasicencodingwithintegrate-and-re(IF)neuron. shiftingthesignaltobepositiveonly,abiphasicpulsetrainwasgeneratedusingtwocomparatorswithdifferentthresholds.ThisledustoemploytwoIFcircuitsthatencodedpositiveandnegativesignals,sothatfurtherbandwidthreductionwouldbeachieved.Thisencodingiscalledthebiphasicpulsetrainsignalrepresentation,andthecorrespondingcircuitiscalledthebiphasicIFcircuit.ThebiphasicIFcircuit,usespulsestorepresenttheinformationwhentheintegralofthewaveform(areaoverthetimeduration),surpassesapositiveornegativethreshold,asdepictedinFigure 1-5 Roughlyspeaking,thebandwidthcanbereducedbyapproximatelysixteentimes,overatraditionalADCsampledsystem(200KHz),ata25KHzsamplingrate,with8-bitofresolutionperchannel.AblockdiagramforthebiphasicencodingsystemisshowninFigure 1-6 .However,thedatareductionachievedbythebiphasicIFcircuit,isstillnotenoughtoscalehundredsofchannelsforasimultaneousneuralrecording.Furtherdatareductionisrequiredtofulllthebandwidth,andpowerconstraints,oftheimplantedwirelessneuralrecordingapplication.Simplyincreasingthethresholdvoltagetoreducethebandwidthcannotretainthesignalinformation.Therefore,thefeatureextractioncircuitdiscussedinthenextsection,isconguredtorefewerpulsesbetweenneuralspikes,whileretainingtheinformationinthespikeregion. 25

PAGE 26

Figure1-7. Blockdiagramofbiphasicencodingwithleakyintegrate-and-re(LIF)circuit. 1.4FeatureExtractionMethod ThefeatureextractionconceptprovidesfurtherdatareductionasproposedbyRogers[ 24 ][?].ThesignalreconstructionprocedureiscomputationalintensiveandanecessarysteppriortospikesortingforpulsesgeneratedfromabiphasicIFcircuit.TheabilityofanIFcircuitallowstheencodingcircuitdesigntoworkatasub-Nyquestrateduetoourinterestinqualifyingspikewaveformregions. Thespikefeatureextractionmethod,canreducethebandwidthevenfurther,byincludingaleakyterminthefront-endfeatureextractor,toformthebiphasicleakyintegrate-and-re(LIF)circuit,asshowninFigure 1-7 .Agreaterreductioninbandwidthresultsfromeliminatingpulsesinnoiseregions,aswellasprovidingsynchronizationforthepulsetrainoutput,atthetimeofthespike. Pulse-basedspikesortingwasexploredandshowntosupporttheback-endprocessingfortheLIFneuralrecordingscheme.Thespikesortingalgorithmusesasimplebutuntraditionaltemplatematchingmethod,becausethewaveformsarepulsetrains.Thefeatureextractorcanreducethebandwidthbyaboutanadditionalfactorofsixteen,lowerthanthebiphasicIFcircuit,makingitmorethantwo-ordersofmagnitudelowerthantraditionalADCsampleddata,at25KHzwith8-bitofresolution. 26

PAGE 27

Table1-1. DataCompressionTechnologies Bandwidth(Hz)/chSpikeSorting TraditionalADC(25KHz,8-bit)200KbpsYBiphasicPulseRepresentation12.5Kpulses/secYLIFFeatureExtraction<1Kpulses/secYSpikedetection<100pulses/secN Asummaryofdifferentdatareductionmechanismscomparingtheirbandwidthperformanceandspikesortingcapabilities,islistedinTable1-1.ThistableillustratesthebiphasicpulserepresentationwhichroughlyreducessixteentimesbandwidthoveratraditionalADC,whiletheLIFfeatureextractionmethod,furtherreducesbandwidthbyanadditionalsixteentimesoverthebiphasicpulserepresentation.Althoughspikedetectiontechnologycanachievethemostbandwidthreduction,spikesortingisnotavailabletoclassifydifferentneuroncellsringactionpotentials.Inourin-vivorecording,thebiphasicpulserepresentationtechniquewasused. 1.5ResearchGoal Althoughlotsoftechniquesforbandwidthreductionhavebeendevelopedandtestedwitharticialsignalsorrecordedneuralsignals,theywereonlybench-toptestings.Nobodyconductedanin-vivorecording.Therefore,alowbandwidthin-vivorecordingbecomesthemaingoalofthisdissertation.Inaddition,furtherimprovementforpackingthelowbandwidthsystemarchitectureisanothergoadaswell. Inordertoexploretheachievablebandwidthreductionforourneuralrecordingsystem,wewillproposenewstrategiesandcomparethemwithexistingstrategiesinordertodemonstratetheirbandwidthefciencyandspikesortingperformance.Thesystemwouldhaveananalogfront-endandadigitalsignalprocessingback-end.The 27

PAGE 28

taskoftheanalogfront-endistoamplifyandcompressthedatabandwidthbeforetransmission,whilethedigitalback-endreconstructsthecompressedinformation.Theprimaryfocusofthisresearchistheanalogfront-enddesign,thatprovidesanewandsimplerdatacompressorwhichreplacesatraditionalADC,sothatthedecodedinformationkeepsadequatequalitywithmuchlowerdatabandwidth.Asimpleexampleexplainsthepromisingperformanceofthisneuroncircuitdesign.Forinstance,supposethata6-bitADCissampledat25KHzforeachrecordingchannel.ThebandwidthconsumedbyeachADCrecordingchannelis150Kbits/sec/channel.Theequivalentresolutionachievesabout42dB.Therefore,theneuroncircuitdesignwearetargetingshouldachievelessthan15Kpulses/sec/channelwithatleast42dBsignalresolution. 1.6DissertationOverview Chapter 2 presentsthedesignofeachbuildingblockoftheanalogfront-endhardware.Theimprovedfront-endhardwaresystemdesignandintegrationincludingimprovedICdesign,canreducebandwidthandnoise.Asystematicmethodforsettingsystemparametersisrequired.Noisereductionmethodswillbediscussedanddemonstrated.Chapter 3 discussestheback-endsoftwaresignalprocessing.Traditionalreconstructionandspikesortingmethodsareusedtoprocesstherecordedpulsesinaneuralsimulatortest.Whenthemaximumrecordedpulseintervalissolarge,thattherecordedpulsescannotbeadequatelyreconstructed,apulse-basedspikesortingcanbeusedfortheback-endprocessing.Animprovedalgorithmtosystematicallygeneratetemplatesforthepulse-basedspikesorting,ispresentedandsimulatedinMATLAB,forneuralsimulatorsignals.Chapter 4 demonstratesthefullsystemintegrationandthesystemthatwillbeusedinin-vivorecordings.Threestrategiesforfurtherdatareduction,includingthetheoreticalneuronencodermodels,andthehardwarecircuitdesign,aredescribedinChapter 5 :therefractoryneuroncircuit,theleakyneuroncircuit,andtheadaptiveneuroncircuit.Theadaptiveneuroncircuitisproposedasasolutiontotheneuralrecordingwithwidedynamicrangeinitsamplitudeofsignals.Simulationswill 28

PAGE 29

bepresentedfortheseproposedandestablishedneuralencodermodels.Finally,theconclusionandcontributionsofthisresearchwillbereviewedinChapter 6 29

PAGE 30

CHAPTER2FRONT-ENDHARDWAREDESCRIPTIONANDINTEGRATIONTOTHEEXISTINGSYSTEM AsdiscussedinChapter 1 ,theUniversityofFlorida(UF)implantedneuralrecordingsystemhasbeenintegratedontoasinglePCBandfullytestedonthebenchtop.Thiscompletesystemoperation,includingtheback-endsignalprocessing,willbedemonstratedviain-vivoneuralrecordinginChapter 4 .Astepforwardtofullsystemintegration,andthetestingsetup,willalsobedescribedinthischapter. 2.1UFSystemOverview TheUniversityofFloridahasestablishedaneuralrecordingsystemtoencodethesignalssothatitcanbe,intheory,perfectlyreconstructedontheback-endwheretraditionalspikesortingcanbeapplied.ThissystemwasoriginallydevelopedbyChen[ 22 ],Li[ 25 ]andRogers[ 24 ]intheirPh.D.studiesattheUniversityofFlorida.Hence,onlyabriefoverviewwillbeprovidedhere,butreaderscanrefertotheirdissertationsforadditionalinformation. ThisUFsystemhasbeenusedinbenchtoprecording.Allelectronicsforamplicationandencodingwerebuiltuponabreadboard.Additionally,thebenchtoptestcouldonlybeconductedwhentheUFrecordingsystemwasplacedinsideabulkyfoilbox,whichwasexpectedtoprotectthehardwaresystemfromnoiseinterference.Thissetuppresentsthreemajorbarrierstothedevelopmentofapracticalimplantedin-vivoneuralrecordingsystem:(1)evenwiththefoilbox,noiselevelswerestilltoolargeforapracticalsystem,(2)theuseofthefoilboxmakesthesystemtoobulkytoimplementarealimplantedrecordingsystem,whichneedstoplaceanalogcircuitsclosetoimplantedelectrodes,and(3)thecurrentrecordingsetuplacksabenchmarktoevaluatethequalityofthein-vivoneuralrecording. Thehugeproblemintherealneuralrecordingexperimentsisnoisewhicharisesfromnumeroussourcesincluding: 1.60Hznoise:noisecoupledfromACpowerintheroom. 30

PAGE 31

Figure2-1. Theexistingneuralrecordingsystem:bio-amplier,V-Iconverter,andbiphasicintegrate-and-re(IF)recordingsystemwiththerefractorycomponentandtheleakycomponent. 2.Groundloopsbetweendigitalandanalogcircuits:becausethebiphasicIFcircuitconsistsofanaloganddigitalcircuits,withveryweakanaloginputsignalsandstrongdigitaloutputsignals.Therefore,thegroundloopissuecouldruintheneuralrecording. 3.Backgroundactivity:notonlydoelectrodespickup3or4neuronsofinterest,buttheyalsomayextractactionpotentialsredbydistantneuroncells,whichwouldbeconsideredbackgroundactivitywhichisnotofinterest. 4.DCdrift:neuralrecordingssometimescanexperienceaslowDCdriftassociatedwithtinyneuralspikes.SincetheamplitudeoftheDCdriftcanbelargerthantheactionpotentialsthemselves,informationofinterestmaynotbeencodedefciently.MorepulsesandmorebandwidthwillbewastedinencodingtheDCdrift. 5.Electricalnoise:thisnoisesourceisfundamentalforallintegratedcircuits,includingthermalnoise,ickernoise,andshotnoise. TheblockdiagramoftheexistingUFneuralrecordingsystemisshowninFigure 2-1 .Thesystemconsistsofthreemainblocks:thebio-amplierampliestheminuteinputsignalby40dB;thevoltage-to-current(V-I)converterconvertsthevoltagetothecurrentforthebiphasicIFcircuit;thebiphasicIFcircuitencodessignalsintobiphasic 31

PAGE 32

Figure2-2. Schematicofabio-amplier. pulserepresentation.TherefractorycomponentandtheleakycomponentareusedinthebiphasicIFcircuitfordatareduction,whicharediscussedinChapter 5 2.2Bio-amplier Therststepofrecordingtheextracellularneuralactionpotentialistoamplifysuchlow-levelamplitudesignals.Integratedcircuitshavebeendesignedforamplifyingtheweakneuralsignalsbeforeanyfurthersignalprocessingsteps[ 26 30 ].Thestructureofthebio-amplierwasoriginallyproposedbyHarrison[ 31 ],asshowninFigure 2-2 .AnACcouplingtechniqueisusedtorejecttheinherentDCoffset.ThemiddlebandgainAmis-C1/C2,thebandwidthisapproximatelygm/(AmCL),wheregmisthetransconductanceoftheoperationaltransconductanceamplier(OTA).Thelowcornerfrequencyisat!11/(RC2),thehighcornerfrequencyisat!2gmC2/(CLC1).Inordertoachievealowcut-offfrequency,asignicantresistanceisneededinthefeedbackloop.Thisresistancecanbeprovidedbyapseudo-resistorwithoutsacricing 32

PAGE 33

Figure2-3. Transistorleveldesignofoperationtransconductanceamplierusedinthebio-amplier. alargediearea.Thefourdiode-connectedPMOStransistorsMaMdinFigure 2-2 actaspseudo-resistors[ 32 ].Thepseudo-resistorelementfunctionsasapairofdiodesinseries,withoppositepolarities.Thecurrentthroughthepseudo-resistorincreasesexponentiallywiththevoltageforeithersignofvoltage,andthereisanextremelyhighresistanceregionforlowvoltagedrops.Figure 2-3 showsthetransistorleveldesignoftheOTAusedinthebio-amplier.TheOTAisatypicaltwostageCMOSamplier,withaP-typeinputdifferentialpairforlowerickernoise.TheinputdifferentialpairM1/M2isactivelyloadedbyWilsoncurrentmirrors.Acascadecurrentmirrorisusedtoconvertthedifferentialoutputintoasingle-endedoutput.Duetothehighoutputimpedanceofthecascadecurrentmirror,andtheWilsoncurrentmirror,thisamplierstageprovidesalargegain.Sincenoiseperformanceisimportantinthebio-amplierdesign,itis 33

PAGE 34

Figure2-4. SchematicofaV-Iconverter. minimizedbycarefullychoosingthewidthandlengthofthetransistors.However,thereisatrade-offbetweenstabilityandlownoise.Sincelownoiseiscriticalinthisapplication,increasingthesizeofM1andM2,anddecreasingthesizeofM7,M8,andthefourtransistorsofthetwoWilsoncurrentmirrors(M3,M4,M13,M14),canminimizethethermalnoise.Flickernoise,whichisimportantinlowfrequencyapplications,isminimizedbyincreasingthesizeoftheinputpair.BecausePMOSdevicesexhibitlessickernoisethanNMOS,PMOSisusedintheinputpair. AdditionalinformationoftheUFbio-amplierdesignisavailable.[ 23 ].ThemidbandgainAmisdesignedtobe40dB,thelowercornerfrequencyisdesignedtobe0.3Hz,andthehighercornerfrequencyisdesignedtobe5KHz.Fortestingpurposes,anexternallumpedcircuithighpasslter(300Hz)followstheoutputofthebio-amplifer,toremovemajorelectromagneticcontaminationof60Hz,anditsharmonicsfortheUFneuralrecordingsystem. 2.3V-IConverter BecausethebiphasicIFcircuitrequiresacurrentinput,aV-Iconverterwasdesignedtoconverttheampliedvoltagefromthebio-ampliertoacurrentinput,and 34

PAGE 35

Figure2-5. Transistorleveldesignofoperationtransconductanceamplier(OTA2). thentotheIFcircuit.Figure 2-4 showstheschematicoftheV-Iconverter.Thiscurrentgenerator,hasacapacitorC1attheinputtorejecttheDCcomponentofthesignal,andconvertsonlytheACvoltagetotheACcurrent.ProperDCbiasingmustbesettoallowOTA2tooperateinthesaturationregion.MaandMbaretwodiode-connectedPMOStransistorsactingaspseudo-resistorswithhugeresistance.ThishugeresistorcanbeviewedastheDCpasspathandtheACstoppath.Alternatively,thiscircuitcanbeseenasaDCclosed-loopandanACopen-loop.Theclosed-loopDCconguration,forcestheDCvoltagesatthenegativeinputnodeandoutputnodeoftheOTA2,tofollowtheDCvoltagexedatthepositiveinputnodeVmid,whiletheopen-loopedACcongurationfullyutilizesthehighopen-loopgainofthisOTA2.ThiscircuitrejectstheDCsignal,andampliestheACsignalofinterest,whichmakesitsuitableforthecurrentgenerator. 35

PAGE 36

Anotheradvantageofthiscircuit,isthatOTA2isconguredasavoltagefollowerforDCoperation,sothattheDCoffsetintroducedbytheinputdifferentialpair,willnotbeamplied.Atthesametime,theoutputresistanceshouldbehighenoughtoreducetheleakycurrent,whichhasbeenfullledbythecascodeoutputstagestructureasshowninFigure 2-5 Figure 2-5 showstheschematicofOTA2usedintheV-Iconverter.ThisOTAisatypicalsingle-stageCMOSamplierwithaP-typeinputdifferentialpair.Theinputdifferentialpairisloadedwithacascodecurrentmirror.Anothercascodecurrentmirrorisusedtoconvertthedifferentialoutputtoasingle-endedoutput.Thecascodecurrentmirrors,giveahighoutputimpedanceforOTA2.OTA2hasasingledominantpole,attributedtothehighoutputimpedance,andtheintegrationcapacitorCL,placedinthefrontendoftheIFcircuitinthenextcircuitstage. Sincelownoiseisthemainconcerninthisapplication,thesizeofthetransistorsmustbedesignedproperly.M7M10,M13andM14inFigure 2-5 arecommongatetransistors,sotheirthermalnoisecontributionisnegligible.TheinputpairM1andM2,NMOScurrentmirrorpairM3M6,andPMOScurrentmirrorpairM11andM12haveidenticalsizeandmatcheachotherindividually.Inordertominimizethenoise,themoststraightforwardapproachistoincrease(W/L)1,2andtodecrease(W/L)36,11,12.However,becauseM36,11,12areassociatedwithnondominantpoles,reducingthesizeofthesetransistorswillreducethetransconductanceofthesetransistors,andpushthenondominantpolesclosertothedominantpole,possiblydegradingthestability. Flickernoise,isanothermajornoisesourceinalow-frequencyCMOScircuits.ItcanbedecreasedbyincreasingtheareasoftheCMOStransistors,especiallyinuencedbytheinputstage.However,theareasarealsorelatedtothenondominantpolesasdiscussedabove,whichconstrainstheickernoiseperformance.Therefore,onlytheinputstageofM1andM2arefocusedtoincreasethearea(WL)forlowickernoise. 36

PAGE 37

Figure2-6. TheschematicofthenewV-Iconverterandintegrator. InthebenchtoptestoftheUFneuralrecordingsystem,thebiphasicIFcircuitfunctionsufferedseverelyfromtheoutputcurrentdriftfromtheV-Iconverter,whichbiasesthebiphasicIFcircuittoreineitherchannel,butnotinbothpositiveandnegativechannels.Wepostulatethattheleakageeffectfromthepseudo-resistorpairintheV-Iconverterfeedbackloopcausedthisproblem.Therefore,anewcircuitdesignwasproposedbysimplyreplacingthepseudo-resistorpairwithasinglepseudo-resistor,asshowninFigure 2-6 .ThisgureshowsthenewdesignoftheV-Iconverter,whereM1isasinglediode-connectedPMOStransistor.Asalreadydiscussed,thisdeviceactsasapseudo-resistorwitharesistancegreaterthan1011[ 23 ].TheoperatingpointoftheOTAisdenedbyVref.TheOTAconvertstheACvoltagesignalsofinterestintoacurrentlinearly,whileblockingDCoffsets,usingthecouplingcapacitorC1.ThecapacitorC2actsasanintegrator.Utilizingasingletransistortoactasthepseudo-resistor,hastheadvantageofleakagecurrentreductionoverthetwodiode-connectedtransistordesign[?],whichleadstoadeviationatVout,andanasymmetricbiphasicpulsetrainsattheIFcircuitoutput. 37

PAGE 38

Figure2-7. SchematicofthecomparatorcircuitusedinthebiphasicIFneuron. 2.4BiphasicIFCircuit Inthebiphasiccircuitdesign(Figure 1-6 ),asinglecapacitorisusedtointegratetheinputcurrentandapositiveandnegativethresholdaresetatoneinputforeachcomparatorseparately.Whenthevoltageacrossthecapacitorrisesabovethepositivethreshold(Vth+),apositivechannelpulseisgenerated;similarly,anegativechannelpulseiscreatedwhenthevoltagedropsbelowthenegativethreshold.Aftereitherspikeisgenerated,thevoltageonthecapacitorisresettoamidrangevoltagevaluebythedigitalcontrolcircuit,aXORgate.Whentheinputcurrentiszero-valued,nopulseisgenerated.Iftheamplitudeoftheinputcurrentishigh,theringratewillbecorrespondinglyhigh. ComparatorDesign Figure 2-7 showsthedesignofahigh-performancecomparatorusedinthebiphasicIFcircuit.Thecomparatorconsistsofthreestages:theinputpreamplier,thedecisioncircuitandtheoutputbuffer.TheinputsignalissensedbytheinputpairM1/M2,anddifferentialcurrentsarecopiedtothenextstageofthedecisioncircuit.Thepreamplierimprovesthecomparatorsensitivity,i.e.,decreasingtheminimuminputsignalwhich 38

PAGE 39

enablesthecomparatortomakeadecision.Theisolationoftheinputpairwiththepositivefeedbackinthedecisioncircuit,helpsdecreasethekickbacknoise.Thecross-connectedpairM7/M8,isusedtoincreasethegainofthecomparator.Sincethecomparatorshouldhavethecapabilitytodealwiththetime-varyinginputsignal,thediode-connectedpairM10/M11,isemployedtoprovidesomehysteresisusedtorejectthenoiseontheinputsignal.M14M18workastheoutputbuffertoconvertthenaloutputofthecomparatorintoalogic-levelsignal. Theinputpreamplierstageisadifferentialpairwithactiveloads.ThedimensionofM1andM2isdesignedbyconsideringthetransconductance,gm,andtheinputcapacitance.Thetransconductancedecidesthegainofthisstage,whiletheinputcapacitancecanbeconsideredasthecapacitancevariationoftheintegrator.Theimpedancelookingintothecross-coupledpairM7andM8is-1 gm7,8,ifthetransconductancesofM7andM8areequal.ThusthesmallsignalgainfromVin+orVin-tonodesXandYisAXY=gm1,2 gm10,11)]TJ /F7 7.97 Tf 6.59 0 Td[(gm7,8.Bycarefullyscalingthedevicesizeandbiascurrentratio,theAXYispreciselydened.ThedimensionofM13isdesignedforaDCshift,toprovidepropercommonmodelevelforthefollowingoutputbuffer.Aninverterwasaddedontheoutputofthebuffer,asanadditionalgainstage. 2.5USBRecordingBoard ThebiphasicpulsesarerecordedusingtheSimpleMonitorUSBxpressboard(USBBoard),developedbyDr.TobiasDelbruck,attheInstituteforNeuroinformatics(INI)inZurich,Switzerland.ForthebiphasicIFcircuit,itschargingcapacitorrestsignalisconnectedtotherequestsignalontheUSBboard.WheneverapulseisgeneratedbythebiphasicIFneuron,therequestpinontheUSBboardispulledlow,andgeneratesaninterruptsignal.Theboardstampsthepulsetiming,andrecordsthetimestampandaddressintoitsinternalRAM.Themicrocontrollerthenpullstheacknowledgesignallow,andtherequestsignalisthenpulledhigh,indicatingthatthehandshakeiscompleted.OncetheinternalRAMisfull,theaddressandtimestampsaresenttothehost,which 39

PAGE 40

Figure2-8. TheParallelRecordingConcept. isacomputerinthiscase.Thesetimestampsarecapturedonthemicrocontrollerusinganinternal1sclockanda16-bittimer,allowingforthemaximuminterpulseintervalof65mswithoutwrappingloss.Timestampsareunwrappedonthehosttoa32-bitresolution,providingamaximumtimestampofabout4300seconds. 2.6BenchTopTest Inordertobenchmarkourcurrentimplantedneuralrecordingsystem,wepresentain-vitrorecordingestablishingaparallelrecordingplatform,whichcomparesinstantrecordingsbetweentheUFrecordingsystem,andacommercialrecordinginstrument,asareliablereference.Fromtheparallelrecordingresults,groundtruthcanbeprovidedtocharacterizetheUFrecordingsystem. 2.6.1ParallelRecordingPlatformDesign TheparallelrecordingplatformwastestedwithaBionic's128-channelhardwareneuralsignalsimulator.Theuseoftheneuralsimulatorallowsforgroundtruth,meaning,thetimeofeachspike,andwhichneuronresthespike,tobeknown.Theneuralsimulatoroutputsarepeated10secpatternofspikesfromthreedifferentactionpotentials,withamplitudesof100150V,andatimewidthof1msec.Theinterspikeintervalis1secforrst9sec,andthenreducesto10msecforthelast1sec,forasetofburstring. 40

PAGE 41

Abench-toplaboratoryrecordingplatform,isusedtoevaluatetheperformanceofthisimplantedrecordingsystem.ToverifytheperformanceoftheUFrecordingsystem,weconstructedtheparallelrecordingplatform,asshowninFigure 2-8 whichillustratestheconceptoftheparallelrecordingplatform.Thesamesignalsourceissimultaneouslyfedintoboththeacommercialdataacquisitioninstrument,andtheUFrecordingsystem. TDTRecordingSystem Tucker-DavisTechnology(TDT,Alachua,Florida)isawell-knowncompany,providingintegratedhardwareandsoftwaresolutions,fordataacquisitionandanalysis.ThePZ-264channelpreamplier,andSystem3real-timesignalprocessingsystem,areinstalledinthecomputer,whichdigitizesampliedanalogsignalsat12.2kHz.Adigitalband-passlter(3005KHz),usedtoisolatesingleneuronactivity,wasimplementedinsoftware,andwasdesignedtohavethesamebandwidthastheUFanalogfront-endrecordingsystem. UFRecordingSystem TheUFrecordingsystemisdividedintothefront-endandtheback-endparts.Themajorstrategyofthisplatform,istoshiftthesignalprocessingloadtothedigitalback-end,whichdoesnothavestrictconstraintsonpowerconsumption,noise,andsize.Thetaskofthefront-endpartistoamplifythesensedsignals,andencodetheinformationintoapulse-basedrepresentation,totransmittotheback-endpart.Theback-enddecodesbyreconstructingtheoriginalsignalsfromthereceivedpulses.Thispulsecodingideafacilitatestheimplementationofwirelessimplantedrecordingelectrodes,whichrequiresasimpleandsmallsizefront-endanalogcircuitdesign,attachedtotheimplantedelectrodes,totransmitinformationinwirelesstotheback-end,forpostsignalprocessingwithlimitedbandwidthforlowpowerconsumption. 41

PAGE 42

Figure2-9. Abenchtoptestingsetupinparallelrecordingwithaneuralsimulatorasthesignalsource. Figure2-10. A20secsegmentoftheTDTsystemrecordingrepresentingatypicalqualityneuralsignalrecording.TheSNRisabout23.5dB. 2.6.2PlatformCongurationandSetup Figure 2-9 illustratestheparallelrecordingplatformcongurationfortheexperiment.Thesamesignalsourceissimultaneouslyfedintoeachrecordingpath.Theoutputsignalsrecordedfrombothsystemsarealignedandcompared.Forthebottompath,theUFrecordingsystemconsistsofthebio-amplier,andthebiphasicIFcircuit,withthebuilt-inV-Iconverter,refractorycomponent,leakycomponent,andtheUSBinterface 42

PAGE 43

Figure2-11. Aphotooftheparallelrecordingexperimentsetupforabench-toptestwithaneuralsimulator. board,connectedtoacomputer,withthereconstructionalgorithmfortheback-endprocess.Thebio-amplierandbiphasicIFcircuitarecustomCMOSintegratedchipsmountedonthesameprintedcircuitboard.Positiveandnegativethresholdsweresetto250mVindividuallyinthebiphasicIFcircuit.Therefractoryperiodwasadjustedto10s. Inthisneuralsimulatortest,thetoppath,theTDTsystemrecording(Figure 2-10 ),isutilizedasareferenceofcomparisontotheUFrecordingsystem.Theactionpotentialisabout150V,andthebackgroundnoiselevelisabout10V.TheUFrecordingsystemperformsSNR17.5dB,withthepulserateonaverage3Kpulses/sec.TheUFsystemprovidesatremendouscompressionfortheneuralsignalrecording,insteadoftheconventionalmethod,whichrequires200Kbits/sec(8-bitand25KHzsamplingrate). 43

PAGE 44

Figure2-12. A30secrecordingbeforethetestingimprovementisinvolved.(A)ThedirectrecordingisfromtheTDTrecordingsystem.ThelowSNRisduetothenoisecoupledfromtheUFrecordingsystem(B)Testpointismeasuredattheinputofthebio-amplier.ThelowSNRisduetothegroundloopandradiationnoiseintheUFsystem.(C)Testpointismeasuredattheoutputofthebio-amplierintheUFrecordingsystem.SNRisbetterthan(B)becausethebuilt-inlterfunctioninthebio-amplier. 2.6.3UFSystemNoise Thebench-toptestasshowninFigure 2-11 wasconductedtorecordsignalsfromtheneuralsimulator.ThemajordifcultyinrecordinghighSNRactionpotentialsinthebench-toptest,wasthattheparallelrecordingincreasesthenoiselevel,asshowninFigure 2-12 GroundLoopNoise AsdisplayedinFigure 2-9 ,thepulsesgeneratedfromtheUFbiphasicIFcircuit,arefedintotheUSBboard,andconvertedtodigitalsignalsintothePC.Becausethe 44

PAGE 45

Figure2-13. Aphotooftheimprovedparallelrecordingexperimentsetupforabenchtoptestwithaneuralsimulator. USBboard,whichdetectspulses,sharesthesamegroundlineastheanalogcircuitsonthePCBandtheTDTsystem,intheparallelrecordingplatform,groundloopnoisecaneasilyspreadoutthroughthewholesystem.Thisruinsthesmallsignalsattheinputportsoftheanalogcircuits,anddegradestheSNRoftheoutputsignalsofthebio-amplier.Onepossiblesolutiontoattenuatethegroundloopnoise,istoplaceopticalcouplerchipsbetweenpulseoutputsfromthePCBtotheUSBpulserecordingboard.Asaresult,thegroundoftheanalogsystemisisolatedfromtheUSBboard.Theopticalcouplersisolatenoisesourcesfromcouplingthroughthegroundloop,whichcoulddegradetheSNR.ThisnoisereductionsolutionwillbeseeninTable2-1. RadiationNoise Althoughtheopticalcouplersimprovedtherecording,thereconstructedsignalsstillshoweddistortion.Themainreasonsareduetotheloosecontacts,unsatisfactoryshieldsofwires,andthelongwireconnections.Unfortunately,thesedetrimental 45

PAGE 46

Figure2-14. Aphotoofacloseviewoftheimprovedparallelrecordingexperimentsetupforabenchtoptestwithaneuralsimulator. Figure2-15. The20secrecordingfromtheimprovedparallelrecordingsetupdirectlyrecordedfromtheTDTrecordingpath(top),andreconstructedsignalsrecordedfromtheUFrecordingsystem(bottom). designswereobservedinoursetup,whichcapturedtheradiationnoiseof60Hznoise,todistortthesmallanaloginputsignals.Therefore,shortcablesandtconnectionsarenecessaryimprovements,asshowninFigure 2-13 ,andacloseviewinFigure 46

PAGE 47

Table2-1. RecordingresultsofSNRfordifferenttestingsetup SNR(dB)backgroundnoiselevel(V) Recordingwithoutnoisereduction5.480Groundloopnoisereduction9.550Radiationnoisereduction17.520Fullyintegratedsystem23.510 2-14 .Consequently,apromisingrecordingresultshowsverylownoiseleveldowntotheamplitudeof20V,closetotheTDTrecording,asshowninFigure 2-15 .Thereconstructionprocesswillbediscussedindetailinthenextchapter.Furthernoisereductionisexpectedifthecircuitsarefabricatedintoasinglechip,andthewholeUFrecordingsystemisintegratedontoasinglePCB.Table2-1liststheSNRandthebackgroundnoiselevelsoftherecordingswithdifferentnoisereductionsolutions. 2.7AnalogFront-endSystemIntegration Figure2-16. Thesystemcongurationofanadaptiveleakyrefractoryintegrate-and-re(ALRIF)neuroncircuit. 47

PAGE 48

Figure2-17. ThePCBdesignoftheanalogfront-endrecordingsystem:(a)thepowersupplyboardprovidesdifferentDCvoltagesforthewholesystem,(b)themainboardofencodingcircuitsincluding(1)theamplierchipsand(2)theV-Iconverterandtherefractoryleakyintegrate-and-reneuroncircuitchip,(c)theadaptivecomponentcircuitboardand(d)theUSBinterfaceboard. Thenalproposedanalogfront-endsystem,includingmoreneuroncircuitdesignforfurthernoisereductionisrepresentedinFigure 2-16 ,isdesignedusingtheAMI0.6mprocesswithasupplyvoltageof2.5V,whichwillbediscussedcompletelyinChapter 5 .Theinputsignalextractedfromtheelectrode,isampliedinthebio-amplierstage,andconvertedintocurrentviatheV-Iconverter,forthebiphasicIFencodingcircuitinthenextstage.ThecurrentisthenintegratedontothecapacitorinthebiphasicIFcircuit.Pulsesaregeneratedoncetheintegratedvoltagereachesthethresholdvoltage.Therearethreeadditionalcomponentsforfurtherdatareduction:theleakycomponent,therefractorycomponent,andtheadaptivecomponent. Thewholeanalogfront-endcircuitsystemisimplementedintoaPCB,asshowninFigure 2-17 .TestingthewholesystemonaPCB,providestheadvantagesofsystemstability,noiserobustness,andeasysetup,whichareallessentialtoarealrecording 48

PAGE 49

test.Inordertoincreasethetestingefciency,theanalogfront-endsystemisdesignedasamodule-basedstructure.Thesystemhasbeendividedintodifferentboards,eachofwhich,hasaspecicfunction.Suchdesignbenetsthedebugandtestprocesses.Thesourceofsystemerrorsorfailurescanbeisolatedmoreeasily.Wecanrapidlyreplacenewboards,orquicklychangedifferentdesignsofboards,fordifferentrecordingpurposes. Thesystemconsistsofonemainboardassociatedwiththreesideboards,asdisplayedinFigure 2-18 (a)Powerboard:thisboardincludesthevoltageregulatortogenerate5VDCvoltagefroma9Vbattery,andprovidesseparateDCvoltagestopowertherestoftheboardsinthesystem. (b)Mainboard:thisboardistheheartofthesystemwherethe(1)bio-amplierchip,(2)V-Iconverterandrefractoryleakybiphasicintegrate-and-rechiparelocated.Attheinputofthisboard,theDB-25portaccepts16differentchannelsofsignalsfromthesensingelectrodes.Thenonechannelofthesignalscouldbeselectedtobeprocessedthroughtheswitch.Inordertoreducetheimpactofthe60Hzanditsharmonics,arst-orderhighpasslterof300Hzissetupattheoutputofthebio-amplier.Thereare4differentvariableresistorstosetthedifferentparametersfortheUFfront-endanalogsystem. (c)Adaptivecomponentcircuitboard:theadaptivecomponentchipisoptionalfortherecordingsystem.Analoganddigitalbuffersarerequiredforcommunicatingwiththemainboard.ThiscomponentwillbediscussedfullyinChapter 5 (d)USBinterfaceboard:thisboardbufferstheencodedpulsesfromthemainboardandsimultaneouslyacknowledgestheUSBpulserecordingboardtorecordthepulsevalue. 49

PAGE 50

Figure2-18. AphotographoftheUFanalogfront-endrecordingsystem. 50

PAGE 51

CHAPTER3BACK-ENDSOFTWARETESTANDCHARACTERIZATION 3.1Reconstruction InordertocomparetherecordedpulsesfromtheUFrecordingsystemagainstthecontinuoussignalsrecordedfromtheTDTrecordingsystem,reconstructionisnecessaryforcomparisonpriortospikesorting.Abriefoverviewisshown,asfollows,andfurtherdetailisprovidedin[ 22 ]and[ 33 ].IntheidealbiphasicIFneuroncircuitmodel,whentheintegratedcurrentx(t)reachesthedenedthresholdvoltageatthistimeti,apulseisgeneratedandthevoltageacrossthecapacitorisresettozero.Theringtimesmustsatisfy: Zti+1tix(t)dt=,8i(3) x(t)needstobebandlimitedto[-s,s],andti,i2Zisatimingsequencewithmaximumadjacentpulseinterval(ti+1-ti)
PAGE 52

Substitutingeq. 3 intoeq. 3 ,weobtain=Zti+1tix(t)dt (3)=Zti+1tiXj!jh(t)]TJ /F14 11.955 Tf 11.95 0 Td[(sj)dt (3)=Xj!jcij (3) wherecijareconstantsthatcanbenumericallycomputedwith: cij=Zti+1tih(t)]TJ /F14 11.955 Tf 11.96 0 Td[(sj)dt(3) TheresultingsetoflinearequationsisgivenbyC W = inmatrixform.ThustheweightvectorW canobtainedas W =C + (3) whereC +isaMoore-PenrosepseudoinverseofC matrix,whichisusuallyanill-conditionedmatrix,andnecessitatestheuseofsomesortofregularizationprocess,suchasthepseudo-inversetechnique(pinvfunctioninMATLAB),tocalculateC Theweightvectorcanbesubstitutedineq. 3 ,tonumericallycalculatethereconstructedsignalx(t)towithinacertainsetprecision.Thenumericalmethodsusedtosolvetheaboveequations,areunsuitableforthereal-timeimplementation,sincetheyarecomputationallyexpensivebothintermsofmemory,andinthenumberofcomputations.Therefore,amorepracticalmethodhasbeendeveloped,andusedintheback-endprocessingforanin-vivorecordingsystem.[ 36 ] Figure 3-1 ,hasthebiphasicpulsetrainsrecordedfromtheUFrecordingsystem,anddisplaysthereconstructionresultinanin-vitrorecording,whichcomparesthereconstructedwaveformsandgroundtruthsignalsfromtheTDTsystem,atazoom-inscopeof25msec.Threedistinctneuronscanbedistinguishedattheinterspikeintervalsof10msec.Reconstructedspikescorrespondtodensepulsering.TheUFrecordingsystemintroducedsomedistortionsintherelativeratioofthedepolarization, 52

PAGE 53

Figure3-1. Alignmentofthreesignals:(top)biphasicpulsetrainsrecordedthroughtheUSBboard,(middle)theoriginalsignalsrecordedthroughtheTDTsystem,and(bottom)reconstructedsignalsattheback-endoftheUFrecordingsystemtestedwiththesetting 2-13 repolarization,andhyperpolarizationphasesoftheactionpotentials.Thisdistortioncanbecontrolledbytheparametersusedtoencodethewaveform,andrepresentsatradeoffbetweenthebandwidth,andtheaccuracyofthereconstruction.Inaddition,thedistortionisalsoattributedtothenoise.Inthisexperiment,alowerbandwidthwaschosenanditshowedthatthedistortionsareuniqueandconsistentwithineachneurontype.Thisobservationopensthepossibilityofspikesorting. 3.2ConventionalSpikeSorting Introduction Multipleneuronsmayberecordedfromthesamechannelofin-vivoextracellularneuralrecordings.Featuresofspikeshape,canbeusedtodistinguishbetweenindividualneuronsignalsthroughaprocesscalledspikesorting.However,because 53

PAGE 54

Figure3-2. ThespikesortingresultforthereconstructedsignalsbasedontemplatematchingandaPCAclustercuttingalgorithm. popularspikesortingmethodsareusuallytoocomputationallyintensiveforimplanteddevices,spikesortingistypicallypushedtotheback-endwherepowerislessstringent.AdetailreviewofthemajorspikesortingmethodscanbefoundinLewicki[ 37 ].Despitethefactthatmanyadvancedspikesorters,suchasthereal-timespikesorting,andtheautomaticspikesorting,claimtoachievecompatiblespikesortingperformance,withlesstimeandhumaninvolvement[ 38 ]and[ 39 ],nocommunitywideagreementhasbeenachievedastowhichspikesorteristhebest. Inordertocomparetherecordingperformance,spikesortingisimplementedonboththereconstructedsignalsfromthebiphasicIFrepresentation,aswellasdirectrecordingsfromacommercialinstrument.Timestampsofeachclassiedneuronfrombothdatasourcesarerecordedtoallowcomparisons.Spikesorting,isaclassicalmethodusedbyneuroscientistswithmanyproposedalgorithmsintheliterature[ 37 ]. 54

PAGE 55

Spike2TM(CED,UK),apopularcommercialprogram,whichcanspikesortofine,isusedtoanalyzethereconstructedsignalsandthedirectrecordedsignals.Spike2TMrstperformspreliminaryspikedetectionbycapturingwindowsaroundspikecandidatesthatcrossauser-denedthresholdvoltage.Thespikesortingisperformedwithacombinationoftemplatematchingandaprincipalcomponentanalysis(PCA)basedclustering[ 37 ].Thisprocess,requiresuserstoselectmanyparameterspriortothetemplatesetup,suchastheminimumnumberoftemplatesneededforclassifying,andallowablevariationwithinthetemplates.Spike2TMprovidesaninteractivevisualdisplayforuserstoassistinsettingthespikesortingparameters. NeuralSimulatorTest WeappliedtheconventionalspikesortingmethodtothereconstructedsignalsshowninFigure 3-1 (bottom).Figure 3-2 showsthreedistinctclustersconstructedfromthePCA,forthecandidatewaveforms,eachofwhichindicatesaclassiedneuronfromtherecordinginFigure 3-1 .EachdotinthedifferentgroupclustersrepresentsasingleactionpotentialclassiedbySpike2TM.Allthreeneuronscanbedistinguished(Table3-1andTable3-2)withhighmatchingrate,lowmissedspikeratio,misdetectionratio,andmisclassicationratio.However,thereisstillsomedistortionobservedforalmostallpeakportionsoftheactionpotentials,whichaffectsthecrosscorrelationcoefcients(CCC)oftheactionpotentials,betweenthereconstructedsignals,andthedirectTDTrecording,asshowninTable3-2.However,theimperfectionofthebiphasicIFcircuitparameters,therecordinghardwaresetting,andtheerrorintroducedfromthereconstructionalgorithms,alsocausesomedistortion.ThereasonforthelowCCC,buthighmatchingrateforNeuron3,isduetotheimpactofwindowsizeselectionsforactionpotentialwaveforms. Whenweappliedtheconventionalspikesortingmethodtothein-vitroparallelrecordingintheneuralsimulatortest,thespikesortingresultsofthereconstructedsignalsandthedirectTDTrecording,asshowninFigure 3-3 ,showedasimilaritytothe 55

PAGE 56

Table3-1. DetectionresultsofrecordingfromFigure 2-13 setting MissedSpikeRatioMisdetectionMisclassication 0.45%0.45%0.45% Table3-2. PatternrecognitionresultsfromFigure 2-13 setting Neuron1Neuron2Neuron3 MatchingRate98.67%100%100%CrossCorrelationCoefcient0.840.830.69 overlappedactionpotentialsofdifferentneuronsbetweenthedirectTDTrecordingandreconstructedsignals. 3.3Pulse-basedSpikeSorting Spikesortingisoftenusedtoclassifydifferentneuronsinneuralrecordings.Conventionalwaveshapeanalysisinthespikesortingprocedure,providesameanstodetectspikes.Owningtohighpowerconsumptionofcomputingthemultiplechannelneuralrecording,spikesortinghasbeenshiftedtotheback-endprocesswherepowerisnotstrictlylimited.Thetransmissionbandwidthtotheback-end,hence,playsanimportantrole.Thepulse-basedrepresentationstrategiesdevelopedatUF,havesuccessfullyreducedtherecordingbandwidthfromthatoftraditionalADC-basedsystems.However,whenmoremultiplechannelneuralrecordingsneedtoberecorded,anexistingrecordingsystemdevelopedbyChen[ 22 ],cannotaffordsuchtremendousincreasesofbandwidthtransmission.Therefore,afurtherdatareductionscheme,aleakyintegrate-and-re(LIF)encodermodel,isemployed,tohavefurtherbandwidthreduction.IftheparametersoftheLIFencodingcircuitaresetupproperly,thelowbandwidthrecordingcouldbereached.However,whenthebandwidthistoolow,the 56

PAGE 57

Figure3-3. AcomparisonofspikesortingresultsfromtheneuralsimulatorbetweenTDTdirectrecordingandreconstructedsignals. conventionalreconstructionmethodfailstoreconstructtheactionpotentialwaveforms,andleadstopoorspikesortingresults.Thissituation,providesthemotivationtodevelopthepulse-basedspikesortingmethod,withoutareconstructionstep.Althoughtheideaofpulse-basedspikesortinghasbeenveriedinsimulationwiththeLIFencodermodel[ 24 ],whenmovingtheseideastoin-vivorecordings,thereisnodatatooptimallyadjusttheleakyandthethresholdvoltageparametersoftheLIFencodingcircuit,foroptimalbandwidthreduction,andperformance.Moreover,thecurrentpulse-basedspikesortingalgorithm,lacksasystematictrainingmethodtogeneratepulse-basedtemplatesforactionpotentials.Instead,expertsneedtogetinvolvedtomanuallyselectasingleactionpotentialforeachneuron.Thesespikesaretransferredintopulsesignaturestobethetemplatesforthepulse-basedspikesortingmethod. FeatureExtraction ThefeatureoftheLIFencodermodel,istoprovideextremelylowbandwidthtomerelyrepresentfeaturesofactionpotentials,whenthepulsebandwidthhasbeenless 57

PAGE 58

thanthelimitoftheNyquistrate.Asaresult,theconventionalwaveshapeanalysisinthespikesortingprocedure,maynotbeanappropriatemethodtoclassifyneurons.Thissituationleadstothepulse-basedspikesortingmethodtospikesortsignalsinthepulsedomain.TheLIFencodingcircuithasaleakycomponent,whichsetsacutofffrequencywiththecapacitor.Theleakvalue,alongwiththeproperthresholdsettings,allowsforveryfewpulseswhichrepresentthenoiseandthemajorityofpulsesthatcontaininformationabouttheneuralspikes.Thetimingoftwoconsecutivepulsesmustsatisfythefollowingequation: 1 CZti+1ti+x()e)]TJ /F12 5.978 Tf 5.75 0 Td[(ti+1 RCd=i(3) wherei2f)]TJ /F4 11.955 Tf 27.59 0 Td[(,gandCisrelatedtotheintegrationcapacitor,andtheRisrelatedtotheleakvalue. SignatureConvolvedwithGaussianFunction Thepulsesrepresentingeachspikeserveasthespikesignature,whereapulse-basedspikesortingalgorithmisusedtoclassifythespikes.Inthisalgorithm,wesetuptheparametersthateachspikesignatureshouldhavefourzerocrossingswithin2ms.Thespikesignaturecandidatesareselectedbasedonthesecriteria. Withthesameprincipleastheconventionalspikesortingalgorithm,thepulse-basedspikesortingalgorithmgeneratestemplatesforeachneuronfromthetrainingphasedataset,andusesthesetemplatestoclassifyspikesinthetestingphasedataset.Therefore,thetechniqueofmatchingtwopulsetrainsisappliedtoclassifytheactionpotentialswithtemplates.Manytechniquesofcomparingtwopulsetrainshavebeendiscussedasfollows.Whiledistortionmetricsforspiketrainshavebeenstudiedinareassuchasneuroscienceandgenetics,manymethodsarecomputationallycomplexandfarfromtherealtimeapplication,suchastheeditdistance[ 40 ].Anotheridea,istolow-passlterthepulsetrainswithafunction,suchasanexponentialfunction,somoretraditionalsignalprocessingcanbeapplied[ 41 ]. 58

PAGE 59

TheoutputpulsetrainsfromtheLIFencodingcircuitaresampledattheback-end,atiminguncertaintyerror,alsocalledjitterorquantizationerror,introducedbyback-endsamplingcannotbeavoided.Thepulse-basedspikesortingmethodconvolveseachpulseinpulsestrainswithaGaussianfunctionfortoleratingthiserrortoformtemplates,wherethedeterminesifthedetectorismoreofacoincidencedetector(muchsmallerthantheinterpulseinterval)orapulsecountdetector(large).AGaussianfunctionwaschosen,asitismoreconcentratedaroundthepeak,allowingthetobettercontrolthedetectortype.OncethepulsesignatureisconvolvedwiththeGaussianfunction,itisthencomparedtoeachofotherGaussian-convolvedspikesignatures.Thetemplatewillbeformedfromtheaveragewaveformsofthosespikesignatures,withthemean-square-error(MSE)lessthanadenedthreshold.ThespikesignatureswithMSEhigherthanthethreshold,areconsiderednoise[ 24 ]. 3.3.1SystematicTrainingforTemplateGeneration Sincedirectspikesortinginpulsetrainsiscomputationallyexpensive,eachpulsesignatureisconvolvedwithaGaussianfunction,toallowtraditionaltemplatematchingsignalprocessingtechniques,tobeapplied.Differentselectionsrepresentdifferentspikedetectors.Spikesortingerror(SSE)ishighlyinuencedwithdifferentselections.Ifistoosmall,templatematchingwouldbeextremelydifcult,becausenoiserepresentedpulsescausejitterstotherstpulsetiming,representinganactionpotential.Thissituationwouldinuencetimingsofalltheotherpulseswithinthesamespike,whichleadstohighSSE.Ontheotherhand,ahighSSEalsotakesplacewhenischosentoolarge.Foralargeselection,apulsesignaturewouldbesmearedmimickingotherpulsesignaturesgeneratedbydifferentneuralspikes.Thisresultdecreasesthedifferentiationamongstdifferentneuralactionpotentials,aswellastherejectiontobackgroundnoise. Alargedrivesthespikesortertoaspikecountdetector,becausepulseratesprovidetheinformation.Hence,morepulsesarepreferredtoencodeaspike.Onthe 59

PAGE 60

Figure3-4. Asegmentofrecordingdatafromtheneuralsimulator. Figure3-5. PulsesgeneratedfromasimpleIFencodermodelandpotentialpulse signaturesaremarkedinred. otherhand,asmall turnsthespikesortertoacoincidencedetector,whichmeans thateachpulsetimingismoreimportantandcarriesmoreinformation.Therefore,less pulsesarepreferredtoencodeaspike.Ifthebestdetectorwassomewherebetweena 60

PAGE 61

Figure 3-6.Truepositiverate(TPR)v.s.different convolvingwithdifferentneuronsfor thecaseofasimpleIFencodingfeatureextractor. Figure 3-7.AveragedtemplatesforeachneuronforthecaseofasimpleIFencoding featureextractor. coincidencedetectorandaspike-countdetector,thecurveinaSSEv.s. gurewould haveaU-shape,withasweetspotforthe ,wherethepulse-basedspikesorterwill applyinordertoachievethebestspikesortingperformance. 61

PAGE 62

Figure 3-8.Potentialpulsesignaturesaremarkedinredinthetestingphaseforthecase ofasimpleIFencodingfeatureextractor. 3.3.2NeuralSimulatorTest Fortestingthepulse-basedspikesortingalgorithminMATLAB,arecordedneural simulatorsignalisshowninFigure 3-4.Itisexamined,consistingofthreesetsof one-secondlongspikebursting,andonespikeineverysecond.InFigure 3-4,therst spikeburstingsetwillbeexaminedasthetrainingphasedataset,whilethesecond spikeburstingsetwillbeexaminedasthetestingphasedataset. Large Selection WerstencodespikeswiththeIFencodermodelfromthetrainingphasedata set.Thepulse-basedsorterscansthroughouttherecordingtoselectpotentialspike signaturesmarkedinred,asshowninFigure 3-5.Next,thespikesorterappliesdifferent andconvolvestheseGaussianfunctionswithallpotentialspikesignatures.Different neuronsareclusteredandeachneuronshowsaU-shapecurveintheSSEv.s. gure, asillustratedinFigure 3-6.Truepositiverate(TPR)isusedastheSSEintheneural simulatortest.IntheFigure 3-6,becausedifferentneuronshavedifferentoptimalvalues 62

PAGE 63

of,averagingthemischosenforthebesttogeneratetemplatesforeachneuron,asshowninFigure 3-7 .Inthelaststep,thesetemplatesareusedtoclassifypulsesignaturesmarkedinredinFigure 3-8 .Inthelargetest,100%spikesortingmatchingrateshasbeenpresentedwiththebestselectionof0.7msecinthissimulation. SmallSelection Withthesametestingprocedure,butwithasmallercaseofasmallselection,theLIFencodermodelisissuedinsteadoftheIFencodermodel,anditisutilizedtoencodespikeswithlesspulsesfromthesamerecording.Thepulse-basedsorterscansthroughoutthetrainingdatasettomarkpotentialpulsesignaturesinred,asshowninFigure 3-9 .Next,thespikesorterappliesdifferentselectionsandconvolvestheGaussianfunctionstoallpulsesignatures.Differentneuronsareclusteredforeachselection.TheSSEv.s.guredemonstratesaU-shapecurve,asillustratedinFigure 3-10 .Truepositiverate(TPR)isusedastheSSEfortheneuralsimulatortest.Again,averagingthebestfromeachneuronistakenastheoptimaltoformtemplatesforeachneuron,asshowninFigure 3-11 .Inthenalstep,thesetemplatesclassifythepotentialpulsesignaturesmarkedinredinFigure 3-12 .Inthesmalltest,99%spikesortingmatchingrateshasbeenpresentedwiththebestselectionof0.16msecinthissimulation. Discussionforpulse-basedspikesorting Performanceofthepulse-basedspikesortinghasbeensimulatedinMATLABwiththeIFandLIFencodermodels.Moreover,asystematictrainingmethodhasbeencreatedtoselectthebestforthepulse-basedspikesortingforeachrecording.TwodifferentencodingcircuitshavetheirownbesttoapproachtheirbestSSE.TheIFencodermodelgeneratesmorepulsesthantheLIFencodermodeldoes,andsomepulsesareencodedforthenoisepartsintherecording.Therefore,largeselectionsarepreferredandthespikesorterismoretowardaspikecountdetector.Onthecontrary,TheLIFencodermodelgeneratesfewerpulses,andveryminimumpulses 63

PAGE 64

Figure3-9. PulsesgeneratedfromaLIFencodermodelandpotentialpulsesignatures aremarkedinred. Figure3-10. Truepositiverate(TPR)v.s.different convolvingwithdifferentneuronsfor thecaseofaLIFencodingfeatureextractor. representthenoisepartsintherecording.Asaresult,small selectionsarechosen, andthespikesorterismoretowardtoacoincidentdetector. 64

PAGE 65

Figure 3-11.AveragedtemplatesforeachneuronforthecaseofaLIFencodingfeature extractor. Figure 3-12.Potentialpulsesignaturesaremarkedinredinthetestingphaseforeach neuronforthecaseofaLIFencodingfeatureextractor. Nevertheless,aproblemmayberaisedfromusingasinglevalueof .Spikes withdifferentamplitudeshavedifferentinterpulseintervals.Theproblemisthatthe valueof ,whichcorrespondstoacoincidencedetectororapulsecountdetector, 65

PAGE 66

dependsontheinterpulseintervalofthepulsetrain.Thus,itispossiblethatwithinapulsesignature,whenonlysingleaveragingisapplied,onepartofthepulsesignaturewouldbeclassiedforonetypeofspike,whileotherpartsofthepulsesignature,wouldmovemoretowardtheothertypeofspike.TheLIFencodermodelmaynotresolvethisproblem,because,theleakvalueonlysynchronizesthebeginningofthepulsesignature,andbytheendofthepulsesignature,theaccumulatednoisewouldcausethelaterpulsetimingstodeviate.Twopossiblesolutionsarediscussedbelow. (1)AnadaptiveLIFencodermodelisutilizedtocreateamoreuniformpulserate,whichisinspiredfromthebiologicalneuron'sadaptivethresholdmechanism,thatkeepstheringratefromsaturatingandinformationfrombeinglost[ 42 ]. (2)Variablevaluesaccordingtothespiketemplateinterpulseintervalleadstoamoreconstantdetectoracrossspikes. Inthisresearch,theALRIFneuronencodermodelisproposedinChapter 5 fortherstsolution.Thefocusofthissectionisthefeatureextractionnotthespikesorting.Thus,thespikesortingprocedurewaskeptsimpletopurelyshowthatthefeatureextractionhaspotentialbenets. 3.4FullSystemConguration Figure3-13. ThecongurationoftheUFneuralrecordingsystem. Thefullsystemcongurationincludingthefront-endandback-endisillustratedinFigure 3-13 .Boththeconventionalspikesortingmethod,andthepulse-based 66

PAGE 67

spikesortingmethod,sharetheexactsamefront-endhardwarebutwithdifferentparametersettings.Whentherecordedpulserateissufcientforreconstruction,theperformanceoftheneuralrecordingsystemcanbeevaluatedbasedonSERv.s.bandwidthconsiderations.Ontheotherhand,whentherecordedpulserateisnotsufcientforadequatereconstruction,theback-endcanbeswitchedtothepulse-basedspikesortingmethod,toevaluatetherecordingsystemintermsofSSEv.s.bandwidth. 67

PAGE 68

CHAPTER4FULLNEURALRECORDINGSYSTEMOPERATION 4.1AcuteIn-vivoExperiment Chapter 2 and 3 discussedthedesignanddevelopmentforthecompleteUFrecordingsystem.Thenextfocusisthein-vivorecordingontherat.Thegoalofthisexperimentistoextractclassiableactionpotentialsfromalivesubject. Figure4-1. Thein-vivorecordingpreparation:(a)theUFanalogfront-endrecordingboardconnectstheDB-25adapterfromtheelectrodebufferandtheUSBporttothePCand(b)theratrsthasanelectrodearrayimplantedintoitsbrainfortheneuralsignaldetectionandabufferhookedonthetopholdssignalstotheelectronicsintheback. Figure 4-1 showsthepreparationofthein-vivorecording.TheratundertesthasanelectrodearrayimplantedviaabrainsurgeryconductedbyatrainedexpertusingapprovedIACUCprotocol.Theactionpotentialsinthehippocampusismonitored.Then,abuffer,alsocalledaheadstage,isrequiredtoconnectfromtheimplantedelectrodearraytotheanalogfront-endrecordingplatform,includingboththeTDTandtheUFrecordingsystems,fortheimpedancematchingpurposes.Asaresult,thesignalcanbefaithfullytransferredtotheelectronicsforfurtheramplicationandencoding.TheUFanalogfront-endboardacceptstheDB-25connectorfromtheheadstageattached 68

PAGE 69

Figure4-2. Thetestingconceptforthein-vivorecordingintheacuteexperiment. Figure4-3. Thewholeacutein-vivoexperimentsetup:theratundertest(top)isconnectedtoeithertheUFrecordingsystem(center)ortheTDTrecordingsystem(bottom).ThePC(left)hastheMATLABprogramtocontrolbothsystemstorecordanddisplayrecordings. totherat'shead,andaUSBcableconnectsfromthesystemoutputforextractingtheencodedpulseeventstothePC. 69

PAGE 70

Figure4-4. Aphotoshowstherathasbeenimplantedwithanelectrodearraydowntothehippocampusareafortheneuralsignalrecordingintheacuteexperiment. Figure4-5. Aphotoshowstheratisanesthetized.Itwouldbekeptasleepbythe1.5%isouranceandplacedonahotplatethroughouttheacuteexperiment. Theacutein-vivoexperimentalprincipleisillustratedinFigure 4-2 .Theacuteexperimentmeansthattheanimalwouldbeexaminedcontinuouslyunderthetestforthewholeday,untiltheexpecteddatahasbeenextracted.Thepurposeoftheacute 70

PAGE 71

experimentistocollectthedatafasterandmoreeasily.TheTDTelectrodeisrstimplantedintherat'sbrain.TheelectrodeisthenattachedtotheTDTheadstagetoretainclearsignals.Thescrewimplantedintheskulloftheratprovidesaneuralreferencevoltagetothebuffer(headstage).ThisheadstageisthenconnectedtoeithertheTDTsystemortheUFrecordingsystem.ThegoalistocomparetheactionpotentialsrevealedintherecordingsfromtheTDTandtheUFrecordingsystemsviathespikesortingprocess.Figure 4-3 demonstratesthewholeacutein-vivoexperimentsetup.Ifbothrecordingscanbeclassiedasthesamedistinctactionpotentials,theUFrecordingwouldbeconsistentwiththeTDTrecording. Figure4-6. TheMATLABinterfaceprogrammonitorsthetotal16channelsfromtheimplantedelectrodearrayfortheneuronringactivityineachchannel. Therststepfortheacutein-vivoexperimentistheanimalpreparation.Animplantationofanelectrode16-channelarrayhasbeenconductedthroughsurgeryasshowninFigure 4-4 .Theratwouldbeanesthetizedby1.5%isouranceduringthewholeexperimentandplacedonthetopofahotplatetomaintainhisbodytemperature,aspresentedinFigure 4-5 71

PAGE 72

Figure4-7. TheUFanalogfront-endsystemboardextractstheneuralsignalsfromtheheadstagethroughaDB-25adapter(leftside)andpassestheoutputpulsestotheback-endthroughaUSBcableattached(rightside). Beforetherealin-vivorecording,weneedtosearchforneuralactivity.TheMATLABprograminthePCprovidesinstantobservationofeachoftheimplantedelectrodechannels.Thisimplantedelectrodearrayhasauniquefeature.Thisfeatureallowstheusercandriveeachelectrodedowntodifferentdepthsinthebraintodetectneuronringsignals,whichprovidesmoreexibilityintheneuralrecordingexperiment.Whiledrivingtheelectrodedown,variousneuralactivitiesappearinallofthechannels.IntheFigure 4-6 ,onlychannels3,14,15,and16haveneuralactivities.Eitherchannelcouldbechosentoextractactionpotentialsinthein-vivorecording.TheUFanalogfront-endboardisconnectedfromtheheadstageontherat'sheadviatheDB-25adaptor.NextaUSBcableisattachedtothisrecordingboardbacktothePCinordertocollecttheencodedpulserepresentationasshowninFigure 4-7 .InthemainboardoftheUFanalogfront-endsystem,thereare16switches,eachofwhichcontrolsthepassingoftherecordinginoneindividualchannel.Onlyonechannelshouldbeturnedonwhiletherestofthechannelsshouldbeturnedoff.Thisswitchdesignmakesthisin-vivorecordingexperimentmuchmoreexible. 72

PAGE 73

4.2DataCollectionPriortoCalibration Figure4-8. Thein-vivorecordingsetup:theratwiththeelectrodeimplantedinitsbrainisanesthetizedforconvenienceandhookedtotherecordingsystem. Figure4-9. Thein-vivorecording:eithertheTDTortheUFrecordingsystemconnectstotheelectrodesimplantedintotherattorecordactionpotentials. Therstdatacollectingexperimentisrstconductedpriortocalibration.AftertheratisanesthetizedasshowninFigure 4-8 ,theheadstageonthetopoftheimplanted 73

PAGE 74

Figure4-10. A33secondlongin-vivorecordingviatheUFrecordingsystem. Figure4-11. A30secondlongin-vivorecordingviatheTDTrecordingsystem. electrodeisconnectedbacktoeitherapre-amplieroftheTDTsystem,ortheUFanalogfront-endcircuitsystem.Figure 4-9 depictsthecompleterecordingsetup. 74

PAGE 75

ThisneuralrecordingwasconductedthroughboththeUFandTDTrecordingsystemsforabout30secondseachasshowninFigure 4-10 andFigure 4-11 .Bothrecordingsrevealclearspikes.Figure 4-12 andFigure 4-13 arethezoom-inviewsoftheUFandtheTDTrecordingsrespectively.Inthesegures,twodistinctactionpotentialsareobservedonbothrecordings. ThebackgroundnoiselevelintheUFrecordingsystem(NppUF350V)ishigherthanthatintheTDTrecordingsystem(NppTDT200V)foranumberofreasons.First,theTDTrecordingsystemhasbetternoiseisolationfortheirdiscretecircuitdesign,andisequippedwithopticalberconnections,toeliminatemostkindsofinterferencesfromtheenvironmentandgroundloop.Second,therearefundamentaltradeoffsbothinintegratingdiscretecomponentsonchipandinreducingpowerconsumption.However,thenoiseleveloftheserecordingishigherthantypicalneuralrecordingnoiselevels(20Vpp),duetoanoisierenvironmentandimperfectelectrodeimplantationintherat'sbrainforthisin-vivorecording.Figure 4-14 andFigure 4-15 arethesamplesoftheactionpotentialsclassiedinthespikesortingsoftware,Spike2TM.Figure 4-16 summarizesthepile-uptemplatesoftheclassiedspikesorts.Tworecordingsarebothclassiedastwodistinctclassesofactionpotentials. 4.3UFSystemCalibration Inthebeginningofoperatinganinstrument,acalibrationprocessisusuallytherststeptoverifythefunction.IntheUFrecordingsystem,properparametersareimportantfortheanalogfront-endtoencodethemostspikeinformation,aswellasforthedigitalback-endtoreconstructthecorrectshapeandamplitudeofactionpotentialsignals.Becausethegroundtruthprovidedbytheneuralsimulatorisknown,theneuralsimulatortestisgood,butnotthebestapproachtocalibratetheparametersrequiredintheanalogfront-endanddigitalback-endoftheUFrecordingsystem.Theoptimalproceduretocalibrateaneuralrecordingsystem,shouldadoptaperiodofrealneuralactionpotentialspre-recordedfromarat.Eventhoughthispre-recordedcalibration 75

PAGE 76

Figure4-12. AcloserviewintoFigure 4-10 .Twodistinctactionpotentialsarerevealed. Figure4-13. AcloserviewintoFigure 4-11 .Twodistinctactionpotentialsarerevealed. signalmaynotbethesameaseveryneuralrecording,thecalibratedparametersintheanalogfront-endandthedigitalback-endoughttobebetterthanthoseparameterscalibratedwiththearticialspikesgeneratedbytheneuralsimulator.Duetothedifculty 76

PAGE 77

Figure4-14. AcloserviewofthespikesortingresultfortheTDTsystemrecording.Twoclassesofactionpotentialsaresortedindifferentcolors. Figure4-15. AcloserviewofthespikesortingresultfortheUFsystemrecording.Twoclassesofactionpotentialsaresortedindifferentcolors. ofobtainingthepre-recordedneuralrecordingbeforeeveryexperiment,weutilizedtheneuralsimulatorspikestocalibrateboththeanalogfront-endanddigitalback-endintheUFsystemtoseeifyoucouldimproveupontheuncalibratedresults. ThecalibrationsetupisshowninFigure 4-17 .Theparametersfortheanalogfront-endthatrequiredoptimization,arethepositiveandnegativethresholdvoltages,the 77

PAGE 78

Figure4-16. Asummaryofpile-upclassiedsortsofactionpotentialsforbothrecordingsfromtheUFsystemandtheTDTsystem. Figure4-17. TheUFrecordingsystemiscalibratedthroughtheneuralsimulatortest. biascurrentfortherefractorycomponent,andthereferencevoltagefortheV-Iconverter.Theadjustableparametersinthedigitalback-end,forthebestreconstruction,aretheactualpositiveandnegativeintegralpeaks,thepositiveandnegativepulsewidths,theTDTsamplingrate,andtheactualcapacitancefortheintegrate-and-reprocess. 78

PAGE 79

Figure4-18. Aphotoofthein-vivorecordingconductedwiththeUFsystem. Figure4-19. Aphotoofthein-vivorecordingconductedwiththeTDTsystem. Thein-vivorecordingfromtheTDTandtheUFsystems: AftertheUFanalogfront-endsystemhasbeencalibratedthroughtheneuralsimulatortest,wethenchangetheconnectiontotheheadstageontopoftherat'sbrain.Wepickuponechanneltorecordfromavailablechannelsscreenedinthespike-searchedstep.First,theUFsystemisconnectedinordertoconductthein-vivorecordingasshowninFigure 4-18 .Then,weswitchtotheTDTrecordingsystem,to 79

PAGE 80

Figure4-20. A20secondslongoftherstin-vivorecordingwiththeTDTsystem. Figure4-21. A27secondslongofthein-vivorecordingthroughtheTDTsystem5minutesaftertherecordinginFigure 4-20 recordthesamechannelagain,asshowninFigure 4-19 .Thepurposeofrecordingthesamechannelfrombothrecordingsystems,isforthecomparisonaftertheback-endsignalprocessing.Ifthesameactionpotentialsortsareclassiedfrombothrecordings,thiswillverifythattheUFrecordingsystemiscapableofextractingthesameneural 80

PAGE 81

Figure4-22. At32secondslongofthein-vivorecording,theUFsystemrecorded5minutesaftertherecordinginFigure 4-21 Figure4-23. At32secondslongofthein-vivorecording,theUFsystem5minutesaftertherecordinginFigure 4-22 spikeswithmuchlessbandwidththantheTDTsystemneedsforfuturewirelesstransmission. Intheacutein-vivorecordingexperiment,wewouldrecorddifferentneuralrecordingsfourtimes.ThersttworecordingsarerecordedwiththeTDTsystem 81

PAGE 82

Figure4-24. ApartoftheTDTdirectrecordingsignals(bottomtrace)showsthesortedspikesaremarkedanddisplayedseparately(toptrace). Figure4-25. ApartofthereconstructedsignalsfromtheUFsystemback-end(bottomtrace)showsthesortedspikesaremarkedanddisplayedseparately(toptrace). (Figure 4-20 andFigure 4-21 ),andtheothertworecordingsarerecordedwiththeUFsystem(Figure 4-22 andFigure 4-23 ).Eachrecordingwas5minutesapartfromthepreviousrecording. 82

PAGE 83

ThenextstepistopickuponerecordingfromeachoftheTDTrecordingsystem(Figure 4-20 )andtheUFrecordingsystem(Figure 4-22 ),andsortthespikesonbothoftheminSpike2TM.ThespikesortingresultsarepresentedinFigure 4-24 and 4-25 respectively.Afterapplyingtheprincipalcomponentanalysis(PCA)algorithmtothetemplatematchedspikes,eachrecordinghasclusteredasinglegroupofspikesandemergesasasimilarspikesortwhichisexhibitedinFigure 4-26 Figure4-26. ThespikesortingresultsforboththedirectrecordingfromtheTDTsystem(left)andthereconstructedsignalsfromtheUFsystem(right).Bothresultsshowthesimilarsinglesort. TheperformanceintherecordingsthroughtheTDTsystemandtheUFsystemareinterpretedinTable4-1andTable4-2respectively.Thesignal-to-noiseratio(SNR)iscalculatedfromtheratioofthevalueofthepeak-to-peakactionpotentialstothepeak-to-peaknoiselevel. Table4-1. Theperformanceofthein-vivorecordingwiththeTDTsystemduringtheacuteexperiment ActionPotentialsVppTDTNoiseLevelNppTDTSNRTDTBW(16-bit,25KHz) Figure 4-20 900V200V13.06dB400Kbits/secFigure 4-21 800V140V15.14dB400Kbits/sec 83

PAGE 84

Table4-2. Theperformanceofthein-vivorecordingwiththeUFsystemduringtheacuteexperiment ActionPotentialsVppUFNoiseLevelNppUFSNRUFPulseRate Figure 4-22 1000V300V11.37dB28.86Kpulses/secFigure 4-23 900V250V11.48dB28.96Kpulses/sec InTables4-1and4-2,boththeTDTandtheUFrecordingsystemsarecapableofrecordingactionpotentialsover800Vforthepeak-to-peakvalue.However,theSNRintheTDTsystemrecordingishigherthantheUFsystemrecording,meaningthenoiselevelinducedfromtheUFrecordingsystemisabout3.6dBhigherthanthenoiselevelintheTDTsystemrecording.ThenoiseiscoupledfromthepulsesattheoutputoftheUFanalogfront-endboardtotherestofthecircuitonthesameboardviathegroundloop. Table4-3. Thedataanalysisoftheacutein-vivoexperimentfortheTDTandtheUFrecordingsystems Recording#DataSortedSpikesRecordingtimeFiringRate 1.Figure 4-20 (TDT)58020.13sec28.80Hz2.Figure 4-21 (TDT)58426.84sec21.75Hz3.Figure 4-22 (UF)63132sec19.72Hz4.Figure 4-23 (UF)31432sec9.81Hz TheUFrecordingsystemhasshownthecapabilityofrecordingactionpotentialsinthisacuteexperiment.ThedataanalysisofferedintheTable4-3describesthestabilityoftheUFrecordingsystem.Intheneuralrecording,avariationoftheringratemorethan10%issignicant,whichcanbecategorizedintodifferentactivities.Fromtherecording1and2fromtheTDTsystem,theringactivityclearlydropsfrom28.8Hzto 84

PAGE 85

21.75Hz.Inaddition,whenswitchingtotheUFrecordingsystem,recordings3and4alsoshowthattheringratedropsfrom19.72Hzdownto9.81Hz. Figure4-27. A30mslongtransientanalysisfor1kHzsinewaveinputsignaltotheexistingsysteminFigure 2-1 .(top)vout-representsthenegativepulsetrain,(middle)vout+representsthepositivepulsetrainand(bottom)vout2representstheintegrate-and-reresponseonthemembranevoltage. 4.4FurtherPossibleImprovements AlthoughtheUFrecordingsystemhasdemonstratedqualityneuralsignalrecordingwithlowbandwidth,thereisstillplentyofroomtofurtherimprovethisanalogfront-endperformance.Oneunstablephenomenonhasbeenobservedinthecircuitsimulationandthebenchtoptestforthecircuitbehavioroftheintegrate-and-reencodingcircuitforthemembranevoltage.Thebiphasicintegrate-and-reprocessisnotsymmetricringwithasymmetricsinewavecurrentinput,butperformsadrifttowardoneside,asillustratedinFigure 4-27 .Inthissimulation,themembranevoltage,vout2,driftstowardsmorepositiveintegrationthannegativeintegration.Asaresult,thepositive 85

PAGE 86

Figure4-28. Proposedfullydifferentialleakyrefractoryintegrate-and-reencoding(FD-LRIF)analogfront-endcircuitsystem. channel,vout+,resmorepulseswhilethenegativechannel,vout-,resfewerpules.Thebiphasicmodeeventuallywouldbecomesinglemodeandcorruptthewholesignalreconstructionprocess.Becausethissituationisnotstationary,thiserrorcannotbecompensatedwiththeback-endsignalprocessing. ThisdriftingproblemcomesfromtheV-Iconverterdesign.AssketchedinFigure 2-6 ,Vout,themembranevoltage,givesaDCbiasthroughanegativefeedbackwiththepseudoresistor,M1.Duetoleakagecurrentoccurringinthepseudoresistor(thebodyofthePMOS),atime-varyingVDSofM1increases.ThiscausesabiasofthenegativeinputpointoftheOTAintheV-Iconverterdrifts.Therefore,thecurrentoutputoftheOTAisnotsymmetricalongwithasymmetricvoltageinput. InordertoeliminatethisproblemofasymmetricV-Iconvertercurrentoutput,onepossiblesolutionistoremovethepseudoresistorusedintheV-Iconverter.TheinputoftheV-Iconverter,however,losesaDCbiasbecauseofthecapacitorinthefront,C1(Figure 2-6 ).Therefore,anadditionalbiasingcircuitisneededtoemployedhere. Anotherkillertoallkindsofbiologicalsignalrecordingsystems,isthe60Hzinterference,whichhasbeendiscussedinSection 2.6.3 .Goodsystemgrounding,indeed,dramaticallydecreasesthe60Hzinterferenceandstabilizestherecordingsystem.Moreover,takingadvantageofthecommonmodepatternofthe60Hz 86

PAGE 87

Figure4-29. A30mslongtransientanalysisfor1kHzsinewaveinputsignaltotheexistingsysteminFigure 2-1 .(top)vout gmrepresentstheintegrate-and-reresponseonthemembranevoltage,(middle)vout+representsthepositivepulsetrainand(bottom)vout-representsthenegativepulse. noise,commonsignaleliminationtechnique,couldbeasolutiontofurtherimprovetherecordingperformance. ThediscussionsabovegivebirthtoanovelsystemarchitecturefortheUFanalogfront-endcircuitsystem,asillustratedinFigure 4-28 .Afullydifferentialleakyrefractoryintegrate-and-re(FD-LRIF)encodingcircuit,ispresentedtodemonstratethecapabilityofcommon-moderejectiontominimizethe60Hznoise.Inaddition,theadventofthecommon-modefeedback(CMFB)circuitoffersDCbiasesonbothoutputsofthebio-amplierandinputsoftheV-Iconverter.Thisdesignnotonlyeliminatestheneedofthefrontcapacitor,whichwasforACcouplingpurposesinordertosavecircuitarea,butalsoremovesthefeedbackpseudoresistor,whichwasforinputDCbiaspurposestogainbettercontroloftheDCbiasthroughCMFBwhiletheoutputoftheV-Iconverter 87

PAGE 88

Figure4-30. Thetransistorleveldesignforthefullydifferentialamplierdesignwithacommon-modefeedbackcircuit. isstillbiasedviatheswitchcapacitordesign.Furthermore,thisbetterDCbiascontrolsolvestheproblemoftheunstableintegrate-and-reprocessonthemembranevoltage,asdemonstratedinFigure 4-29 .Thetransistorlevelcircuitdesignofthefullydifferentialbio-amplierandCMFBareshowninFigure 4-30 88

PAGE 89

CHAPTER5NEURONCIRCUITDESIGNFORFURTHERBANDWIDTHREDUCTION 5.1Background AswediscussedinChapter 1 ,datareductionisourmainpriorityinimplementingtheimplantedneuralrecordingsystem.Theseencodingcircuitswillbecalledneuroninthischapterbecausetheyareallneuroninspiredencodingcircuits(NIEC).AlthoughtheplainbiphasicIFneurondesigncompressesalotoftransmissionbandwidth,itstilldoesnotmeettherequirementofverylowbandwidth.ThreeexistingandproposedneuroncircuitsarepresentedasdatareductionstrategiesfortheplainIFneuron:arefractoryneuron[ 22 ],aleakyneuron[ 24 ]andanadaptiveneuron.Inthischapter,weusethetermsimpleneurontorepresenttheplainIFneuroncircuitwithoutanyofthesethreecomponentsattached. Forin-vivoextracellularneuralrecordings,multipleneuronsarerecordedonthesameelectrode.Thespikeshapecanbedistinguishedforeachoftheindividualneuralsignalsthroughaprocesscalledspikesorting.Hence,reconstructionbecomesessentialtotransferthepulsesgeneratedfromthebiphasicIFneurontothecontinuous-timepulsetrains,inordertorunconventionalspikesortingalgorithms.Sufcientconditionsforperfectreconstructionarebothbandlimitedcontinuoussignals,andmaximuminterspikeperiodsmallerthantheNyquistperiod.WhenthetimeintervalinbetweensomepulsesexceedstheNyquistperiod,onlylocalreconstructionispossible.Iftheinterspikeintervalsaretoolarge,evenwithinthesignalregions,pulse-basedspikesortingwillberequired. Insteadoffocusingonmimickingactualbiologicalneuronsfordeterminingneuroncircuitperformance[ 42 ][ 43 ],wearetheonlyresearchgroupprovidingquantitativeperformancemetricsforneuronmodelsandhardware.Ourmethodinevaluatingtheperformanceoftheneuroncircuitsistoconsiderthetradeoffofbandwidthandsignal-to-errorratio(SER)orspikesortingerror(SSE).TheSERisdenedasthe 89

PAGE 90

Figure5-1. Asimulationofpulseratev.s.inputsinewaveamplitudefordifferentrefractoryperiodsettingsofarefractoryneuron. ratioofthepowerofthesignalstothepoweroftheerror(thedifferencebetweenthecontinuoussignalsandreconstructedsignals).Fortheentireneuralrecording,onlytheactionpotentialregionsareusedtocalculatetheSERforthespikereconstructionperformanceregardlessofreconstructionaccuracyofthenoiseregions. 5.2RefractoryIntegrate-and-Fire(RIF)Neuron Inbiology,thereisamaximumringrateforneuronsbecausetheinactivationofthesodiumchannelsrestricttheneuronsfrominitiatinganotheractionpotentialforabout1ms,calledtherefractoryperiod.Additionally,anactionpotentialisrelativelydifculttoinitiateduringhyperpolarizationoftheactionpotential,alsocalledtherelativerefractoryperiod,furtherreducingtheringfrequencyofneurons.Technicallyspeaking,everyhardwareneuroncircuitmusthaveanonzerorefractoryperiod,butwedesignarefractorycomponenttolengthenandexplicitlysetthevalueoftherefractoryperiod.Todemonstratethefunctionofarefractoryneuron,weapplieddifferentamplitudesinewavesinaMATLABsimulationtotherefractoryneuronmodel,andacquireddifferent 90

PAGE 91

Figure5-2. Transistorleveldesignofrefractorybufferusedintheintegrated-and-recircuit. pulserates,asshowninFigure 5-1 .Therefractoryneuronstillretainsthetheoreticalperfectreconstructionandwon'tdegradetheSERaslongasthepulsedensityassumptionsaremet.Thesecurvesshow,thatastherefractoryperiodincreases,thepulseratedecreaseswiththesameinputsignalamplitude.Thus,therefractorymechanismcancompressthesignalbandwidthwithoutsacricingSER.Whentheinner-spikeintervals(ISIs)exceedtheNyquistperiod,theSERdegrades. ThecircuitimplementationoftherefractorycomponenthasbeenaddressedinLi'swork[ 25 ].Thiscircuitneedstofeaturevariablerefractoryperiodsettings,whichcanbedenedpost-fabrication.Therefractorycomponentisrealizedbytheasymmetriccurrent-starvedinvertershowninFigure 5-2 .ItconsistsofaCMOSpairM1/M2withanadditionalseries-connectedPMOStransistorM3inthepull-up.APMOStransistorM4isemployedfortestingpurposes.ThecontrolvoltageVbias-refractoryisusedtoadjusttherisetimeattheoutputpulse.ByadjustingVbias-refractory,wecandesigndifferentrefractoryperiodsbasedonthebandwidthconstraintsofthetransmissionchannel.Whentheinputchangesfromlowtohigh,thiscomponentworkslikeatypicalinverter 91

PAGE 92

Figure5-3. Schematicofaleakyintegrate-and-re(LIF)circuit. andthetimingofthefallingedgeispropagatedtotheoutputpreciselywithnegligibledelay.Whentheinputchangesfromhightolow,thechargingcurrentfromVddtothedrainofM2isdecidedbyM3andtherisetimeiscontrolledbythegatevoltageofM3.Comparedtootherdelaycomponents,suchastransmissiongatedelay,andcascadedinverterdelay,thispull-uptimecanbedenedpost-fabrication. 5.3LeakyIntegrate-and-Fire(LIF)Neuron Actualbiologicalneuronsinherentlypossessaleakagemechanismwhilegeneratingspikes.Thecellmembranevoltagedecreaseswhentheinputcurrentisturnedoff.Hardwareneuronsimplicityhavesomeleakcomponentsbutitistypicallytoosmalltobeusefulandnotcongurable.Inspiredfromrealneurons,theleakycomponentwasaddedinthebiphasicIFneuroncircuitstoenhancethedatareduction.Theleakycomponenteliminatesacertainamountofinputsignalandslowsdownthepulserate.TheleakycomponentturnstheIFneuronintotheleakyintegrate-and-re(LIF)neuron. 5.3.1LeakyComponentImplementation Theleakycomponentisconnectedtotheinput(membranevoltage)ofthesimpleIFneuron,inparallelwiththeintegratingcapacitor.Eitheraconstantcurrentsourceoralinearresistorcanbeusedtoimplementtheleakycomponent.Thelinearresistoractsasanadaptivecurrentsourcebytakingmorecurrentforlargersignals.Inthe 92

PAGE 93

Figure5-4. Aschematicoftheleakycomponentcircuitdesign. integratedcircuitdesign,thelinearresistordesignisnotfeasibleduetoalargeareathatisrequiredandhighthermalnoise.Thedesignoftheconstantcurrentsourceisalsonotthebestchoice,becauseiftheinputcurrentintegratedinthecapacitorismuchlargerthantheconstantleakycurrentsource,theleakycomponentisineffective.Therefore,theleakycomponentisrealizedusinganoperationaltransconductanceamplier(OTA),whichisconguredasaunity-gainfollower,asshowninFigure 5-3 ,andtheleakycurrentcanbeadjusted.Theleakycomponentsinkscurrentforapositiveinputandsourcescurrentforanegativeinput.TheOTA'sschematicisdisplayedinFigure 5-4 .TheOTA'sbiascurrent(Ib)isadesignparameterdeningthemaximumleakycurrent.Forprocessingasmallsignalscale,suchasneuralactionpotentials,asmallleakycurrentisrequiredandcanbeimplementedinaCMOSsubthresholdcircuitdesign[?].Theleakycurrentislinearwhenthevoltageonthecapacitorissmall.Wewanttheleaky 93

PAGE 94

Figure5-5. Afamilyofcurrent-to-voltagecurvesfordifferentlinearinputvoltagerangesinthesub-thresholdoperationmodefortheleakycomponent. componenttoactasaresistor,becausemoreleakycurrentresultsinhigherintegratedvoltageinthecapacitor(membranevoltage),meaningfurtherbandwidthreduction.Increasingthelinearrangeoftheleakycomponentcircuitcanleadtofurtherdecreasesinbandwidth. Insubthresholdoperation,theleakycurrent(Ileak)dependsonmembranevoltage(Vc),thebiascurrents(Ib)oftheOTAintheleakycomponent,andsomefabricationtechnologyparameters,asgiveninequation 5 Ileak=Ibtanh( 2VTVc),(5) Largerleakycurrentsleadtofurtherdatareduction.However,thechangeofthebiascurrentwillnotinuencethelinearrangeoftheI-Vcurves,asshowninFigure 5-5 .ThelinearregionoftheseI-Vcurvesrepresentaresistor-likeleakycomponent.Theslopes,gm,oftheI-Vcurvesinthelinearregioncanbeapproximatelyexpressedbyequation 5 94

PAGE 95

Figure5-6. Asimulationofpulseratesv.s.inputsinewaveamplitudesfordifferentbiascurrentdesignsettingsinaleakycomponentwithalinearinputvoltagerangeof0.5V. gm=Ib 2VT(5) ThisequationalsoexplainswhytheslopevarieswiththebiascurrentoftheOTA.However,therangeofVcinthelinearregionisnotsufcientinourapplication.IncreasingthelinearrangeofVcispreferredforfurtherdatareduction.whereisthesubthresholdslopefactorandVTisthethermalvoltage. 5.3.2LeakyNeuronModelSimulation Inordertodemonstratetheoperationoftheleakyneuronmodel,differentamplitudesinewavesat1kHzwereappliedtogeneratedifferentpulserateswithdifferentleakvaluesettings,asshowninFigure 5-6 .Thehigherbiascurrentledtofurtherdatareduction.Thelargerthebiascurrentintheleakycomponent,themorethebandwidthwasreduced. 5.4AdaptiveIntegrate-and-Fire(AIF)Neuron Theideaoftheadaptiveneuronhasbeenstudiedformanyyearsintheeldofneuroncircuitdevelopment[ 42 45 ].Earlyon,Mahowaldetal.,mimickedabiological 95

PAGE 96

Figure 5-7.Asimulationdemonstratestheoperationofanadaptiveneuronmodel. neuronadaptionmechanism,[ 42].Later,Schultzetal.[43 ],andIndiveri[ 45],also proposedadaptivefunctionsfortheIFneuroncircuit.Thelasttwopapersclaimsimilar mechanismstoadaptneuroncircuitringbyplacingaconstantleakycurrentsource inanadditionalfeedbacklooptotheinputintegralcapacitor.Wheneverapulseis generatedattheoutput,thecircuitwilltriggeraswitchtoturnonaleakycurrent sourcetosinkacertainamountofcurrent,whichextendstheintegrationtime.The disadvantageofthismechanism,isthattheirintegratedvoltagesareadaptiveintwo levels,notmultiplelevels,becauseaconstantcurrentistriggeredtodropfromthe capacitorwhenapulseres.Theneuronthresholdvoltageisxedduringoperation.All oftheseconditionsmakethereconstructionprocessdifcultinourapplicationwithspike sortingintheback-end. Tosolvethisproblem,weproposeadifferentconceptofdesignfortheadaptive componentofthebiphasicIFneuron.Thisadaptivecomponentcanadjustthering ratedynamicallydependingonthepulserate.Therearetwopossiblesolutions.One 96

PAGE 97

Figure5-8. Pulseratesv.s.inputsignalamplitudeswithan1KHzinputsinwaveanddifferentincrementalvoltagethresholdadaptation(Vthstep)foreachpulseintheproposedadaptiveneuronmodel. istovarytheV-IconvertergainfordifferentpulseringratesandtheotheristoadaptthethresholdvoltageofthecomparatorsintheIFneuroncircuit.Wechosethesecondmethodduetoitssimplerimplementation. TheadaptivecomponentisdesignedtoadaptthethresholdvoltagefortheIFneurondependingontheoutputpulserate.Thisadaptivecomponentispulsecountsensitive,andincreasesthethresholdvoltagebyaxedamount,whenapulseisgenerated.InthisbiphasicIFneuronarchitecture,ringinonechannelwillresetthethresholdvoltageoftheotherchannelbacktoitsinitialvalue.Figure 5-7 demonstratestheoperationalconceptoftheadaptiveneurondesign.A1kHzsquarewavewasusedasaninputsignal.Thedottedlinemeansthesignalisnotshowninscale.Thebiphasicencodedpulsesspreadoutwiththesameamplitudeofinputsduetotheadaptingthresholdvoltages. 97

PAGE 98

Figure5-9. Acomparisonofpulseratesgeneratedfromdifferentthresholdvoltagesbetweenanadaptiveneuronandasimpleneuronforasinewaveinputwithanamplitudeof500V. Inordertocharacterizethedesignedadaptiveneuronmodel,differentamplitudesinewavesat1KHzwereappliedinaMATLABsimulation,todemonstratethedatareductionachievedwithdifferentparametersettings,asshowninFigure 5-8 .Vthstepistheamountthethresholdischangedateachstepofadaptation.Themorethethresholdvoltageincreasesateachstep,themorethepulserateisreduced.Duetohardwarelimitations,theadaptedthresholdvoltageshavetheirownhighandlowbounds.ItisnotefcienttosetVthsteptoohighsincetheboundsarereachedtooquickly,andthesignalinformationwouldnotbefullyretained.Onedisadvantageoftheadaptiveneuroncurrentdesign,isthatinformationcanbelostwhensignalsdropfromtheirpeakvalue,beforethesignalpolarityswitches,becausetheadaptivethresholdvoltageswouldstillremainhigh.Theadaptivethresholdvoltageinonechannelcouldberesetbackonlywhenanotherchannelisactivatedinthismodel.Therefore,furtherimprovementsarerequiredandwillbediscussedinsection 5.5.2 98

PAGE 99

Figure5-10. Asegmentofrecordedactionpotentialsignalsusedfortestingdifferentneuronmodelsinthischapter.Twoactionpotentialsaremarkedinred. Figure5-11. AcomparisonofSERv.s.pulseratesbetweenanadaptiveneuron(blue)andasimpleneuron(red)forapplyingdifferentthresholdvoltages. Nextwekeepthesameinputsinewaveamplitude,500V,andchangetheinitialthresholdvoltagesforboththesimpleneuronandtheadaptiveneuron.Figure 5-9 indicatesthattheadaptiveneuroncansavemorebandwidththanthesimpleneuron,especiallyatalowthresholdvoltagesetting.However,thetruebandwidthsavingscan 99

PAGE 100

Figure5-12. AcomparisonofSERv.s.pulseratesbetweenanadaptiveneuron(blue)andasimpleneuron(red)forapplyingdifferentscalingfactorstotheactionpotentials. Figure5-13. Anoverlapplottingofrecordedandreconstructedsignals(top)andpulsesgeneratedbyanadaptiveneuron(bottom). onlybejustiedifnecessaryinformationispreserved.Therefore,thetradeoffbetweenSERandpulseratemustbeconsideredtounderstandthetrueperformanceofthedatareduction. 100

PAGE 101

Figure 5-14.Anoverlapplottingofrecordedandreconstructedsignals(top)andpulses generatedbyasimpleneuron(bottom). Aftertestingtheadaptiveneuronwithsinewavesforproofofconcept,wewill examinetheadaptiveneuronwithrealneuralrecordingsignalsasshowninFigure 5-10, wheretwoactionpotentialsarerecordedwithassociatednoise.Figure 5-11 represents theperformanceofdatareductionbetweentheadaptiveneuronandthesimpleneuron withdifferentthresholdvoltagesettings.Theadaptiveneuronoutperformsthesimple neuronintermsofdatareduction,whilestillretainingasimilarSER.Thisbehavior becomesmoreapparentathighpulserates.Inthisexamination,theadaptiveneuron reducesthepulserateover40 % fromthesimpleneuronathighSER. 5.4.1PerformancewithDifferentSNR Inordertocomparetheperformancebetweentheadaptiveneuronandthesimple neuronwhentheSNRoftheactionpotentialsarescalable,Figure 5-12 ,revealsthatthe adaptiveneuronhasbetterbandwidthefciencythanthesimpleneuronathighSER. BoththeadaptiveneuronandthesimpleneuronachievesimilarpulseratesatlowSNR, whiletheadaptiveneuronreducesmorepulseratesthanthesimpleneuronathigh 101

PAGE 102

Figure 5-15.Anoverlapplottingofrecordedandreconstructedsignalswithaction potentialsmarkedinred(top)andpulsesgeneratedbyanadaptiveneuron (bottom). Figure 5-16.Anoverlapplottingofrecordedandreconstructedsignalswithaction potentialsmarkedinred(top)andpulsesgeneratedbyasimpleneuron (bottom). SNR.Figure 5-13 and 5-14 showthatbothneuronscanperformhighSERwhilethe adaptiveneurongeneratesfewerpulses. 102

PAGE 103

Figure5-17. AcomparisonofSERv.s.pulseratesbetweenanadaptiveneuron(blue)andasimpleneuron(red)forapplyingdifferentthresholdvoltages. Figure5-18. Acomparisonofpulseratesgeneratedwithdifferentthresholdvoltagesbetweenanadaptiveneuronandasimpleneuronfortherecordedactionpotentialsassociatedwitha90Hz,100Vsinewave. 5.4.2PerformancewithSlowDCDrift Inarealneuralrecording,aslowDCdriftiscommonandcancauseexcessivebandwidthintheencodedsystem.Inthissimulation,weadda100V,90HzsinewavetotherecordedsignalsinFigure 5-10 ,toformanewtestinput.Figure 5-15 and 103

PAGE 104

Figure 5-19.Anoverlapplottingofrecorded(blue)andreconstructed(red)signals wheretherstactionpotentialisscaledbythreetimes(top)andpulses generatedbyasimpleneuron(bottom). Figure 5-20.Anoverlapplottingofrecorded(blue)andreconstructed(red)signals wheretherstactionpotentialisscaledbythreetimes(top)andpulses generatedbyanadaptiveneuron(bottom). 5-16,showthepulsesgeneratedbytheadaptiveneuronandthesimpleneuronand theircorrespondingreconstructionresults.Again,theadaptiveneurongeneratesfewer pulsesthanthesimpleneuron.InFigure 5-17 ,theadaptiveneurondominatesthe 104

PAGE 105

Figure5-21. AcomparisonofSERv.s.pulseratesbetweenanadaptiveneuron(blue)andasimpleneuron(red)forapplyingdifferentinitialthresholdvoltageswiththeinputtestingdatawheretherstactionpotentialisscaledbythreetimesforthelargedynamicrangetest. Figure5-22. Acomparisonofpulseratesgeneratedwithdifferentinitialthresholdvoltagesbetweenanadaptiveneuronandasimpleneuronfortheinputtestingdatawheretherstactionpotentialisscaledbythreetimes. datareductionoverthesimpleneuronatasimilarSER.Figure 5-18 ,emphasizesthepromisingdatareduction,whichshowsthattheadaptiveneuronproducedaverylimitedincreaseofpulseratewhenthethresholdvoltagedecreases. 105

PAGE 106

5.4.3PerformancewithLargeDynamicRange Foranextracellularneuralrecording,anelectrodemayrecordmultipleneuralactionpotentialsfromdifferentdistancestotheelectrode.Thisleadstoasituationwhererecordedactionpotentialsmayhavealargedynamicrange.Theadaptiveneuroncanencodesmallactionpotentialsefciently,withoutrequiringanexcessivepulseratetoencodemuchlargeractionpotentials.ReferringtoFigure 5-19 ,theadaptiveneurongeneratesfewerpulsesshowninFigure 5-20 ,thanthesimpleneuron,andreducesthebandwidthefcientlytoretaintheinformationofbothactionpotentials.Figure 5-21 veriesthattheadaptiveneuroncansave50%ofthebandwidthat0.5Vinitialthresholdvoltage(themostrightdatapointforeachneuroncircuits),atasimilarSERtothesimpleneuron.AlowSERatlowpulseratesisexpectedandwillnotsupportreconstructionwell.Inaddition,agreaterdifferenceinpulseratesbetweenthesetwoneuronscanbefoundwhenalowerthresholdvoltageisapplied,asshowninFigure 5-22 5.5AdaptiveNeuronCircuit Figure5-23. Thecircuitcongurationofanadaptivecomponentcombinedtothedevelopedneuralrecordingsystem. 106

PAGE 107

Figure5-24. Theconceptualcongurationofthecircuitdesignforanadaptivecomponent. AnadaptivecomponenthasbeendesignedforthebiphasicIFneuronasdepictedinFigure 5-23 .ThetwooutputpulsetrainsoftheIFneuronarebothfedintotheadaptivecomponent,whichformsfeedbackloopstodenetheinstantthresholdvoltagesforeachcomparator.Theadaptivecomponentadjuststhepositivethresholdvoltagesaccordingtothepositivepulsetrain,andthenegativethresholdvoltagesaccordingtonegativepulsetrain.Theintensiveringincreasesthethresholdvoltagerapidlytorestrainfurtherring.Thepositivethresholdvoltageraisesbyastepvoltageforeachcomingpositivechannelpulse.Thenegativethresholdvoltagedecreasesbyastepvoltageforeachcomingnegativechannelpulse.Eachofpositiveandnegativeadaptivevoltagescanreachitssaturationvalueduetothehardwarelimitation. 5.5.1AdaptiveCircuitDesign ThedesignprincipleoftheadaptivecomponentistoprovideanegativefeedbacktothebiphasicIFneuroncircuit.Thenoveltyofthisdesignistoself-adjustthethresholdvoltageaccordingtothepulsetrain.Thedynamicthresholdvoltagewillberesettotheinitialthresholdvoltagewhentheotherchannelstartsringpulses.TheconceptualcongurationofthecircuitdesignisshowninFigure 5-24 .InthisFigure,IpandInare 107

PAGE 108

Figure5-25. Theschematicdesignofthebiphasicneuroncircuitwithanadaptivecomponentandarefractorycomponent. user-denedcurrentswhichdeterminehowrapidthedynamicthresholdvoltagescanchange.M1andM4serveasswitchescontrolledbyincomingpulses,andM2andM3serveasresetdevices.Eachnegativepulsetriggerstheswitch,M4,tosinkaxedamountofchargefromthecapacitorduringpulseperiodanddecreasethenegativethresholdvoltage,Vref N.Moreover,thenegativepulsetriggersM2toresetpositivethresholdvoltage,Vref P,backtotheinitialpositivethresholdvoltage,Vref0 p.PositivepulsesfollowexactlythesameprincipleforthecounterparttoinuenceVref PandresetVref N. ThecircuitsimulationinCadenceRhasveriedthedesign,asillustratedinFigure 5-25 andFigure 5-26 .Anadvantageofsuchadaptiveneuroncircuitdesignistoalleviatedifcultyofadjustingtwoparameters:theleakycurrentandthethreshold 108

PAGE 109

Figure5-26. Asimulationresultofthebiphasicneuroncircuitwithanadaptivecomponentandarefractorycomponent:(A)asingletonesinewaveinput,(B)theoutputofthebio-amplier,(D)apulsetraingeneratedinthepositivechannel,(E)apulsetraingeneratedinthenegativechannel,(F)thedynamicpositivethresholdvoltagerespondingtotheinputpositivechannelpulsetrain,and(G)thedynamicnegativethresholdvoltagerespondingtotheinputnegativechannelpulsetrain. voltageforoptimalperformanceatthesametime.Instead,onlytheleakycomponentneedstobefocusedduringasignalrecordingbecausethethresholdvoltagesareself-adjusted. 5.5.2ImprovedAdaptationMechanism Thekeyadvantageofthisadaptivethresholdvoltagedesignistoprovideadiscretemulti-levelthresholdadaptation,whichcanberecoveredintheback-endforaccuratesignalreconstruction.Unlikeotheradaptiveneurondesigns,thetimingofadaptationmaynotbeclearlyknown,leadingtotheproblemsintheback-endsignalreconstruction.Thisdesign 5-27 (a),however,mayneedanimprovementfortherecordingsignalsthat 109

PAGE 110

Figure5-27. Circuitdiagramoftheadaptivecomponents. arenotzeromean,butwithaDCoffsetinonepolarity.Thiswillcauseringinonlyonechannelandthethresholdoftheotherchannelmayneverberesetbacktoitsinitialvoltage.Inoursystem,theinputiscapacitivelycoupledsothatweexpecttheinputsignaltobeapproximatelyzeromean.Inthecaseofanon-zero-meansignal,arevisedadaptivecomponentdesigncanbeusedasshowninFigure 5-27 (b).Anadditionalleakycurrentsourceisaddedandstaysonallthetime.Thebiphasicchannelsareindependentandeachchannelisnotinuencedbytheringintheotherchannel.BecauseIpandIncanbesetmuchlargerthanIp subandIn subindividually,thethresholdadaptationwillnotbeinuencedverymuchandthethresholdvoltagewilldecaybacktotheinitialvalueifnospikeisgeneratedforalongtime. Figure 5-28 illustratestheactualthresholdadaptionmechanism.First,currentsignals,i(t),startbeingintegratedwiththetimeconstant,V,onthecapacitortoformthemembranevoltage,V,asshownin( 5 ). dV dt=)]TJ /F14 11.955 Tf 10.68 8.09 Td[(V v+i(t)(5) 110

PAGE 111

Figure5-28. Thediagramshowstheconceptoftheimprovedadaptivecomponentmechanism.Theplotinredrepresentstheadaptivethresholdvoltageandtheplotinbluerepresentstheintegrationprocessonthecapacitor.Pulsesaregeneratedwhentheintegratedvoltagereachestheinstantthresholdvoltage. Figure5-29. TheSPICEsimulationoftheproposedlow-bandwidthadaptivebiphasicneuronusingtheadaptivecomponent 5-27 (b). Whentheintegratedvoltage,V(t1),reachestheinitialthresholdvoltage,S0,anewpulsewillbegeneratedatt1.Inthemeantime,thethresholdvoltagewillraisebya 111

PAGE 112

pre-denedvoltageStobeS(t1),asshownin( 5 )and( 5 ). V(t1)=S0(5) S(t1)=S0+S(5) Beforethenextpulseisgenerated,thedynamicthresholdvoltageisdecayingwiththetimeconstant,S,duetoabuilt-inleakysourceuntiltheintegratedvoltage,V(t),reachesagainS(t2)att=t2,asshownin( 5 ).Thenthenextpulseisgenerated. dS dt=S(t2))]TJ /F14 11.955 Tf 11.96 0 Td[(S(t) S(5) AsimulationinCadenceRhasveriedthedesigninFigure 5-27 (b),asillustratedinFigure 5-29 Figure5-30. AcomparisonofSERv.s.pulseratesbetweenanadaptiveleakyneuron(blue)andasimpleneuronwithaleakycomponent(red)fordifferentbiascurrentsfrom10nA100nAandtheleakycomponenthasalinearinputvoltagerangeof0.5V. 112

PAGE 113

5.6AdaptiveLeakyRefractoryIntegrate-and-Fire(ALRIF)Neuron Ifwecombinetheleakycomponentwiththeadaptiveneuronandasimpleneuron,thedatareductionismorepronounced.Figure 5-30 showstheperformanceoftheadaptiveleakyneuronandthesimpleneuronwithaleakycomponentattheleakvaluesrangingfrom10nAto100nAandthelinearrangeissetto0.5V.Theadaptiveleakyneuronshowsmorepulseratereductionthanthesimpleneuronwithaleakycomponent.Thisgurealsoshowsthatwhenthepulseratedropstoolow,theSERstartsdroppingforbothneurons.Asaresult,reconstructionmaybeproblematicwiththeleakycomponent. Figure 5-31 and 5-32 showthattheleakycomponentdramaticallyreducesthepulseratebydroppingmanypulsesforthenoiseregionoftherecording.Nevertheless,thesignalscannotbereconstructedduetolowpulserates,whichforcesustodevelopthepulse-basedspikesortingmethod,withoutsignalsbeingreconstructedrst. Theoptimalleakinesssettingfortheleakycomponentisrelatedtothenoiseooroftherecording,andtheoptimalthresholdvoltagedependsontheamplitudesoftheactionpotentials.Theoreticalmodelsimulationshavebeengivenaboveforthesethreebandwidthreductionstrategies.Consequently,wedecidetocombinealloftheiradvantagesandintegratethemintotheplainIFneurontoformanadaptiveleakyrefractoryintegrate-and-re(ALRIF)neuronusedintheUFrecordingsystem.Therefractorycomponentsetsamaximumringrate.Theleakycomponentformsalowpasslterwiththecapacitortodroppulsesrepresentinghighfrequencynoise,leadingtofurtherdatareduction.Theadaptivecomponentcanrestrainpromptincreaseofthepulserateduetodifferentspikeamplitudesinthesamerecording,toenlargethedynamicrangeandreducethebandwidthefciently. 5.7ALRIFNeuronBenchTopTest Afterintroducingtheanalogfront-endhardwarecircuitdesignandintegrationinChapter 2 ,andback-endsignalprocessinginChapter 3 ,thewholeUFneuralrecording 113

PAGE 114

Figure 5-31.Anoverlapplottingofrecorded(blue)andreconstructed(red)signals. Signalscannotbereconstructed(red)perfectlywhenpulsesarenot sufcient(top)andpulsesgeneratedbyaleakysimpleneuron(bottom). Figure 5-32.Anoverlapplottingofrecorded(blue)andreconstructed(red)signals. Signalscannotbereconstructed(red)perfectlywhenpulsesarenot sufcient(top)andpulsesgeneratedbyanadaptiveleakyneuron(bottom). systemiscomplete.Thenextstepwouldbetofocusonverifyingtheperformanceof theUFneuralrecordingsystem.Abench-toptestshouldbeconductedrstinthesame mannerasshowninFigure 5-33.Signalssourcessuchasasignalgeneratororthe 114

PAGE 115

Figure5-33. Abench-toptestingsetup:anindependentsignalsourcepumpsanalogsignalsintothecircuitboardundertesting,ampliedoutputcouldbemonitoredintheoscilloscopeandencodedpulsesarerecordedthroughtheUSBinterfaceintothecomputerfortheback-endsignalreconstruction. neuralsimulatorarepreparedtoprovidetheinputsignaltotheUFanalogfront-endrecordingsystemboard.Thebio-amplierchipoutputandencodedpulsesattheARLIFchipoutputontheboardcouldbemonitoredintheoscilloscope.TheoutputpulseswillberecordedintothePCaseventsthroughtheUSBboardforthereconstructionandspikesortingintheback-end. AdoptingtheimprovedadaptivecomponentintotheleakyrefractoryIFneuroncircuit,Figure 5-34 ,showstheARLIFchipmeasurementresultforasine-wavebenchtoptest. Figure 5-35 ,performsthenewadaptivecomponentfunctionemployedintotheARLIFneuroncircuitwiththesourceoftheneuralsimulator.Inthisoscilloscopedisplay,onlyoneactionpotentialisshown.Thetopandbottomtraces,whichrepresentpositiveandnegativethresholdvoltagesrespectively,adaptuponeachintegrate-and-recycle.Themiddletraceshowsthemembranevoltageresponse,whereeachintegration 115

PAGE 116

Figure5-34. TheimprovedadaptivemechanismemployedintoanALRIFneuroncircuitmanifestsontoaoscilloscopedisplayforabenchtoptestwithasinewaveinput:topandbottomtracesrepresentboththepositiveandnegativeadaptivethresholdvoltagesrespectively.Themiddletracerepresentsthemembranevoltagetoshowtheintegrate-and-reprocessundertheadaptationcondition. Figure5-35. Anoscilloscopedisplayshowsthethresholdvoltageadaptationcorrespondingtoeachintegrate-and-recycleonthemembranevoltagefortheALRIFneuroncircuit. cyclereachesdifferentpeaksaccordingtoadaptivethresholdvoltages.Figure 5-36 ,demonstratesabenchtoptestforexaminingtheARLIFneuronanalogfront-endcircuitsystem.Threedistinctactionpotentialsaresuccessfullyreconstructed.Biphasicpulsesappearonpositiveandnegativesidesclearlycenteredatactionpotentials. 116

PAGE 117

Figure5-36. Top:recordedbiphasicpulsesfortheALRIFanalogfront-endrecordingsystemwiththeneuralsimulatorinput.Bottom:reconstructedsignalsintheback-end.Threeseparateandclearactionpotentialsarerevealed. Aspikesortingprocesswouldthenbenecessarytoverifythattherecordingtrulyhasthreedifferentclassesofactionpotentials.Figure 5-37 ,showsthespikessortedinSpike2TMandFigure 5-38 presentsthesethreepile-upsortsofclassiedactionpotentialsfromtheneuralsimulatorviatheALRIFanalogfront-endrecordingsystem,concludingthatthisrecordingissuccessfulforthepurposeofspikesorting. 5.8Discussion AlthoughtheALRIFneuronhasdemonstratedreliablerecordingfromtheneuralsimulatortest,thereisstillroomtofurtherimproveitsperformance.Thefollowingsectionswillpresentaproposedideatothecircuitimprovement,discusslimitationsretainingrecordingquality,systemcalibrationandpulsedensityestimation.Thesolutiontominimizereconstructionerrorisintroducedinthecalibrationprocess. 5.8.1WideDynamicRangeAdaptiveComponentDesign AnOTAself-connectedasabufferisutilizedintheleakysourcedesign,whosenegativeinputisshortedbacktothecapacitorstoringtheinstantthresholdvoltagewhile 117

PAGE 118

Figure5-37. Theback-endspikesortingprocessisexecutedinSpike2TM.Threekindsofsortedspikesaremarkedindifferentcolors. Figure5-38. Threedistinctpile-upwaveformsofsortedspikesfromthereconstructedsignalsrecordedfromtheALRIFanalogfront-endrecordingsystem. thepositiveinputisconnectedtoaxedvoltage,asillustratedinFigure 5-39 .Whenthedynamicthresholdvoltageisverymuchdifferentfromthexedvoltagesettothepositiveinput,theOTAdoesnotstayinthelinearoperationmodebutswitchestolargesignaloperation.Onesideofthedifferentialinputstageturnsoffcompletely,whiletheothersideconductsthemajoritybiasingcurrent.Asaresult,theleakysourcecannotconductconstantleakycurrent,butdumpslargernonlinearcurrenttoretainfurtherthresholdvoltageadaptation,whichleadstoanerrorforthethresholdvoltagepredictionintheback-end. 118

PAGE 119

Figure5-39. Theschematicofthewidedynamicrangeadaptivecomponentdesign. ThekeytoincreasingtheadaptivethresholdvoltagedynamicrangereliesonthedesignoftheOTAdesignusedintheleakysourceasshowninFigure 5-39 .Inordertopreventturningtoalargesignalmode,thisOTAshouldhavealargerlinearrangedesign.ThetechniqueweusedisthesourcedegenerationintheinputdifferentialstageoftheOTA.AsshowninFigure 5-40 ,NMOStransistors,N5andN6,playtheroleofsourcedegeneration.DiodeconnectedN5andN6bothhavetheresistance,1 gm. 119

PAGE 120

Increasingresistanceforlargelinearitynecessitatesthesmallersize(W L)designforthesetwotransistors,whereW=1.5mandL=15m. Figure5-40. TransistorleveldesignforthelargerlinearityOTAdesign. Figure 5-41 demonstratestheappealingimprovementofthresholdvoltageadaptation.ComparedtothesimulationinFigure 5-26 ,themaximumdynamicrangeofthethresholdvoltageadaptationwaspushedfrom400mVto1.3V. 5.8.2ALRIFNeuronSystemFundamentalLimitation FurtherdatareductionhasbeenobservedinSection 5.7 fortheadaptiveneuronencodingcircuit.However,theissuewhichshouldbediscussed,iswhatfundamentallimitationcausedtheSERtostartdegradationupondatareduction.Thisfundamentallimitationraisesthefollowingdifferentfactors: 120

PAGE 121

Figure5-41. ThesimulationofthetransientresponseoftheALRIFneuroncircuitwithawiderdynamicrangeadaptivecomponentdesign:(top)vout)]TJ /F1 11.955 Tf 7.08 1.79 Td[(gmrepresentsthemembranevoltageforadaptedintegrate-and-reresponse,(middle)vref-representstheadaptivethresholdvoltageforthenegativechanneland(bottom)vref+representstheadaptivethresholdvoltageforthepositivechannel. ErrorEstimationofthresholdvoltage Therearetwosystemiclimitationsources.Therstsourceistheinstantthresholdvoltage.Thethresholdvoltageappliedinthereconstructionprocessdoesnotmatchtheactualinstantthresholdvoltagethustriggeringtherecordedoutputpulse.Thereasonforthis,isbecausethedynamicthresholdvoltageusedintheback-endsignalprocessingiscalculatedfromthetheoreticalcircuitbehaviormodelasshownin( 5 )and( 5 ).Inthiscircuitbehaviormodel,thetimeconstant,RC,isattingparameterwhichmaynotttheactualexacttimeconstant.Thismismatchwouldresultintheerrorestimationoftherealthresholdvoltagetriggeringthenextpulse. 121

PAGE 122

Vthnew=Vthold+V(5) Vth(t)=Vthnewexp(t RC)(5) Thesecondsourceisthenon-idealleakycurrentsourcedesignedintheadaptivecomponent.Section 5.8.1 discussedthisproblemandprovidedasolutiontothisissue.ThesolutionshowedthatanimprovedOTAwithalargerlinearrangeprovidedconstantleakycurrents,whilethedifferenceoftheOTAinputincreasessothatthedynamicrangeoftheadaptivethresholdvoltageincreases. CircuitHardwareImperfection Thecomparatorcircuit(Figure 2-7 )usedintheencoderdesignisnotperfect.Oneinputofthecomparatorisconnectedtoaxedthresholdvoltagewhileanotherinputisconnectedtothemembraneintegratedvoltagewhichisdynamic,asshowninFigure 1-6 .Thisconditionsigniestheunbalanceddifferentialinputinthecomparatorduetodifferentbiasestoinputs.Asaresult,thecomparatorwouldswitchthestateandgenerateanewpulsebeforethemembraneintegratedvoltagereachesthethresholdvoltage.Therefore,thethresholdvoltageforthesignalreconstructionwouldnotbethesameastheactualthresholdvoltageforeachgeneratedpulse,whichdegradesSER.Theotherreasoncausingthethresholdvoltageerrorisduetothehysteresiseffectinthedecisionstageofthecomparator.Propertransistorsizedesigninthedecisionstagecouldminimizethehysteresiseffect. Therearealsootherlimitations.Forinstance,amismatchcouldoccurduetovariablefabricationprocessesanderrorintroducedfromthereconstructionalgorithmitself,allofwhichwoulddegradetherecordingquality. 5.8.3Calibration ThefactorsresultinginthedistortionoftherecordingqualityhavebeendiscussedinSection 5.8.2 .Thesediscussionsmagnifytheimportanceofthecalibrationprocess 122

PAGE 123

forarecordingsystem.Calibrationisalwaystherststepbeforeusinganytypeinstrument.Theprincipleofcalibrationintherecordingsystem,istotargetaknownsignalinputandoutput.Forinstance,atypicalsinewave,whoseamplitudeandfrequencycanbecontrolled,playstheinputtotherecordingsystem.Comparingthereconstructedsignaltotheknowninputsinewave,appropriatetuningofallparametersusedinthereconstructionalgorithmcanmakereconstructedsignalssimilartotheinputsignalinordertokeepthequalityrecording.Thentherecordingsystemiscalibrated.Forfurtherrecordings,parametersshouldbekeptthesameforsignalreconstruction. 5.8.4PulseDensityAnalysis Themainpurposeofthisproposedanalogfront-enddesignistoreducethepulserate.Therefore,derivingthemodelofthepulsedensityisessentialtorealizethefactorsreducingthedatabandwidth,asdepictedasfollows: Ifaspikehasredattimet0,anotherwilloccuratt0+t.tconsistsoftwoparts,asshownin( 5 ). t=t1+t2(5) t1isthepulsewidthlastingfromt0tot1,whichiscontrolledbytherefractorycomponent[ 25 ].Therefractorycomponentsetsupthetimeforarefractorycurrent,Iref,tointegrateoveraparasiticcapacitor,Cparbeforeapulselevelswitches,Vpulse. Vpulse=1 CparZt0+t1t0Irefdt(5) t1=VpulseCpar Iref(5)t2representstheintegrationtimetoreachtheadaptivethresholdvoltage,adapwiththeleakagecurrent,Ileakbeforethenextspike. adap=1 CZt1+t2t1[i(t))]TJ /F14 11.955 Tf 11.95 0 Td[(Ileak]dt(5) 123

PAGE 124

Deningf(t)tobetheintegralofi(t)andusingarst-orderTaylorseriesexpansion,wecanrewrite( 5 )as: adap1 C[f(t1)+t2df(t1) dt)]TJ /F14 11.955 Tf 11.96 0 Td[(f(t1))]TJ /F8 11.955 Tf 11.96 0 Td[(t2Ileak](5) t2adapC [i(t1))]TJ /F14 11.955 Tf 11.95 0 Td[(Ileak](5) Therefore,wecanrelatetheresultingpulsedensityintermsofallcircuitdesignparametersas: PulseDensity/1 t1 VpulseCpar Iref+adapC [i(t0))]TJ /F7 7.97 Tf 6.59 0 Td[(Ileak](5) Becauseencodedspikesneedtobereconstructedbacktocontinuoussignalsintheback-end,thereconstructionprocessdeterminesthelowestbandwidthboundfortheneuroncircuitdesigntomaintainacertainleveloftheSER.Indeed,simplyincreasingthexedthresholdcaneasilyreducethebandwidth;however,theSERofthereconstructionsignalwoulddegrade. 124

PAGE 125

CHAPTER6CONCLUSIONS Alowbandwidthpulse-basedneuralrecordingsystemhasbeenfabricatedandtested.Thissystemconsistsofthefront-endrecordinghardwareandtheback-endsignalprocessingsoftware.Theuseofapulserepresentationintheneuralrecordingsystemperformspromisingdatareduction.AnalogVLSIcircuitrychosentoimplementthefront-endhardware,ensuressmallsizeandlowpowerconsumptionfortheimplantedneuralrecording. Abench-toptestingwasconductedfortheasynchronousbiphasicpulse-basedleakyrefractoryIFneuroncircuitsystem,onabreadboard,withaneuralsimulatorasasignalsource.OpticalcouplersandbettersignalcontactsareappliedintherecordingsystemforreducinginputreferrednoisedowntoNpp40V.Thissystemisexaminedinaparallelrecordingplatformalongwithacommercialrecordingsystem(TDT).Thespikesortingresultsofhighdetectionandpatternrecognitionrates,showninSpike2TM,veriesthattheperformanceoftheUFrecordingsystemiscomparablewiththecommercialproduct,andahighperformancespikesortingispossible.TheLow-powerandnoise-resistantIFneuroncircuitonabreadboardcanpreformaslowas3Kpulses/secfora10-secondlongneuralsimulatorrecordingtest,andcouldreplacean6-bitADCatasamplingrateof25KHz,whichneedsabandwidthof150Kbits/sec.Thisdemonstrationtakesusastepforwardtoafullyintegratedcircuitandbattery-poweredsystem. Theimprovedpulse-basedspikesortingalgorithmhasbeenpresentedindetail.Itshowscompatibledetectionandpatternrecognitionrateswiththeconventionalspikesortingmethodwhichrequiresreconstruction.Recordedneuralsimulatorsignalswereusedtocharacterizetheperformanceoftheimprovedalgorithm.Thenoveltyofthisproposedalgorithmistosystematicallygeneratetemplatesbasedontheincomingdataset.Averagedtemplatesaregeneratedwiththesortingerrortolerancesetinthetraining 125

PAGE 126

phase.Thepulsestrainsofspikescanbeclassiedaccordingtothedenedtemplates.Thesystematictrainingalgorithmdemonstrates99%detectionand100%classicationrates,withhighSNRneuralsignalsintheneuralsimulatortest. Thepurposeofthisresearchistoestablishalowbandwidthpulse-basedneuralrecordingsystemforneurophysiologyresearch.Themaincontributionofthisdissertation,istorevealthecapabilityandfeasibilityofwirelesslytransmittinglargeamountsofrecordingchannelsfromanelectrodeimplantedintoarat'sbrain.Fromtheexperimentalresults,theIFneuroncircuithasbeenshowntoreplaceatraditionalADCintheneuralrecordingapplication.Thebenchtoptestdemonstratesthefunctionoftheanalogfront-endhardwarecircuitsystem.Thein-vitroparallelrecordingtestconrmsthecapabilityoftheUFrecordingsystem.Furthermore,theanimalrecordingfurtherveriesthefeasibilityofain-vivoneuralrecording.Inaddition,thespikesortingresultsuggeststhattheUFrecordingsystemrecordssimilaractionpotentialstotheTDTrecordingsystemwithlowerbandwidthexpenselessthan30Kpulses/secinthein-vivoexperiment.TheUFsystemiscapableofrecordingclearactionpotentials,aswellasrecordingSNRatabout11.43dB,whiletheTDTsystemcanrecordactionpotentialsatabout15.03dB.ThenoiselevelintheUFsystemisabout3.6dBwhichismorethanthatintheTDTsystem.Thisnoisesource,isattributedtothehighlyintegratedcircuitboardintheUFanalogfront-end,wheretheoutputpulserepresentationcouplesbacktherestofcomponentsontheboard. Thepulse-baseddatacompressionmodelofadaptiveleakyrefractoryintegrate-and-reneuroncircuit(ALRIF),isrealizedbyaddingthreedifferentneuroncircuitstoaplainbiphasicIFneuroncircuitdesign:refractoryneuron,leakyneuron,andadaptiveneuron.Theadaptiveneuronimprovestherecordingsystembysupportingthewidedynamicrangeofspikeswithlowbandwidth.Therefractoryneuronsetsthemaximumbandwidthoftherecording,whileretainingtheneuralinformation.Theleakyneuroncanreducethedatabandwidthbylow-passlteringthehighfrequencynoiseandeliminatingpulsesin 126

PAGE 127

thenoiseregion.TheadaptiveneuroncircuitwasdesignedinCadenceR.Thisneuronissensitivetothepulserateandself-adjuststhethresholdvoltagesforfurtherbandwidthreduction.ThebiphasicIFneuroncircuitwithrefractory,leaky,andadaptivecomponentsdemonstratesproperfunctionalityinthesimulation. Thisdissertationcontributestotheintroductionofnoveladaptiveneuroncircuitimplementationforfurtherdatareductionintheneuralrecordingapplication.Theconceptofadaptationfunctionforneuroncircuitshasbeenusedforalongtime,butourproposedadaptivecomponentdesignisnewforthebiphasicintegrate-and-reneuroncircuit.AlthoughotherkindsofadaptivefunctioncircuitdesignshavebeenimplementedforIFneuroncircuits,noneofthemfocusonthresholdvoltageauto-adaptation.Inadditiontheiradaptivemechanismcannotsupportreconstructionasrequiredbyourapplication. Theideaofthefullydifferentialadaptiveleakyrefractoryintegrate-and-reneuroncircuit(FD-ALRIF)fortheanalogfront-enddesign,wouldimprovetherecordingquality,andthenewwidedynamicrangedesignoftheadaptivecomponentfortheALRIFencodingcircuit,areexpectedtoreducemoredatabandwidthinthenextgenerationofanalogfront-endrecordingsystems. Overall,theMATLABsimulationoftheneuronmodeldesign,systemlevelsimulationandchipleveldesignshowthatthepulse-basedneuralrecordingsystemcouldbeapossiblecandidatetodetecttheimplantedwirelessin-vivoneuralrecording.Thecoreoftheanalogfront-endhardwarehasbeenintegratedandimplementedintoasingleprintedcircuitboard.TheneuralsimulatortestfortheALRIFneuroncircuitrecordingsystemdemonstratesthatthereconstructedspikescanbespikesorted.Thein-vivoneuralrecordingafrmsthattheUFrecordingsystemiscapableofextractingactionpotentialsfromarat'sbrain. 127

PAGE 128

REFERENCES [1] J.C.Williams,R.L.Rennaker,andD.R.Kipke,Long-termneuralrecordingcharacteristicsofwiremicroelectrodearraysimplantedincerebralcortex,BrainResearchProtocols,vol.4,no.3,pp.303,1999. [2] K.C.Cheung,P.Renaud,H.Tanila,andK.Djupsund,Flexiblepolyimidemicroelectrodearrayforinvivorecordingsandcurrentsourcedensityanalysis,BiosensorsandBioelectronics,vol.22,no.8,pp.1783,2007. [3] E.Patrick,M.Ordonez,N.Alba,J.C.Sanchez,andT.Nishida,Designandfabricationofaexiblesubstratemicroelectrodearrayforbrainmachineinterfaces,inEngineeringinMedicineandBiologySociety,2006.EMBS'06.28thAnnualInternationalConferenceoftheIEEE,2006,pp.2966. [4] S.Suner,MRFellows,C.Vargas-Irwin,GKNakata,andJPDonoghue,Reliabilityofsignalsfromachronicallyimplanted,silicon-basedelectrodearrayinnon-humanprimateprimarymotorcortex,IEEETransactionsonNeuralSystemsandRehabili-tationEngineering,vol.13,no.4,pp.524,2005. [5] D.R.Kipke,R.J.Vetter,J.C.Williams,andJ.F.Hetke,Silicon-substrateintracorticalmicroelectrodearraysforlong-termrecordingofneuronalspikeactivityincerebralcortex,IEEETransactionsonNeuralSystemsandRehabilitationEngineering,vol.11,no.2,pp.151,2003. [6] J.Mavoori,A.Jackson,C.Diorio,andE.Fetz,Anautonomousimplantablecomputerforneuralrecordingandstimulationinunrestrainedprimates,JournalofNeuroscienceMethods,vol.148,no.1,pp.71,2005. [7] KSGuilloryandRANormann,A100-channelsystemforrealtimedetectionandstorageofextracellularspikewaveforms,JournalofNeuroscienceMethods,vol.91,no.1-2,pp.21,1999. [8] J.Wessberg,C.R.Stambaugh,J.D.Kralik,P.D.Beck,M.Laubach,J.K.Chapin,J.Kim,S.J.Biggs,M.A.Srinivasan,andM.A.L.Nicolelis,Real-timepredictionofhandtrajectorybyensemblesofcorticalneuronsinprimates,Nature,vol.408,no.6810,pp.361,2000. [9] J.K.Chapin,K.A.Moxon,R.S.Markowitz,andM.A.L.Nicolelis,Real-timecontrolofarobotarmusingsimultaneouslyrecordedneuronsinthemotorcortex,NatureNeuroscience,vol.2,pp.664,1999. [10] S.Suner,M.R.Fellows,C.Vargas-Irwin,G.K.Nakata,andJ.P.Donoghue,Reliabilityofsignalsfromachronicallyimplanted,silicon-basedelectrodearrayinnon-humanprimateprimarymotorcortex,IEEETransactionsonNeuralSystemsandRehabilitationEngineering,vol.13,no.4,2005. 128

PAGE 129

[11] J.C.Williams,R.L.Rennaker,andD.R.Kipke,Long-termneuralrecordingcharacteristicsofwiremicroelectrodearraysimplantedincerebralcortex,BrainResearchProtocols,vol.4,no.3,pp.303,1999. [12] D.R.Kipke,R.J.Vetter,J.C.Williams,andJ.F.Hetke,Silicon-substrateintracorticalmicroelectrodearraysforlong-termrecordingofneuronalspikeactivityincerebralcortex,IEEETransactionsonNeuralSystemsandRehabilitationEngineering,vol.11,no.2,pp.151,2003. [13] P.Irazoqui-Pastor,I.Mody,andJWJudy,In-vivoEEGrecordingusingawirelessimplantableneuraltransceiver,inNeuralEngineering,2003.ConferenceProceed-ings.FirstInternationalIEEEEMBSConferenceon,2003,pp.622. [14] R.R.Harrison,P.T.Watkins,R.J.Kier,R.O.Lovejoy,D.J.Black,B.Greger,andF.Solzbacher,Alow-powerintegratedcircuitforawireless100-electrodeneuralrecordingsystem,IEEEJournalofSolid-StateCircuits,vol.42,no.1,pp.123,2007. [15] D.Cheney,A.Goh,J.Xu,K.Gugel,J.G.Harris,J.C.Sanchez,andJ.C.Principe,Wireless,invivoneuralrecordingusingacustomintegratedbioamplierandthepicosystem,in3rdInternationalIEEE/EMBSConferenceonNeuralEngineering,2007.CNE'07,2007,pp.19. [16] T.Akin,K.Naja,andRMBradley,Awirelessimplantablemultichanneldigitalneuralrecordingsystemforamicromachinedsieveelectrode,IEEEJournalofSolid-StateCircuits,vol.33,no.1,pp.109,1998. [17] Q.HuangandM.Oberle,A0.5-mWpassivetelemetryICforbiomedicalapplications,IEEEJournalofSolid-StateCircuits,vol.33,no.7,pp.937,1998. [18] P.Mohseni,K.Naja,SJEliades,andX.Wang,WirelessmultichannelbiopotentialrecordingusinganintegratedFMtelemetrycircuit,IEEETransactionsonNeuralSystemsandRehabilitationEngineering,vol.13,no.3,pp.263,2005. [19] R.R.Harrison,P.T.Watkins,R.J.Kier,R.O.Lovejoy,D.J.Black,B.Greger,andF.Solzbacher,Alow-powerintegratedcircuitforawireless100-electrodeneuralrecordingsystem,IEEEJournalofSolidStateCircuits,vol.42,no.1,pp.123,2007. [20] AALazarandLTToth,Timeencodingandperfectrecoveryofbandlimitedsignals,in2003IEEEInternationalConferenceonAcoustics,Speech,andSignalProcessing,2003.Proceedings.(ICASSP'03),2003,vol.6. [21] D.Wei,Time-basedanalog-to-digitalconverters,Ph.D.dissertation,DepartmentofElectricalandComputerEngineering,UniversityofFlorida,Gainesville,FL,2005. 129

PAGE 130

[22] D.Chen,AnUltra-LowPowerNeuralRecordingSystemUsingPulseRepresenta-tion,Ph.D.dissertation,UniversityofFlorida,2006. [23] D.Chen,JGHarris,andJCPrincipe,Abio-amplierwithpulseoutput.,inCon-ferenceproceedings:...AnnualInternationalConferenceoftheIEEEEngineeringinMedicineandBiologySociety.IEEEEngineeringinMedicineandBiologySociety.Conference.ConfProcIEEEEngMedBiolSoc,2004,vol.6,p.4071. [24] C.L.Rogers,Ultra-lowpoweranalogcircuitsforspikefeatureextractionanddectectionfromextracellularneuralrecordings.,Ph.D.dissertation,UniversityofFlorida,2007. [25] Y.Li,AnIntegratedMultichannelNeuralRecordingSystemWithSpikeOutputs,Ph.D.dissertation,UniversityofFlorida,2007. [26] P.MohseniandK.Naja,AfullyintegratedneuralrecordingamplierwithDCinputstabilization,IEEETransactionsonBiomedicalEngineering,vol.51,no.5,pp.832,2004. [27] W.Wattanapanitch,M.Fee,andR.Sarpeshkar,Anenergy-efcientmicropowerneuralrecordingamplier,IEEETransactionsonBiomedicalCircuitsandSystems,vol.1,no.2,pp.136,2007. [28] M.YinandM.Ghovanloo,Alow-noisepreamplierwithadjustablegainandbandwidthforbiopotentialrecordingapplications,inIEEEInternationalSymposiumonCircuitsandSystems,ISCAS2007,pp.321. [29] M.Dagtekin,W.Liu,andR.Bashirullah,Amultichannelchoppermodulatedneuralrecordingsystem,inEngineeringinMedicineandBiologySociety,2001.Proceedingsofthe23rdAnnualInternationalConferenceoftheIEEE.IEEE,2001,vol.1,pp.757. [30] J.Aziz,R.Karakiewicz,R.Genov,AWLChiu,BLBardakjian,M.Derchansky,andPLCarlen,Invitroepilepticseizurepredictionmicrosystem,inIEEEInternationalSymposiumonCircuitsandSystems,ISCAS2007,pp.3115. [31] RRHarrisonandC.Charles,Alow-powerlow-noiseCMOSamplierforneuralrecordingapplications,IEEEJournalofSolid-StateCircuits,vol.38,no.6,pp.958,2003. [32] T.DelbruckandCAMead,Adaptivephotoreceptorwithwidedynamicrange,in1994IEEEInternationalSymposiumonCircuitsandSystems,ISCAS'94.,vol.4. [33] D.WeiandJ.G.Harris,Signalreconstructionfromspikingneuronmodels,inISCAS'04,Proceedingsofthe2004InternationalSymposium,vol.5. [34] RJDufnandACSchaeffer,AclassofnonharmonicFourierseries,TransactionsoftheAmericanMathematicalSociety,pp.341,1952. 130

PAGE 131

[35] H.G.FeichtingerandK.Grochenig,Theoryandpracticeofirregularsampling,Wavelets:MathematicsandApplications,pp.305,1993. [36] J.Harris,T.Nishida,J.C.Principe,E.Patrik,V.Sankar,A.Singh,J.Xu,M.Rastogi,S.F.Yen,B.Mahmoudi,andC.S.Sanchez,Animplantablelow-bandwidthmicroelectroderecordingsystembasedonpulsetimingrepresentationforbrain-machineinterfaces,NeuralEngineering,2011. [37] M.S.Lewickietal.,Areviewofmethodsforspikesorting:thedetectionandclassicationofneuralactionpotentials,Network:ComputationinNeuralSystems,vol.9,no.4,pp.53,1998. [38] F.T.Abu-Nimeh,M.Aghagolzadeh,andK.G.Oweiss,VLSI-friendlyalgorithmforreal-timespikesortinginBrainMachineInterfaceapplications,inIEEEBiomedicalCircuitsandSystemsConference,2008.,pp.265. [39] G.Santhanam,M.Sahani,SIRyu,andKVShenoy,Anextensibleinfrastructureforfullyautomatedspikesortingduringonlineexperiments,inEngineeringinMedicineandBiologySociety,26thAnnualInternationalConferenceoftheIEEE,2004,vol.2. [40] J.D.Victor,Spiketrainmetrics,CurrentOpinioninNeurobiology,vol.15,no.5,pp.585,2005. [41] M.C.W.VanRossum,Anovelspikedistance,NeuralComputation,vol.13,no.4,pp.751,2001. [42] M.MahowaldandR.Douglas,Asiliconneuron,Nature,vol.354,no.6354,pp.515,1991. [43] EMIzhikevich,Simplemodelofspikingneurons,IEEETransactionsonNeuralNetworks,vol.14,no.6,pp.1569,2003. [44] M.J.Chacron,K.Pakdaman,andA.Longtin,Interspikeintervalcorrelations,memory,adaptation,andrefractorinessinaleakyintegrate-and-remodelwiththresholdfatigue,NeuralComputation,vol.15,no.2,pp.253,2003. [45] G.Indiveri,Alow-poweradaptiveintegrate-and-reneuroncircuit,inCircuitsandSystems,Proceedingsofthe2003InternationalSymposium,2003,vol.4. 131

PAGE 132

BIOGRAPHICAL SKETCH Sheng-FengwasborninTaipei,Taiwan,in1977tohislovingparentsMing-Shen YenandShin-MeiHsienYen.Hehasoneolderbrother,Da-JuneYenandoneolder sister,June-LingYen.Sheng-FengreceivedtheB.S.degreeintheDepartmentof ElectricalEngineeringfromNationalTaiwanUniversity,Taipei,Taiwan,in1999.He wasawarded1stplaceintheNationalCompetitionofMicrocomputerApplications inTaiwan,1999andAcademicDean'slistintheCollegeofEngineering,atNational TaiwanUniversityin1995and1996. AfterSheng-Fengnishedtwoyearsofmilitaryservice(1999)Tj/T1_0 11.955 Tf[(2001)hedecidedto pursuefurthereducationintheUnitedStates.Sheng-FengreceivedaM.S.degreein ElectricalEngineeringfromtheUniversityofCaliforniaatLosAngeles(UCLA),in2003. HewasanElectricalEngineerinVIATechnology,Fremont,CA,in2005. Since2007,Sheng-Fenghasbeenateachingandresearchassistantinthe DepartmentofElectricalandComputerEngineeringattheUniversityofFlorida.He joinedtheComputationalNeuroEngineeringLaboratory(CNEL),attheUniversity ofFlorida,whereheworkedwithDr.JohnHarrisontheBrainMachineInterface Project.Hisresearchinterestsarebiologicallyinspiredanalogsignalprocessing andmixed-signalintegratedcircuitdesign.Specially,hisinterestslieindevelopinga low-poweranalogVLSIandmixed-signalVLSIforbiomedicalapplications.Heinterned atTucker-DavisTechnologyinGainesvillein2010.HereceivedhisPh.D.degreeatthe UniversityofFlorida,Gainesville,FLinMay2011. 132