using a common average reference to improve cortical neuron recordings from microelectrode arrays

101:1679-1689, 2009. First published Dec 24, 2008; doi:10.1152/jn.90989.2008 J NeurophysiolDavid J. Anderson and Daryl R. Kipke Kip A. Ludwig, Rachel M. Miriani, Nicholas B. Langhals, Michael D. Joseph,

You might find this additional information useful...

34 articles, 4 of which you can access free at: This article cites http://jn.physiology.org/cgi/content/full/101/3/1679#BIBL

including high-resolution figures, can be found at: Updated information and services http://jn.physiology.org/cgi/content/full/101/3/1679

can be found at: Journal of Neurophysiologyabout Additional material and information http://www.the-aps.org/publications/jn

This information is current as of March 17, 2009 .

http://www.the-aps.org/.American Physiological Society. ISSN: 0022-3077, ESSN: 1522-1598. Visit our website at (monthly) by the American Physiological Society, 9650 Rockville Pike, Bethesda MD 20814-3991. Copyright © 2005 by the

publishes original articles on the function of the nervous system. It is published 12 times a yearJournal of Neurophysiology

on March 17, 2009

jn.physiology.orgD

ownloaded from

http://jn.physiology.org/cgi/content/full/101/3/1679#BIBL

http://jn.physiology.org/cgi/content/full/101/3/1679

http://www.the-aps.org/publications/jn

http://www.the-aps.org/

http://jn.physiology.org

Innovative Methodology

Using a Common Average Reference to Improve Cortical Neuron RecordingsFrom Microelectrode Arrays

Kip A. Ludwig,1 Rachel M. Miriani,1 Nicholas B. Langhals,1 Michael D. Joseph,1 David J. Anderson,2,3

and Daryl R. Kipke1

1Department of Biomedical Engineering, 2Professor Emeritus of Biomedical Engineering, and 3 Professor Emeritus of ElectricalEngineering and Computer Science, University of Michigan, Ann Arbor, Michigan

Submitted 3 September 2008; accepted in final form 15 December 2008

Ludwig KA, Miriani RM, Langhals NB, Joseph MD, AndersonDJ, Kipke DR. Using a common average reference to improvecortical neuron recordings from microelectrode arrays. J Neuro-physiol 101: 1679 –1689, 2009. First published December 24,2008; doi:10.1152/jn.90989.2008. In this study, we propose andevaluate a technique known as common average referencing (CAR) togenerate a more ideal reference electrode for microelectrode recordings.CAR is a computationally simple technique, and therefore amenable toboth on-chip and real-time applications. CAR is commonly used in EEG,where it is necessary to identify small signal sources in very noisyrecordings. To study the efficacy of common average referencing, wecompared CAR to both referencing with a stainless steel bone-screwand a single microelectrode site. Data consisted of in vivo chronicrecordings in anesthetized Sprague-Dawley rats drawn from priorstudies, as well as previously unpublished data. By combining the datafrom multiple studies, we generated and analyzed one of the morecomprehensive chronic neural recording datasets to date. Referencetypes were compared in terms of noise level, signal-to-noise ratio, andnumber of neurons recorded across days. Common average referenc-ing was found to drastically outperform standard types of electricalreferencing, reducing noise by �30%. As a result of the reduced noisefloor, arrays referenced to a CAR yielded almost 60% more discern-ible neural units than traditional methods of electrical referencing.CAR should impart similar benefits to other microelectrode recordingtechnologies—for example, chemical sensing—where similar differ-ential recording concepts apply. In addition, we provide a mathemat-ical justification for CAR using Gauss-Markov theorem and thereforehelp place the application of CAR into a theoretical context.

I N T R O D U C T I O N

Individual unit recordings from cortical neurons are de-pendent on separating the recorded extracellular actionpotential of a neuron from ambient sources of noise. Thesesources of noise can be correlated or uncorrelated across theelectrode array and include motion artifact, 60-Hz noise,instrumentation noise, thermal noise, and biological sourcesof noise (Horsch and Dhillon 2004; Kovacs 1994; Schmidtand Humphrey 1990; Shoham and Nagarajan 2003; Webster1998). As cortical recording experiments move away fromclosed environments designed to reduce noise (e.g., Faradaycages) to more real-world situations (e.g., neuroprostheticdevices), these sources of noise become increasingly prob-lematic (Horsch and Dhillon 2004; Kovacs 1994; Schmidtand Humphrey 1990; Shoham and Nagarajan 2003; Webster1998).

Recording useful signal is dependent on minimizingsources of noise through the use of an appropriate referenceelectrode (Horsch and Dhillon 2004; Kovacs 1994; Schmidtand Humphrey 1990; Shoham and Nagarajan 2003; Webster1998). Typically either an additional microelectrode, or alarge electrode such as a stainless-steel bone screw, orstripped wire, is placed in a location with minimal corticalactivity and used as a reference to subtract out correlatedsources of noise (Blanche et al. 2005; Henze et al. 2000;Ludwig et al. 2006; Nelson et al. 2008; Vetter et al. 2004;Webster 1998; Williams et al. 1999). Both microelectrodeand large electrode references have their own specific ad-vantages and disadvantages.

A large electrode is often preferred as a reference tominimize impedance (Horsch and Dhillon 2004; Kovacs1994; Schmidt and Humphrey 1990; Shoham and Nagarajan2003; Webster 1998). Thermal noise is proportional toimpedance; therefore less thermal noise is “subtracted into”the recordings when using a low-impedance reference elec-trode (Horsch and Dhillon 2004; Kovacs 1994; Schmidt andHumphrey 1990; Shoham and Nagarajan 2003; Webster1998). Unfortunately, there is a significant size and imped-ance mismatch between the reference and the recording siteson the microelectrode array when using a large referenceelectrode (Horsch and Dhillon 2004; Kovacs 1994; Schmidtand Humphrey 1990; Shoham and Nagarajan 2003; Webster1998). Consequently, the representation of correlated sources ofnoise (such as motion artifact and 60-Hz noise) is differentbetween the reference and the microelectrode sites and there-fore is not fully removed by reference subtraction (Horsch andDhillon 2004; Webster 1998).

Because of the sheer size of the large electrode, the referenceis typically placed on top of the dural surface, as opposed toimplantation in cortical tissue (Ludwig et al. 2006; Vetter et al.2004). By placing the reference in a location distal from themicroelectrode sites, the difference between the voltage repre-sentation of correlated sources of noise at the reference and themicroelectrode sites is increased (Horsch and Dhillon 2004;Webster 1998). In addition, there remains the possibility of thereference adding an ECoG (electrocorticogram) signal at thedural surface into the recordings.

As opposed to a large reference electrode, the use of anadditional microelectrode implanted in cortical tissue as a refer-ence presents alternative problems. Although microelectrode ref-

Address for reprint requests and other correspondence: K. A. Ludwig, Univ.of Michigan, 1101 Beal Ave., 2247 LBME, Ann Arbor, MI 48109 (E-mail:[email protected] or [email protected]).

The costs of publication of this article were defrayed in part by the paymentof page charges. The article must therefore be hereby marked “advertisement”in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

J Neurophysiol 101: 1679–1689, 2009.First published December 24, 2008; doi:10.1152/jn.90989.2008.

16790022-3077/09 $8.00 Copyright © 2009 The American Physiological Societywww.jn.org

on March 17, 2009

jn.physiology.orgD

ownloaded from


erences match recording sites on the array in terms of geometryand impedance, they have greater impedance than their largeelectrode counterparts and therefore “subtract in” more thermalnoise to the recordings (Horsch and Dhillon 2004; Kovacs 1994;Ludwig et al. 2006; Schmidt and Humphrey 1990; Shoham andNagarajan 2003). This problem is exacerbated in chronic appli-cations, where the microelectrode reference becomes encapsu-lated in fibrous tissue, further increasing its impedance (Hochberget al. 2006; Ludwig et al. 2006; Nicolelis et al. 2003; Otto et al.2006; Polikov et al. 2005; Rennaker et al. 2007; Seymour andKipke 2007; Spataro et al. 2005; Suner et al. 2005; Szarowski etal. 2003; Turner et al. 1999; Vetter et al. 2004).

When placed on the electrode array itself, a microelectrodereference may actually record neural signal or uncorrelatedbiological noise caused by the activity of distal neural sources;this uncorrelated activity is subtracted into the recordings(Horsch and Dhillon 2004; Webster 1998). If placed in alocation to minimize the probability of recording neural activ-ity (e.g., corpus callosum), the increased physical separation ofthe reference electrode from the microelectrode array decreasesthe correlation between noise at these two locations and there-fore decreases the utility of the reference (Horsch and Dhillon2004; Webster 1998).

In this study, we introduce a technique known as commonaverage referencing (CAR) to generate a more ideal electrodereference for single-unit neural recordings. CAR is commonlyused in EEG, where it is necessary to identify small signalsources in very noisy recordings (Cooper et al. 2003; Offner1950; Osselton 1965). Unlike more complex methods of de-noising recorded signals post hoc (Aminghafari et al. 2006;Bierer and Andersen 1999; Oweiss and Anderson 2001), CARis a computationally simple technique and therefore amenableto both on-chip and real-time applications. As the name im-plies, an average of all the recordings on every electrode site istaken and used as a reference (Cooper et al. 2003; Offner 1950;Osselton 1965). Through the averaging process, only signal/noise that is common to all sites (correlated) remains on theCAR.1 Signal that is isolated on one site (single-unit activity)does not appear on the CAR, unless the signal is so large as todominate the average2 (Cooper et al. 2003; Offner 1950;Osselton 1965). Uncorrelated random noise with a zero meanis minimized through the averaging process.3 Because theCAR provides an accurate representation of correlated noise atthe location of the microelectrode array, but minimizes thecontribution of uncorrelated noise sources, we hypothesize thatcommon average referencing will improve neural recordingquality with respect to both large and microelectrode refer-ences.

To study the efficacy of common average referencing, wecompared CAR to both referencing with a stainless-steel bonescrew and a single microelectrode site, using in vivo chronicrecordings from a prior study (Ludwig et al. 2006), as well aspreviously unpublished data. By combining the data frommultiple studies, we generated and analyzed one of the largestchronic neural recording datasets to date. Reference types werecompared in terms of peak-to-peak noise, signal-to-noise ratio,

and number of units recorded across days. Moreover, weprovide a mathematical justification for common average ref-erencing based on Gauss-Markov theorem and therefore builda theoretical context for future CAR applications.

M E T H O D S

Microelectrodes

Twenty-one male Sprague-Dawley rats were implanted with 2616-channel chronic silicon “Michigan” microelectrode arrays, usingexperimental procedures outlined previously (Ludwig et al. 2006;Vetter et al. 2004). Arrays consisted of four shanks, each with fourevenly spaced iridium electrodes. Site and shank spacings weresufficient (100 �m or greater) to limit the probability of an individualneuron being recorded from multiple sites (Henze et al. 2000). All ofthe electrodes on a specific array were the same site size; electrode sitesizes on an individual array were either 703 or 1,250 �m2.

The data for this study was drawn from previous (Ludwig et al.2006) and ongoing studies aimed at evaluating the efficacy of theconductive polymer poly(3,4-ethylenedioxythiophene) (PEDOT) forimproving neural recording quality. Toward that end, eight of the siteson each array were coated with PEDOT using various depositionmethods and counter-ions, the details of which are beyond the scopeof this study (Ludwig et al. 2006). The remaining eight sites were leftuncoated as controls. Sites were stagger coated to prevent bias causedby cortical depth or location (Ludwig et al. 2006).

Overall, PEDOT sites perform similarly to control sites in recordingneural activity, with one exception. As noted in other studies, sitescoated with PEDOT recorded activity from a slightly larger number ofneurons, primarily as a result of reduced thermal noise (Cui andMartin 2003; Ludwig et al. 2006; Yang et al. 2005). These slightdifferences in recording performance did not affect the results in thispaper (see RESULTS and DISCUSSION for details).

Surgical techniques

All of the arrays in this study were implanted in motor cortex,targeting cortical layer V. Initial anesthesia was administered viaintraperitoneal injections of a mixture of 50 mg/ml ketamine, 5 mg/mlxylazine, and 1 mg/ml acepromazine at an injection volume of 0.125ml/100 g body weight. Updates of 0.1 ml ketamine (50 mg/ml) weredelivered as needed to maintain anesthesia during the surgery. Ani-mals were secured to a standard stereotaxic frame, and three stainless-steel bone screws were inserted into the skull. The electrode connectorwas grounded to a bone screw over parietal cortex using a stainless-steel wire.

A craniotomy �3 � 2 mm was made over the target area (targetlocation: 3.0 mm anterior to bregma, 2.5 mm lateral from bregma, and1.4 mm deep from the surface of the brain). Two incisions were madein the dura mater to create four flaps, which were subsequently foldedback over the edge of the craniotomy. The electrodes were handinserted using a microforceps into the approximate target cortical area.Cortical depth was estimated using the known location of the elec-trode sites on the individual shanks in conjunction with the knownlength of the individual shanks. Next, the surface of the brain wascovered with GelFoam (Henry Schein, Miami, FL) for protection. Thesilicon cable connector was covered with either remaining Gelfoam orKwik-Sil silicone polymer (World Precision Instruments). The entireassembly excluding the connector was enclosed using dental acrylic(Co-Oral-Ite, Dental Mfg., Santa Monica, CA). Finally, sutures wereused to close the skin around the acrylic, and triple-antibiotic ointmentwas applied. All procedures complied with the U.S. Department ofAgriculture guidelines for the care and use of laboratory animals andwere approved by the University of Michigan Animal Care and UseCommittee.

1 For example, 60-Hz noise and motion artifact.2 Given n electrodes, this occurs if the signal is n times larger than the

peak-to-peak value of the noise floor.3 For example, thermal noise and distal neural sources.


1680 LUDWIG, MIRIANI, LANGHALS, JOSEPH, ANDERSON, AND KIPKE

J Neurophysiol • VOL 101 • MARCH 2009 • www.jn.org

on March 17, 2009

jn.physiology.orgD

ownloaded from


Neural recordings and data analysis

For eight of the animals in this study, recorded neural signals wereacquired using a Plexon Multi-channel Neural Acquisition Processor(MNAP; Plexon, Dallas, TX). For the remaining animals, signals wereacquired using a TDT multi-channel acquisition system (Tucker-Davis Technologies, Gainesville, FL). Neural electrophysiologicalrecordings for all 16 channels were amplified and band-pass filtered;single- and multi-unit recordings were sampled at either 40 kHz(Plexon) or 24,414 Hz (TDT), and band-pass filtered from 450 to5,000 Hz. During recording sessions, animals were placed in anelectrically shielded recording booth, and multiple 30-s segments ofcontinuous neural recordings were taken.

Neural recording segments were analyzed off-line using customautomated MatLab (Mathworks) software, as described in detailelsewhere (Ludwig et al. 2006). In summary, an amplitude thresholdwindow was set 3.5 SD above and below the mean of the sampledistribution. For each peak exceeding the threshold window, a 2.4-mscandidate waveform snippet centered on the absolute minimum of thewaveform was removed from the recorded segment and stored. Theamplitude of the noise voltage for every recording site in eachrecorded segment was calculated after all candidate waveforms hadbeen removed.

After initial principal component analysis, individual clusters wereidentified using Fuzzy C-Means clustering (Bezdek 1981; Dunn 1974;Ludwig et al. 2006). When compared with hard clustering, fuzzyclustering reduces classification errors resulting from the synchronousfiring of multiple neurons (Zouridakis and Tam 2000). To determinethe optimum number of clusters, the number of clusters was iterativelyincreased until the value for the objective function calculated for k �1 number of clusters was �55% of the value for the objective functioncalculated for k number of clusters (Karkkainen and Franti 2002).

After clustering, waveforms with a cluster membership index of�0.8 were used to determine a mean waveform for a cluster. Contri-butions of white noise and waveforms created by the simultaneousfiring of multiple neurons generally do not have a membership indexof �0.8 for a particular cluster and therefore were limited using thisprocedure (Zouridakis and Tam 2000). An interspike interval histo-gram for each cluster was generated and visually inspected for anobvious absolute refractory period as an additional measure of noiserejection. Signal amplitude for a cluster was defined as the peak-to-peak amplitude of the mean waveform for each cluster.

The signal-to-noise ratio (SNR) for a given cluster was defined asfollows

SNR � signal amplitude/peak-to-peak amplitude of the noise floor

The peak-to-peak amplitude of the noise on a given site wascalculated as six times the SD of the recording after thresholdedwaveforms were removed, spanning �99.7% of normally distributednoise data (Blanche et al. 2005). By using this method, the calculatedSNR and peak-to-peak noise amplitude on a given site was moreconsistent with a visual inspection of the recorded voltage traces (seeFigs. 2 and 3). For example, an SNR of 2 would indicate that the meanpeak-to-peak amplitude of the signal was twice as large as thepeak-to-peak amplitude of the noise floor. Because the peak-to-peakamplitude of the noise floor for neural recordings is typically between6 and 10 times larger than the root mean square (RMS) value of thenoise floor (Blanche et al. 2005), SNR calculations for neural record-ings based on the RMS of the noise floor raise the SNR with respectto peak-to-peak values.

Clusters with a mean SNR of 1.1 or greater were considereddiscriminable units, because the signal amplitude of these clusters wassufficient to be reliably differentiated from the noise floor. Con-versely, clusters generated by random outlying perturbations fromsources of noise had mean SNR values of 0.9 or less. Althoughnormally distributed noise sources will occasionally exceed the 3.5SD threshold by random chance, the average waveform generated by

these noise sources returns to zero after crossing threshold (instead ofexhibiting an immediate opposing peak). Consequently, the meanwaveform of a noise cluster spans �6 SD of the noise floor, resultingin a calculated SNR of �1. When adjusted for the difference betweencalculating SNR using peak-to-peak amplitude of the noise floorinstead of RMS, an SNR of 1.1 or greater corresponded well withobservations of “moderate or better” unit quality based on SNR valuesfrom similar recording studies (Henze et al. 2000; Ludwig et al. 2006;Suner et al. 2005).

Isolating action potentials from an individual neuron using anindividual recording site is inherently prone to classification errors(Harris et al. 2000; Lewicki 1998). The methodology used in thisstudy was intended to minimize these errors and should accuratelyparallel the true number of underlying neural sources (Ludwig et al.2006). The sorting routine produces similar results to manual sortingperformed by experienced researchers over the same datasets, but withthe advantage of being objective and automated (Ludwig et al. 2006).

Referencing techniques

As noted previously, all recordings in this study were initiallyreferenced to a stainless-steel bone screw (screw) located over parietalcortex. Both the microelectrode reference and the CAR were imple-mented digitally after this initial reference subtraction.

CAR. The most intuitive implementation of a CAR would be to refer-ence a specific site to the sample by sample average of all of theremaining sites on the array. Unfortunately, this approach presents twoproblems with respect to real-time cortical recordings. First, each of the16 sites would have a unique reference, instead of a global referenceshared by all sites. This presents an additional layer of complication intranslating a CAR from a digital reference to an on-chip analog reference.Second, individual sites on a microelectrode array occasionally fail tofunction properly. Consequently, these bad sites must be identified andremoved from the dataset before generating a CAR.

To address these problems, the common average reference for eacharray was generated by taking the sample by sample average of all“good” recording sites, creating one global reference for all sites(CAR-16). For a site to be considered good, the RMS of the noisefloor on the site was required to be between 0.3 and 2 times theaverage RMS of the noise floor across all 16 sites on the array. Sitesidentified as “bad” through impedance spectroscopy (Ludwig et al.2006) typically exhibited noise floors with an RMS value 3–6 timeslarger than the average RMS of the noise floor on good sites. Thissimple methodology for eliminating bad channels proved sufficientfor removing the occasional bad site from further analysis in anautomated fashion.

Because each good recording site contributes to the average calculatedfor the global CAR, the amplitude of all samples on a good site is slightlydecreased when referenced to the CAR. More specifically, given n goodsites, the amplitude of each sample will be (n � 1)/n times the originalvalue (Cooper et al. 2003; Offner 1950; Osselton 1965). As n becomeslarger, this scale factor approaches a value of 1. Because this scale factoraffects both signal and noise equally, it does not affect the calculated SNRon each site.4 For comparison with other types of referencing, all CARpeak-to-peak noise values denoted in this study have been appropriatelyadjusted to compensate for this scale factor.

To assess the contribution of recording sites electrochemicallydeposited with a conductive polymer to the overall data, two addi-tional common average references were created and applied to all datasets. One CAR was generated from only the conductive polymer siteson a given array, and the second CAR from only the control iridiumsites. The CAR generated from conductive polymer sites was onlyapplied as a reference to the conductive polymer sites; the CAR

4 Although SNR is not affected by this scale factor, CAR improves SNR byremoving noise common to all sites without adding uncorrelated noise.


1681COMMON AVERAGE REFERENCING TO IMPROVE NEURAL RECORDINGS


on March 17, 2009

jn.physiology.orgD

ownloaded from


generated from control sites was only applied as a reference to thecontrol sites. In this paper, data obtained using this secondary methodare referred to as CAR-8.

SINGLE-BEST MICROELECTRODE REFERENCE. Cortical recordings areoften taken in reference to an individual microelectrode site; conse-quently, we implemented an algorithm to identify the single-best micro-electrode reference on an array to compare with common averagereferencing. Because a single microelectrode may pick up single-unitactivity, we first used our automated sorting algorithm in conjunctionwith CAR to identify candidate sites with no discernible unit activity.5

Each of the candidate sites were used as a reference for the entire arrayon the original data set. The candidate microelectrode reference thatcreated the lowest average noise floor across the array on a given day wasselected as the single-best microelectrode reference. By identifying thesingle-best microelectrode reference for each array on a daily basis, wecreated a conservative comparison for CAR.

Theoretical justification of CAR usingGauss-Markov theorem

Multi-channel neural recordings can be described as an observed noisysignal, Y, which is composed of an unknown true signal, X, and anunknown noise, N. This general situation can be described by thefollowing equation (Albert 1972; Stark and Woods 2002)

Y � AX � N (1)

where Y is the n � 1 vector of observed neural recordings across nchannels, X is a k � 1 vector representing the true signal (with krepresenting the number of underlying neural sources), A is an n � kmatrix (n � k) that maps X onto Y, and N is an n � 1 normallydistributed random noise vector with zero mean and covariance matrix K.

One useful approach for obtaining a good estimate, X̂, of X from theobserved values of Y is to restrict X̂ to be a linear function of Y

X̂ � BY (2)

Gauss-Markov theorem states that, in a linear model in which theerrors have expectation zero and equal variances, a best linear unbi-ased estimate of the coefficients is given by a least squares estimator(Albert 1972; Stark and Woods 2002). In this formulation, we want todetermine the matrix B, a spatial filter that provides the best linearunbiased estimate (BLUE) of X based on Y.

The filter, Bi, for each channel, i, can be calculated as follows(Albert 1972; Stark and Woods 2002)

Bi � �AiTK�1Ai�

�1AiTK�1 (3)

There are two simplifications that align this general solution with thisstudy. 1) Site spacing is sufficiently large to ensure that neurons areexpressed on only one site. Consequently, Ai for each channel has theform [1, 0, 0, . . . , 0]. 2) The noise model for each channel is identical,producing a covariance matrix K with equal diagonal values (1) andequal off-diagonal components (c).

Consequently, B for each channel has the form

Bi � 1, �, �, . . ., � (4)

where � � �c/[1 � (n � 2)c].As n increases, � approaches �1/(n � 1), leaving Bi as

Bi � 1, � 1/�n � 1�, � 1/�n � 1�, . . ., � 1/�n � 1� (5)

Placing this result into Eq. 2, obtaining the best linear unbiasedestimate of X is equivalent to subtracting the average of all other

channels from the current channel—in other words, use a CAR. Thisformulation is generally true even if the diagonal and off-diagonalvalues vary across channels. Note that when c is large (near 1), capproaches �1/(n � 1) more quickly. This result is intuitive; whenthere is high-amplitude correlated noise across channels, referencingis particularly effective. When there is very little or no correlatednoise across channels, using a reference only subtracts in uncorrelatednoise from the reference to the other channels. Although commonaverage referencing minimizes uncorrelated noise as a function of n,it does not remove it entirely. Consequently, referencing becomescounterproductive when there is no correlated noise between thereference and recordings sites.

Statistical analysis

Because the four types of referencing in this study were performeddigitally on the same data sets, these data sets are highly correlated.To accommodate the correlations between comparison groups, arepeated-measures ANOVA was performed; a Bonferroni correctionwas applied to the tests for each group pair to control for experiment-wise error rate. The Bonferroni correction is regarded as a verystrict/conservative test of significance when comparing multiplegroups. At an � value of 0.01, the Bonferroni correction required P �0.0017 for significance. Equality of variance was verified by theLevene statistic.

R E S U L T S

Noise

Over the course of this study, the average peak-to-peak noisefloor across all sites using CAR was 32.5 �V (Fig. 1), asignificant improvement over referencing to either a stainlesssteel ground screw (45.9 �V; P � 10�15) or to the single-bestmicroelectrode site on the array (40.7 �V; P � 10�15). Morespecifically, every single recording site in the study (n �5,010) exhibited a lower noise floor when referenced to theCAR in comparison to the single-best microelectrode refer-ence. In all but three instances, the noise floor using CAR waslower than referencing to a ground screw.

Using either CAR or single-best references, the trend in theaverage noise floor across sites following surgery paralleled theresults of other microelectrode recording studies. The noisefloor typically decreased over the days immediately followingsurgery and gradually increased over time. An increase in noiseover time is expected, as the impedances of the electrodes areknown to increase over time due to fibrous encapsulation,increasing the thermal noise on a site (Ludwig et al. 2006;Vetter et al. 2004; Williams et al. 1999).

In contrast to CAR or single-best references, the noisefloor using the ground screw as a sole reference was con-siderably more variable. This variability can be explained byseparating the recordings into two distinct cases. In the firstcase, the recordings made using the ground screw as areference were contaminated by one or more large sources ofcorrelated noise (Fig. 2). Both the CAR and single-best referenceminimize the contribution of correlated noise sources and there-fore exhibit lower noise floors and more distinct unit activitythan ground screw referencing when large sources of correlatednoise are evident. Because the ground screw is a poor matchfor the implanted microelectrode in terms of location, imped-ance, and geometry, it does not lower the contribution ofcorrelated noise sources as effectively. The noise floor onelectrode sites was larger when using the single-best electrode

5 Using a common average reference was necessary to identify sites withoutunit activity, because the noise floor using the ground screw as a sole referencewas sufficient to mask unit activity on many days.




on March 17, 2009

jn.physiology.orgD

ownloaded from


reference as opposed to the CAR, because the single-bestelectrode subtracts in the uncorrelated noise from the singleelectrode site. Conversely, uncorrelated sources of noise with azero mean tend toward zero when averaged over all the sites on anarray using CAR.

In the second case, the recordings were only minimallycontaminated by correlated noise (Fig. 3). In this case, theground screw reference actually lowered the noise floor withrespect to the single-best microelectrode reference. Groundscrews used in this study had impedances that were multipleorders of magnitude lower than implanted microelectrodes andconsequently subtracted less thermal noise into the record-ings.6 Because only recordings referenced to the ground screwwere significantly altered by the absence or presence of corre-lated noise, recordings referenced to the ground screw wereconsiderably more variable.7 The CAR, which minimized bothcorrelated and uncorrelated sources of noise, dramaticallyoutperformed the single-best and ground screw references inboth cases.

Signal-to-noise and unit activity

Calculating a meaningful average SNR of unit recordings forcomparison between reference types was complicated by thefact that sites registered an overall greater number of discern-ible units with CAR. More specifically, the reduction in noiseon CAR sites was sufficient to separate previously unapparentsignals from the noise floor—instead of just increasing theSNR of already discriminable units. Therefore comparing theaverage SNR of all discriminable units was not representativeof the true difference in SNR between reference types.

To calculate the true difference in SNR between CAR andother types of referencing, we first used CAR in conjunctionwith our automated sorting algorithm to identify the location intime of all unit activity. The known location in time of unitactivity was used to calculate the mean signal amplitude for allreference types. Through this methodology, we were able toidentify the location of signal that was previously obscured bythe noise floor when not using CAR.

Over the course of this study, sites referenced to CARexhibited a higher SNR than sites referenced to either a groundscrew or single-best microelectrode reference (P � 10�10;Fig. 4). The average SNR on CAR sites was 1.59, in compar-ison to 1.31 for sites referenced to the single-best microelec-trode reference and 1.24 for sites referenced to a ground screw.Because CAR does not alter the relative size of signal on agiven site, this improvement in performance is attributable tothe decreased CAR noise floor.

6 Many students in our laboratory were surprised to note that digitallyreferencing an individual electrode site would often increase the noise flooracross the array and initially attributed this result to an equipment malfunction.

7 Within array variances when using all types of referencing in this studywere equivalent, as verified by the Levene statistic. Ground screw peak-to-peak noise values varied greatly between arrays, depending on the presence ofcorrelated noise sources. As noted in METHODS, both the single-best referenceand CAR were implemented after initial referencing to a ground screw.

20

25

30

35

40

45

50

55

60

Days from Surgery0−

1 (n=

406)

2−3 (

n=35

8)

4−5 (

n=34

4)

6−7 (

n=34

3)

8−9 (

n=32

8)

10−1

1 (n=

342)

12−1

3 (n=

357)

14−1

5 (n=

342)

16−1

7 (n=

326)

18−2

1 (n=

358)

22−2

5 (n=

340)

26−2

9 (n=

314)

30−3

3 (n=

285)

34−3

7 (n=

299)

38−4

1 (n=

268)

Peak

−to−

Peak

Noi

se (µ

V)

ScrewSingle−BestCAR−8CAR−16

FIG. 1. Noise across days. Bars denote SE of the dataset on a given block of days. Number of sites on a given block of days, n, is listed on the x-axis. Overthe course of the study, sites referenced to a common average reference exhibited 30% less noise than standard methods of referencing (P � 10�15). Sitesreferenced to a ground screw placed over parietal cortex exhibited increased noise floor variability in comparison to referencing with common average referencing(CAR) or the single-best microelectrode site. Noise floor amplitude calculated using CAR or single-best microelectrode site references decreased immediatelyfollowing surgery and increased afterward, consistent with the trend in noise level found in prior studies (Ludwig et al. 2006; Schwartz 2004; Vetter et al. 2004.;Williams et al. 1999).




on March 17, 2009

jn.physiology.orgD

ownloaded from


Referencing with CAR both 1) increased the SNR of unitsalready evident when referencing to the single-best microelec-trode reference or the ground screw and 2) decreased the noisefloor sufficiently to show units that were previously obscuredwhen referencing to the single-best microelectrode or theground screw. Consequently, there was a greater number ofdiscernible units when referencing to CAR in comparison to a

ground screw or single-best microelectrode reference (P �10�10; Fig. 5). Over the course of the study, an average of46.5% of CAR sites exhibited discriminable unit activity, incomparison to 29.6% of the single-best sites and 27.46% of theground screw sites. CAR sites recorded an average of 0.842units per day, whereas single-best sites recorded an average of0.447 units per day and ground screw sites recorded an average

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100 Common Average Reference0 0.5 1 1.5 2

-50

-25

0

25

50

CAR Site #1SNR 2.45

Pk-Pk Noise 25.2

0 0.5 1 1.5 2 -50

-25

0

25

50

Screw Site #1SNR 1.07

Pk-Pk Noise 57.7

0 0.5 1 1.5 2 -50

-25

0

25

50

Single Best Site #1SNR 1.79

Pk-Pk Noise 34.5

0 0.5 1 1.5 2 -50

-25

0

25

50

CAR Site #2SNR 3.65

Pk-Pk Noise 27.7

0 0.5 1 1.5 2 -50

-25

0

25

50


Pk-Pk Noise 57.8

0 0.5 1 1.5 2 -50

-25

0

25

50


Pk-Pk Noise 35.7

0 0.5 1 1.5 2 -50

-25

0

25

50

CAR Site #3SNR 1.77

Pk-Pk Noise 24.4

0 0.5 1 1.5 2 -50

-25

0

25

50


Pk-Pk Noise 57.3

0 0.5 1 1.5 2 -50

-25

0

25

50


Pk-Pk Noise 32.9

Volt

age

(µV)

0 0.5 1 1.5 2 -50

-25

0

25

50 Common Average Reference

SecondsFIG. 2. Representative example of recordings contaminated by correlated noise. All columns depict the same 2 s of high-speed recordings taken

simultaneously across 3 sites on the same array. Column 1 data are referenced to a CAR. Column 2 data are referenced to a stainless-steel screw placed aboveparietal cortex. Column 3 data are referenced to the single-best microelectrode reference on the array. Sites referenced to CAR exhibit a lower noise floor andhigher signal-to-noise ratio than standard electrical references. When sites are referenced to the ground screw, an intermittent correlated noise source is evidentthat does not appear when using either the CAR or single-best microelectrode reference. This noise source is sufficient to completely obscure unit activity. Alsonote that, despite the presence of large action potentials on site 2, traces of this signal are not evident on sites 1 and 3 when referenced to CAR (see sites 1 and3 screw and single-best reference over the same data set for comparison).




on March 17, 2009

jn.physiology.orgD

ownloaded from


of 0.415 units per day. These daily numbers correspond to4,045 total units registered on CAR sites during this study, incomparison to 2,152 units registered on single-best sites and1,998 units registered on ground screw sites. As with theimprovement in SNR, this improvement is attributable to themarkedly decreased CAR noise floor. Consistent with previ-

ous studies (Ludwig et al. 2006; Santhanam et al. 2007;Schwartz 2004; Vetter et al. 2004; Williams et al. 1999),there was a notable decrease in unit activity in the daysimmediately following surgery, followed by an increase inunit activity after day 15. Although the exact reason for thisincrease in unit activity at the 2- to 3-wk point is unknown,

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100

CAR Site #1SNR 1.65

Pk-Pk Noise 34.6

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100


Pk-Pk Noise 38.9

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100


Pk-Pk Noise 43.2

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100

CAR Site #2SNR 1.99

Pk-Pk Noise 36.7

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100


Pk-Pk Noise 41.3

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100


Pk-Pk Noise 45.1

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100

CAR Site #3SNR 3.58

Pk-Pk Noise 34.7

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100


Pk-Pk Noise 39.5

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100


Pk-Pk Noise 43.7

Seconds

Volt

age

(µV)

0 0.5 1 1.5 2 -100

-75

-50

-25

0

25

50

75

100 Common Average Reference

FIG. 3. Example of recordings with low correlated noise. All columns depict the same 2 s of high-speed recordings taken simultaneously across 3 sites onthe same array. Column 1 data are referenced to a CAR. Column 2 data are referenced to a stainless-steel screw placed above parietal cortex. Column 3 dataare referenced to the single-best microelectrode reference on the array. Sites referenced to CAR exhibit a lower noise floor and higher signal-to-noise ratio thanstandard electrical references. In cases where large sources of correlated noise are not evident, the ground screw reference routinely outperformed the single-bestreference in terms of noise level, signal-to-noise, and number of discriminable units. In these cases, common average referencing still outperformed referencingto either a ground screw or the single-best microelectrode site on the array. Arrows denote a neural signal evident on site 2 when referenced to either CAR orsingle-best microelectrode reference. Note that the waveform has been distorted on the voltage axis, presumably a result of the increased noise floor when usingthe single-best microelectrode reference. Even if a neural unit is discernible from the noise, an increased noise floor means more waveform variability, limitingthe efficacy of common sorting algorithms (Lewicki 1998).




on March 17, 2009

jn.physiology.orgD

ownloaded from


this result has been previously reported in other studies(Ludwig et al. 2006; Santhanam et al. 2007; Schwartz2004). At 2–3 wk after implantation, the fibrous encapsula-tion around an implant becomes more defined, and theimmune response begins a transition into the chronic phase(Turner et al. 1999).

D I S C U S S I O N

Conductive polymer CAR

As noted in the METHODS, two additional CARs were createdand applied to all data sets to assess the contribution ofrecording sites electrochemically deposited with a conductivepolymer to the overall data. One CAR was generated from onlythe conductive polymer sites on a given array, and the secondCAR from only the control sites (CAR-8). The CAR generatedfrom conductive polymer sites was only applied as a referenceto the conductive polymer sites; the CAR generated fromcontrol sites was only applied as a reference to the control sites.

If the noise between conductive polymer sites was morecorrelated than the noise between conductive polymer andcontrol sites, one would expect the noise floor to decrease bygenerating separate CARs for conductive polymer and controlsites. Instead, creating separate CARs marginally increased thenoise floor (Fig. 1). This increase in noise is a result of usingonly 8 sites to calculate the CAR, instead of 16. As shownpreviously, using CAR more closely approximates the bestlinear unbiased estimate of the signal as the number of chan-nels increases. The average of uncorrelated sources of noisewith a zero mean tends toward zero; the contribution ofuncorrelated sources of noise becomes closer to zero when

more sites are included to generate the average. Consequently,decreasing the number of sites used to calculate the CARincreases the contribution of uncorrelated noise sources.

Although the correlation in noise between pairs of sitesacross a single array was similar on a given day, this correla-tion value could vary dramatically from day to day. Theaverage cross-channel correlation on an array was 0.66 overthe course of this study, but this average value ranged from 0.2to 0.95 depending on the day. These values are similar toresults reported by Rebrik et al. (1999), who found thatcross-channel correlation coefficients within the same tetrodeimplanted in cat visual cortex ranged between 0.8 and 0.92,whereas the correlation coefficients across channels of differ-ent tetrodes ranged between 0.47 and 0.51.

Application of CAR to tetrodes

Under certain circumstances, there is a possibility of regis-tering neural signal on a single site so large that it dominatesthe average of all sites, and therefore appears on the CAR(Cooper et al. 2003; Offner 1950; Osselton 1965). Accord-ingly, this neural signal would appear as a small false unit onall sites referenced to the CAR. For this to occur, given nrecording sites, the average amplitude of the signal on a givensite must be n times larger than the peak-to-peak amplitude ofthe noise floor (Cooper et al. 2003; Offner 1950; Osselton1965). In this study, the signal amplitude on one site was neversufficient to dominate the average (Fig. 4); given 16 recordingssites on every array, the signal on a single site would have toregularly exceed 16 times the peak-to-peak amplitude of thenoise floor (Figs. 2 and 3). Based on the largest peak-to-peak

0.8

1

1.2

1.4

1.6

1.8

2

2.2

Days from Surgery

Sign

al to

Noi

se

0−1 (

n=32

9)

2−3 (

n=18

5)

4−5 (

n=15

8)

6−7 (

n=12

7)

8−9 (

n=70

)

10−1

1 (n=

78)

14−1

5 (n=

82)

16−1

7 (n=

62)

18−2

1 (n=

104)

22−2

5 (n=

112)

26−2

9 (n=

223)

30−3

3 (n=

243)

34−3

7 (n=

253)

38−4

1 (n=

186)

12−1

3 (n=

89)

CAR−16CAR−8Single−BestScrew

FIG. 4. Signal to noise across days. Bars denote SE of the data set on a given block of days. Number of units recorded on a given block of days, n, is listedon the x-axis. Over the course of the study, sites referenced to CAR exhibited a signal-to-noise ratio of 1.59, a significant improvement over referencing to eitherthe single-best microelectrode site (1.31, P � 10�10) or ground screw (1.24, P � 10�10). Variability in signal-to-noise ratio across all types of reference increasedtoward the end of the study, concurrent with an increase in number of units recorded across all arrays. An increase in number of units recorded starting at 3 wkafter implantation has been noted in prior studies (Ludwig et al. 2006; Santhanam et al. 2007; Schwartz 2004).




on March 17, 2009

jn.physiology.orgD

ownloaded from


amplitudes for action potentials observed in this study incomparison to the peak-to-peak amplitude of the noise floor, aminimum of five electrodes must be used to calculate the CARto limit the possibility of an action potential dominating thecommon average reference. As sites were separated by �100�m in this study, the probability of recording signal from anindividual neuron on multiple sites simultaneously was low(Henze et al. 2000).

For some neural recording applications, microelectrodes aremanufactured with groups of four recording sites closelyspaced to deliberately detect neural activity from an individualneuron simultaneously, known commonly as a tetrode config-uration (Gray et al. 1995). As a result, the probability of signalfrom an individual neuron dominating the CAR average be-comes much greater. If m sites of n total electrodes recordsimilar signal from an individual neuron, the average signalacross m electrodes needs only to exceed n/m times the peak-to-peak amplitude of the noise floor to dominate the CAR aver-age. One solution to this problem would be to increase n, thenumber of electrode sites used to calculate the CAR. Inaddition, a site or sites recording aberrantly large signal couldsimply be removed when calculating the common averagereference (Cooper et al. 2003; Offner 1950; Osselton 1965).

When using a CAR, the potential contribution from a neuralsource on one electrode in a tetrode could diminish the measuredpotential from the neural source on the three remaining electrodesin the tetrode. This distortion should be constant across all fourelectrodes and therefore should not adversely affect standardamplitude-based tetrode clustering techniques (Gray et al. 1995).

The distortion could be problematic, however, when using math-ematical techniques to locate the neural source in space.

Comparing CAR to alternative correlation-baseddenoising methods

A number of post hoc methods have been previously sug-gested that identify and use the correlated noise across chan-nels to reduce overall noise and improve spike sorting (Bierer1999; Musial et al. 2002; Rebrik et al. 1999). These post hocanalysis techniques require initial segments of data to definethe correlation in noise between multiple channels. Conse-quently, these strategies are computationally intensive in com-parison to CAR and are not easily performed in real time.Moreover, these methods require additional recalibration stepsas the correlation in noise across the array changes over time.As a result, these techniques are not easily implementedon-chip. CAR is not a post hoc analysis technique but insteadan alternative electrical reference, which can be implementedin real time without the need for calibrations to account fortransient changes in noise content. In addition, CAR can beimplemented with a simple circuit either on-chip or at therecording headstage of the microelectrode array, thereby min-imizing the contribution of noise before subsequent amplifica-tion and digitization.

To implement CAR on-chip, it is necessary to ensure that theamplitude of the incoming signal is within the dynamic range ofthe operational amplifier (op-amp). In some cases, noise tran-sients are large enough to exceed the dynamic range of theop-amps due to the large amplification that is traditionally

0

10

20

30

40

50

60

70

80

90

100

Days from Surgery

0−1 (

n=26

)

2−3 (

n=23

)

4−5 (

n=22

)

6−7 (

n=22

)

8−9 (

n=21

)

10−1

1 (n=

22)

12−1

3 (n=

23)

14−1

5 (n=

22)

16−1

7 (n=

21)

18−2

1 (n=

23)

22−2

5 (n=

22)

26−2

9 (n=

20)

30−3

3 (n=

18)

34−3

7 (n=

19)

38−4

1 (n=

17)

Perc

enta

ge o

f Site

s w

ith D

isce

rnib

le U

nits

CAR−16CAR−8Single−BestScrew

FIG. 5. Percentage of sites with units across days. Bars denote SE of the data set on a given block of days. Number of arrays included in analysis, n, is listedon the x-axis. Over the course of the study, sites referenced to a common average reference yielded almost 60% more discernible units than when reference tostandard electrical references. This increase in performance is attributable to a reduced noise floor, enhancing signal-to-noise ratio, and therefore increasing thenumber of discernible units. Unit recordings were strong initially, dipped dramatically in the days following surgeries, and returned to initial levels after the 4-wkpoint. This trend in recording performance has been noted in previous recording studies (Ludwig et al. 2006; Santhanam et al. 2007; Schwartz 2004).




on March 17, 2009

jn.physiology.orgD

ownloaded from


required for front-stage neural acquisition hardware, causing“saturation” or “clipping” of the output signal. To avoid thisproblem, a large electrode such as a ground screw is normallyused as an electrical reference to minimize incoming noisetransients. To successfully implement real-time hardwareCAR, it would be necessary to either 1) increase the dynamicrange of the op-amps to exceed maximum values of expectednoise transients given the gains necessary for signal reconstruc-tion or 2) first use a ground screw as an electrical reference toprevent amplifier saturation.

Additional benefits of reducing noise

In some cases in this study, CAR provided a significantimprovement in recording performance, in other cases CARsalvaged recordings that would have otherwise been unusable.Decreasing the noise floor not only increases the total numberof units recorded across an array but also decreases the vari-ability of those units. Additional noise can alter the instanta-neous shape of an action potential, potentially distorting therecorded waveform in principal component analysis (PCA)space, deleteriously affecting common sorting techniques(Lewicki 1998) (Fig. 3).

Perhaps even more significant, CAR mitigates the contribu-tion of intermittent sources of correlated noise (Fig. 2), such asmotion artifact. These noise sources distort the relationshipbetween neural firing rates and underlying physiological pro-cesses. This is especially problematic in real-time applicationssuch as brain–machine interfaces, where intermittent artifactcannot be removed through post hoc analysis. Reducing vari-ability in the noise floor can also be critical for comparisonstudies of electrode performance. Significant differences inelectrode performance as a function of electrode design or adrug treatment could potentially be masked by the large vari-ance in noise floor from day to day.

Conclusions

According to standard differential recording theory, thereference electrode and the recording electrode must bematched as closely as possible in terms of geometry, electrodematerial, location, and electrical characteristics (Kovacs 1994;Schmidt and Humphrey 1990; Shoham and Nagarajan 2003;Webster 1998). These requirements would seem to suggest theuse of a single microelectrode as a reference, instead of anunmatched larger electrode. Unfortunately, because of the smallsize of a microelectrode, coupled with fibrous encapsulation en-demic to long-term cortical implants (Edell et al. 1992; Johnsonet al. 2005; Ludwig et al. 2006; Polikov et al. 2005; Schmidt etal. 1993; Szarowski et al. 2003; Turner et al. 1999; Vetter et al.2004; Williams et al. 1999; Xindong et al. 1999), the imped-ance of a microelectrode reference adds significant uncorre-lated thermal noise to cortical recordings (Kovacs 1994;Schmidt and Humphrey 1990; Shoham and Nagarajan 2003;Webster 1998). Moreover, the microelectrode reference couldpotentially register neural signal as well as uncorrelated bio-logical noise—adding both to neural recordings. In contrast, acommon average reference minimizes uncorrelated sources ofsignal and noise through averaging, while eliminating sourcesof noise common to all sites. Therefore a common averagereference more closely approximates the theoretical differentialrecording ideal.

Not surprisingly, CAR was found to drastically outperformstandard types of electrical referencing over the course of thisstudy, reducing noise by �30%. As a result of the reducednoise floor, arrays referenced to a CAR yielded almost 60%more discernible units than traditional methods of electricalreferencing. However, these results need to be considered incontext of the experimental methodology; recordings in thisstudy were taken from anesthetized animals placed in a Fara-day cage designed to reduce ambient noise. Recordings takenfrom awake and behaving animals in nonshielded environ-ments would have much more noise. Consequently, the draw-backs of either single microelectrode or large electrode refer-ences would be further exacerbated. Under these circum-stances, the CAR would likely exhibit an even more dramaticincrease in recording performance. CAR may impart similarbenefits to other microelectrode recording technologies—forexample, chemical sensing (Burmeister and Gerhardt 2001;Dressman et al. 2002; Johnson et al. 2007)—where similardifferential recording concepts apply.

A C K N O W L E D G M E N T S

We thank all of the members of the Neural Engineering Laboratory at theUniversity of Michigan for assistance in this study; S. Richardson-Burns forcoating some of the probes used in this study; and E. Purcell, J. Seymour, M.Gibson, and H. Parikh for invaluable advice and manuscript review.

D. R. Kipke and D. J. Anderson have significant financial interest inNeuroNexus Technologies.

G R A N T S

This work was supported by the National Institutes of Health P41 Center forNeural Communications Technology (EB002030), Biotectix, and the WhitakerFoundation.

R E F E R E N C E S

Albert AE. Regression and the Moore-Penrose Pseudoinverse. New York:Academic Press, 1972.

Aminghafari M, Cheze N, Poggi JM. Multivariate de-noising using waveletsand principal component analysis. Comput Stat Data Analy 50: 2381–2398,2006.

Bezdek JC. Pattern Recognition With Fuzzy Objective Function Algorithms.New York: Plenum Press, 1981.

Bierer SB, Andersen DJ. Multi-channel spike detection and sorting using anarray processing technique. Neurocomputing 26–27: 947–956, 1999.

Blanche TJ, Spacek MA, Hetke JF, Swindale NV. Polytrodes: high-densitysilicon electrode arrays for large-scale multiunit recording. J Neurophysiol93: 2987–3000, 2005.

Burmeister JJ, Gerhardt GA. Self-referencing ceramic-based multisite mi-croelectrodes for the detection and elimination of interferences from themeasurement of L-glutamate and other analytes. Analyt Chem 73: 1037–1042, 2001.

Cooper R, Binnie CD, Osselton JW, Prior PF, Wisman T. EEG, paediatricneurophysiology, special techniques and applications. In: Clinical Neuro-physiology, Vol. 2, edited by Cooper R, Mauguiere F, Osselton JW, PriorPF, and Tedman BM. Amsterdam: Elsevier, 2003, p. 8–103.

Cui X, Martin DC. Electrochemical deposition and characterization ofpoly(3,4-ethylenedioxythiophene) on neural microelectrode arrays. SensorsActuators B 89: 92–102, 2003.

Dressman SF, Peters JL, Michael AC. Carbon fiber microelectrodes withmultiple sensing elements for in vivo voltammetry. J Neurosci Methods 119:75–81, 2002.

Dunn JC, A fuzzy relative of the ISODATA process and its use in detectingcompact well compact well-separated clusters. J Cybern 3: 32–57, 1974.

Edell DJ, Toi VV, McNeil VM, Clark LD. Factors influencing the biocom-patibility of insertable silicon microshafts in cerebral cortex. IEEE TransBiomed Eng 39: 635–643, 1992.

Gray CM, Maldonado PE, Wilson M, McNaughton B. Tetrodes markedlyimprove the reliability and yield of multiple single-unit isolation from




on March 17, 2009

jn.physiology.orgD

ownloaded from


multi-unit recordings in cat striate cortex. J Neurosci Methods 63: 43–54,1995.

Harris KD, Henze DA, Csicsvari J, Hirase H, Buzsaki G. Accuracy oftetrode spike separation as determined by simultaneous intracellular andextracellular measurements. J Neurophysiol 84: 401–414, 2000.

Henze DA, Borhegyi Z, Csicsvari J, Mamiya A, Harris KD, Buzsaki G.Intracellular features predicted by extracellular recordings in the hippocam-pus in vivo. J Neurophysiol 84: 390–400, 2000.

Hochberg LR, Serruya MD, Friehs GM, Mukand JA, Saleh M, CaplanAH, Branner A, Chen D, Penn RD, Donoghue JP. Neuronal ensemblecontrol of prosthetic devices by a human with tetraplegia. Nature 442: 164,2006.

Horsch KW, Dhillon GS. Neuroprosthetics Theory and Practice. River Edge,NJ: World Scientific, 2004.

Johnson MD, Kao OE, Kipke DR. Spatiotemporal pH dynamics followinginsertion of neural microelectrode arrays. J Neurosci Methods 160: 276–287, 2007.

Johnson MD, Otto KJ, Kipke DR. Repeated voltage biasing improves unitrecordings by reducing resistive tissue impedances. IEEE Trans RehabEngineer 13: 160–165, 2005.

Karkkainen I, Franti P. Dynamic Local Search for Clustering with UnknownNumber of Clusters. International Conference on Pattern Recognition,Quebec, Canada, August 11–15, 2002.

Kovacs GTA. Introduction to the theory, design, and modeling of thin-filmmicroelectrodes for neural interfaces. In: Enabling Technologies for Cul-tured Neural Networks, edited by Stenger DA and McKenna T. New York:Academic, 1994, p. 121–165.

Lewicki MS. A review of methods for spike sorting: the detection andclassification of neural action potentials. Network Comput Neural Syst 9:R53–R78, 1998.

Ludwig KA, Uram JD, Yang J, Martin DC, Kipke DR. Chronic neuralrecordings using silicon microelectrode arrays electrochemically depositedwith a poly(3,4-ethylenedioxythiophene) (PEDOT) film. J Neural Engineer3: 59, 2006.

Musial PG, Baker SN, Gerstein GL, King EA, Keating JG. Signal-to-noiseratio improvement in multiple electrode recording. J Neurosci Methods 115:29–43, 2002.

Nelson MJ, Pouget P, Nilsen EA, Patten CD, Schall JD. Review of signaldistortion through metal microelectrode recording circuits and filters. J Neu-rosci Methods 169: 141–157, 2008.

Nicolelis MAL, Dimitrov D, Carmena JM, Crist R, Lehew G, Kralik JD,Wise SP. Chronic, multisite, multielectrode recordings in macaque mon-keys. Proc Natl Acad Sci USA 100: 11041–11046, 2003.

Offner FF. The EEG as potential mapping: the value of the average monopolarreference. Electroencephalogr Clin Neurophysiol 2: 213–214, 1950.

Osselton JW. Acquisition of EEG data by bipolar unipolar and averagereference methods: a theoretical comparison. Electroencephalogr Clin Neu-rophysiol 19: 527–528, 1965.

Otto KJ, Johnson MD, Kipke DR. Voltage pulses change neural interfaceproperties and improve unit recordings with chronically implanted micro-electrodes. IEEE Trans Biomed Eng 53: 333–340, 2006.

Oweiss KG, Anderson DJ. Noise reduction in multichannel neural recordingsusing a new array wavelet denoising algorithm. Neurocomputing 38–40:1687–1693, 2001.

Polikov VS, Tresco PA, Reichert WM. Response of brain tissue to chroni-cally implanted neural electrodes. J Neurosci Methods 148: 1–18, 2005.

Rebrik SP, Wright BD, Emondi AA, Miller KD. Cross-channel correlationsin tetrode recordings: implications for spike-sorting. Neurocomputing 26–27: 1033–1038, 1999.

Rennaker RL, Miller J, Tang H, Wilson DA. Minocycline increases qualityand longevity of chronic neural recordings. J Neural Eng 4: L1–L5, 2007.

Santhanam G, Linderman MD, Gilja V, Afshar A, Ryu SI, Meng TH,Shenoy KV. HermesB: a continuous neural recording system for freelybehaving primates. IEEE Trans Biomed Eng 54: 2037–2050, 2007.

Schmidt E, Humphrey DR. Extracellular single-unit recording methods. In:Neurophysiological Techniques, edited by Boulton B, Baker B, and Vander-wolf H. Clifton, NJ: Humana Press, 1990, p. 1–64.

Schmidt S, Horch K, Normann R. Biocompatibility of silicon-based elec-trode arrays implanted in feline cortical tissue. J Biomed Mater Res 27:1393–1399, 1993.

Schwartz AB. Cortical neural prosthetics. Annu Rev Neurosci 27: 487–507,2004.

Seymour JP, Kipke DR. Neural probe design for reduced tissue encapsulationin CNS. Biomaterials 28: 3594–3607, 2007.

Shoham S, Nagarajan S. The theory of central nervous system recording. In:Neuroprosthetics: Theory and Practice, edited by Horch KW, Dhillon GS.Singapore: World Scientific Publishing, 2003, p. 448–465.

Spataro L, Dilgen J, Retterer S, Spence AJ, Isaacson M, Turner JN, ShainW. Dexamethasone treatment reduces astroglia responses to inserted neu-roprosthetic devices in rat neocortex. Exp Neurol 194: 289–300, 2005.

Stark H, Woods JW. Probability and Random Processes With Applications toSignal Processing. Upper Saddle River, NJ: Prentice Hall, 2002.

Suner S, Fellows MR, Vargas-Irwin C, Nakata GK, Donoghue JP. Reli-ability of signals from a chronically implanted, silicon-based electrode arrayin non-human primate primary motor cortex. IEEE Trans Rehab Eng 13:524–541, 2005.

Szarowski DH, Andersen MD, Retterer S, Spence AJ, Isaacson M, Craig-head HG, Turner JN, Shain W. Brain responses to micro-machined silicondevices. Brain Res 983: 23–35, 2003.

Turner JN, Shain W, Szarowski DH, Andersen M, Martins S, Isaacson M,Craighead H. Cerebral astrocyte response to micromachined silicon im-plants. Exp Neurol 156: 33–49, 1999.

Vetter RJ, Williams JC, Hetke JF, Nunamaker EA, and Kipke DR. Spikerecording performance of implanted chronic silicon-substrate microelec-trode arrays in cerebral cortex. IEEE Trans Neural Syst Rehab Eng 52:896–904, 2004.

Webster JG. Medical Instrumentation: Application and Design. New York:John Wiley, 1998.

Williams J, Rennaker R, Kipke D. Long-term neural recording characteris-tics of wire microelectrode arrays implanted in cerebral cortex. Brain ResBrain Res Protocol 4: 303–313, 1999.

Xindong L, McCreery DB, Carter RR, Bullara LA, Yuen TGH, AgnewWF. Stability of the interface between neural tissue and chronically im-planted intracortical microelectrodes. IEEE Trans Neural Syst Rehab Eng 7:315–326, 1999.

Yang J, Kim DH, Hendricks JL, Leach M, Northey R, Martin DC. Orderedsurfactant-templated poly(3,4-ethylenedioxythiophene) (PEDOT) conduct-ing polymer on microfabricated neural probes. Acta Biomaterialia 1: 125–136, 2005.

Zouridakis G, Tam DC. Identification of reliable spike templates in multi-unit extracellular recordings using fuzzy clustering. Comput Methods Pro-grams Biomed 61: 91–98, 2000.




on March 17, 2009

jn.physiology.orgD

ownloaded from


using a common average reference to improve cortical neuron recordings from microelectrode arrays

Documents