a semi-chronic motorized microdrive and control algorithm ... · cham et al. (jn-00369-2004.r1)...

Cham et al. (JN-00369-2004.R1)

Final Accepted Version - of 14

1. Introduction

A fundamental challenge in experimental neurophysiol-ogy and the development of brain-machine interfaces isto record high quality action potentials from neuronalpopulations. Since the development of the recording elec-trode (Adrian 1926, Hubel 1957, Green 1958), two gen-eral approaches have emerged: acute and chronicrecording methods. In acute recordings, individual elec-trodes are advanced into tissue at the beginning of eachrecording session through a cranial chamber, commonlyby devices termed microdrives. The electrodes areadvanced carefully until high quality neural activity isfound. During the course of an experiment, readjustmentof the electrode's position is often required to re-optimizeand maintain signal quality. While this method allows theexperimentalist flexibility in exploring different celltypes, the process of signal optimization consumes a sig-nificant amount of time, even from an experienced opera-tor. This is especially true for manual microdrives.Although motorized microdrives are commercially avail-able (for example, Thomas Recording GmbH, Germany;FHC Inc., USA; Narishige Inc., Japan), the process ofadvancing the electrodes is still guided by human intu-ition, and it can be tedious or even impractical as thenumber of electrodes increases (Baker et al. 1999).

In chronic recordings, stationary multi-electrode assem-blies, which are typically bundles or arrays of thin wiresor silicon probes, are surgically implanted in the regionof interest (for example, Porada, et al. 2000, Williams et

al. 1999, Rousche and Normann 1998). The signal yieldof the implant array, i.e. the percentage of the array'selectrodes that record active cells, depends upon the luckof the initial surgical placement. Moreover, small tissuemigrations, inflammation, cell expiration or reactive glio-sis can all cause subsequent loss of signal, thereby reduc-ing or disabling the function of the recording array overtime.

Chronic microdrives have been reported by severalinvestigators. Variations of a common design in whichmanual turning of lead screws advances individual orsmall bundles of electrodes include Wall et al. (1967),Kubie (1984), Vos et al. (1999), Venkatachalam et al.(1999), Kralik et al. (2001), Tolias, et al. (2002) andKeating et al. (2002). Hoffman and McNaughton (2002)and deCharms et al. (1999) presented designs for chronicimplants with large arrays of movable electrodes (49 and144 respectively), in which a conventional microdrive ismanually used to push or pull each electrode. Again, theamount of manual operation required to reposition eachelectrode in these devices can be impractical if not intrac-table. Moreover, it is difficult to ensure with such probesthat specific neurons are tracked over time, as neuronalsignals may be lost between adjustments. However,arrays with this many (or more) electrodes are likely to beneeded for future working neuroprostheses. In this case,it is clearly not practical to require periodic manualadjustments of microdrives implanted on paralyzedpatients.

A semi-chronic motorized microdrive and control algorithm for autonomously isolating and maintaining optimal extracellular action potentials

Jorge G. Cham,1 Edward A. Branchaud,1 Zoran Nenadic,1 Bradley Greger,2 Richard A. Andersen,2 and Joel W. Burdick1

1 Division of Engineering and Applied Science, California Institute of Technology, Pasadena, CA 911252 Division of Biology, California Institute of Technology, Pasadena, CA 91125

AbstractA system was developed that can autonomously position recording electrodes to isolate and maintain optimal quality extracel-lular signals. The system consists of a novel motorized miniature recording microdrive and a control algorithm. The micro-drive was designed for chronic operation and can independently position four glass-coated Pt-Ir electrodes with micronprecision over a 5mm range using small (3mm diameter) piezoelectric linear actuators. The autonomous positioning algorithmis designed to detect, align and cluster action potentials, and then command the microdrive to optimize and maintain the neuralsignal. This system is shown to be capable of autonomous operation in monkey cortical tissue.

Articles in PresS. J Neurophysiol (June 30, 2004). 10.1152/jn.00369.2004

Copyright © 2004 by the American Physiological Society.

Cham et al. (JN-00369-2004.R1)


Fee and Leonardo (2001) described a motorizedchronic microdrive with two movable electrodes that issuitable for freely behaving small animals such as thezebra finch. This device, which uses two miniature elec-tric motors, was still operated under human control. Feeand Leonardo reported that the ability to easily adjust theelectrodes led to significant improvements in experimen-tal productivity. This insight is promising for the activi-ties of this paper that go beyond manual electrodeadjustment to automated adjustment of multiple elec-trodes. Related work includes the stabilization of elec-trodes for intracellular recordings over a period of a fewminutes by Fee (2000), and the control architecture for acommercial acute multielectrode microdrive by Baker etal. (1999), which autonomously advances electrodes untiltarget cells are detected, at which point human operatorsoptimize the signal. To date no miniature microdrive hasbeen reported that can seek out, optimize and maintainextracellular action potentials in a fully autonomous fash-ion.

This paper describes a motorized microdrive that canautonomously position four independent electrodes andthat is suitable for semi-chronic operation in monkeys.By the term semi-chronic, we mean that it can be usedwithin a recording chamber for a period of a few weeks.We describe a novel microdrive design, which uses min-iature piezoelectric linear actuators. The underlying the-ory for the autonomous algorithm used to control thedevice is only summarized here, and is described in moredetail in Nenadic and Burdick (2004). Finally, we presentexperimental results including autonomous isolations ofcells in monkey cortex.

The autonomous probe positioning algorithm can beimplemented on a wide variety of devices. Additionally,the mechanical microdrive design can also be used as aconventional human-guided microdrive, and has theadvantage of extremely small size. While the microdriveand control algorithm presented here are independent ofeach other, the system described in this paper serves as atest-bed for developing future "smart" neural recordingimplants that are fully autonomous. The full realizationof the dream of a miniaturized chronically implantableautonomous electrode array will await further advancesin micro-machine technology. In the short term, byreducing the amount of time required for an operator toisolate a cell, by allowing many cells to be isolated inparallel, and by maintaining high signal quality, autono-mously operated microdrives can greatly increase theefficiency with which neurophysiological studies areconducted.

Figure 1. Design of the prototype microdrive. a) Renderedmodel of the design; b) standard cranial chamber used in labo-ratory; c) cross-section of the device inside of chamber, illus-trating relative position to skull and brain tissue; d) front viewof the device: rotation of the motor assembly and positionerallows X-Y positioning of the guide tube.

Skull

Dura

BrainTissue

StandardChamber

AcrylicCover

PiezoelectricActuators

PositionSensor

VerticalPositioner

Knob

FluidDeliveryChannels

Electrodes

Guidetube

ChamberAdapter

Positioner

Motor Assembly

10mm

Guidetube

MotorAssembly

PositionerStandardChamberChamber

Adapter

a) b)

c)

d)

10mm

10mm

VerticalPositioner

KnobChamber

Adapter

Positioner

Guidetube

Motor assemblyand electrodecarrier tubes

StandardChamber

BrassManifold

MotorAssemb.Cap

BrassBushings

Cham et al. (JN-00369-2004.R1)


2. Methods

2.1 Motorized microdrive design

The basic design of our prototype is shown in Figure1a. The prototype is designed to fit inside a standard lab-oratory cranial chamber, used for acute experiments innon-human primates, to allow semi-chronic operation. Asemi-chronic design has the advantage that the device canbe repositioned over a different region with minimaleffort and without need for additional surgeries. Thechamber, shown in Figure 1b, is a hollow plastic cylinderthat is placed over a craniectomy and embedded inacrylic. The chamber can be sealed airtight using a cap,as shown in the figure.

As illustrated in Figure 1c, the device consists of a coremotor assembly that fits inside a gross vertical and hori-zontal positioner. The positioner in turn is fitted insidethe chamber through a chamber adapter. The motorassembly consists of four piezoelectric actuators, a cylin-drical brass manifold and a cylindrical cap, as shown inthe figure. A short guide tube emerges from the cap, withfour inner channels spaced 500 microns apart throughwhich the electrodes are lowered. This guide tube is off-center relative to the motor assembly, while the motorassembly and positioner are arranged as non-concentriccylinders. Rotation of both the motor assembly and thepositioner adjusts the horizontal or X-Y position of theguidetube relative to the chamber over an 6mm diameterarea, as shown in Figure 1d. The positioner is constrainedto move only in the vertical direction relative to thechamber adapter by two slots on its sides that match twoset screws on the adapter. This allows gross vertical or Zpositioning of the electrodes by turning a knob thatengages the outside threads at the top of the positioner, asshown in Figure 1c. Once the horizontal location of theguidetube has been determined and the device placedinside the chamber and lowered vertically by the posi-tioner knob, set screws lock all the parts together, and theelectrodes are then advanced by the actuators.

The electrodes are positioned by custom-made "Nano-motor" piezoelectric linear actuators (Klocke Nanotech-nik, Germany, part NMSB0T10). The actuators, shown inFigure 2a, are 3mm in diameter and 22.5mm in length.These actuators were chosen for their accuracy(unloaded, they can be positioned with nanometer accu-racy), high range of motion (5mm), relatively high forceoutput (up to 0.03N of force), and direct linear drive (nogears or lead screws are needed to convert rotary to linearmotion), thereby avoiding inaccuracies in positioning dueto gearing backlash. The actuators are activated by a saw-

tooth-shaped voltage signal with a minimum peak-to-peak amplitude of 30V. The frequency and amplitude of

5mm

a)

b)

c)

e)d)

BrassManifold

CarrierTubes

PiezoelectricActuators

GuidetubeFront View

Bentelectrodesare back-loaded intothe motorassembly...

PolyimideChannels

Guidetube and capBack View

…and frontloaded intothe assemblycap andguidetube

Loaded motor assembly

0.5mm

Electrode

1 2

3

1

g)f)

ElectrodeWire

Pin and socketconnectors

To head-stage

Electrodesare securedto the brassbushings byset screws.

Figure 2. a) “Nanomotor” piezoelectric actuators; b) frontclose up of the guide tube, showing the four inner channels; c)back view of the motor assembly cap, showing the back of theguide tube; d) loading procedure for electrodes; e) loadedmotor assembly; f) a jig that consists of three machined alumi-num blocks that fit together allows electrodes to be bent at pre-cise pre-determined locations; g) electrode connector.

Cham et al. (JN-00369-2004.R1)


the sawtooth wave determine the speed of positioning(maximum of approximately 2mm/s). The piezoelectricdrives are mounted on a brass manifold that helps absorbthe reaction forces that occur during activation of thedrives. This prevents the motion of each drive fromaffecting the position of the others. In this design, themanifold increases the height and weight of the over-alldevice, though it may be possible to further reduce itssize since its dimensions were chosen conservativelythrough direct experimentation by the actuator manufac-turer.

The piezoelectric element in each actuator drives a hol-low steel carrier tube through its center. Each electrode ispassed through the center of one of these tubes andattached at the top to small brass bushings by set screws.The electrodes (FHC Inc., USA, part UE-RA1), are Pt-Irwires (125 micron diameter), sharpened and glass coateda length of 5mm at the tip, and insulated by .008" ODpolyimide tubing the rest of the length, with 10mmexposed at the end for electrical connection. Typicalimpedance is 1.5 to 3 MOhms at 1kHz. Each electrode isplaced inside a 27ga steel tube, and the tube is pre-bentby 90 degrees in two locations such that one end isaligned with the actuator, and the sharpened end isaligned with the guide tube, as shown in Figure 2d. Sincethe guide tube is off-center to the carrier, each electrodemust be bent a different amount. The electrodes and steeltube are bent using a custom-made system of jigs foraccuracy and consistency, as shown in Figure 2f. A smalllength of the flexible coated wire is left between the benttube and the electrode tip to allow for misalignments inassembly. The guide tube is made from hypodermic steeltubing (0.051" ID diameter) cut to size with four 0.012"ID .016" OD polyimide tubes arranged inside (see Figure2b). The guide tube can be sharpened to penetrate thickdural tissue if necessary.

An alternate cap for the motor assembly can be used forsingle-electrode operation. This cap has a 23ga hypoder-mic stainless steel guide tube aligned with the piezoelec-tric motor drive, such that bending of the electrode is notnecessary.

Electrical connection is made to the electrode by crimp-ing the end of the wire between a pin and socket for D-sub-type connectors, to which a thin wire is soldered, asshown in Figure 2g. This wire is routed to a connectorthat carries the signal to a pre-amplifier.

A position sensor, shown in Figure 1c, was mounted onthe device to track the movement of one of the electrodesfor calibration and initial testing. The position sensorconsists of a Hall-effect magnetic field sensor microchip

(Micronas GmbH, Germany, part HAL401) and a smallmagnet attached to the brass bushing of the electrodedrive. The output voltage of the sensor is proportional toits position relative to the magnet. This sensor is capableof sensing changes in position of one micron over a rangeof 5mm.

2.2 Microdrive fabrication and preparation

The chamber adapter, positioner, turning knob andparts of the motor assembly were machined from Ultempolyethermide (McMaster Carr Supply Co., part8686K76). This material matches the chamber material,and exhibits high temperature and chemical resistance,biocompatibility and machinability properties. Figure 2eshows a picture of the fabricated motor assembly, andFigure 3 shows the final prototype device. Engineering

10mm

Figure 3. Assembled final prototype device and standard cham-ber.

Cham et al. (JN-00369-2004.R1)


drawings of the individual parts are shown in Figure 11.The overall device weighs approximately 40g.

Figure 2d shows how electrodes are loaded into themicrodrive. After the electrodes have been placed insidethe 27ga tubing and pre-bent using the jig system, theyare first front-loaded into the cap of the motor assemblythrough the guide tube. This is facilitated by a small setof plastic cross plates, as shown in Figure 2c, which helpguide the electrode tips into the individual polyimidechannels inside the guide tube. Then, the electrodes andcap are attached to the rest of the motor assembly, back-loading the electrodes through the piezoelectric drive car-rier tubes. For the single electrode configuration, theelectrode is simply back-loaded into the guide tube,through the motor assembly, and attached at the top.

Movement of the piezoelectric microdrives was charac-terized and calibrated by observing and measuring itsmotion under a standard microscope. The actuators canbe commanded to move one step by sending a singlepulse of the sawtooth wave. At full voltage, the actuatorsmove approximately one micron per commanded step.By reducing the voltage amplitude of the pulse, the stepsize could be observably reduced to 0.5 microns. Whilethere is an offset between moving forwards and movingbackwards that must be accounted for when commandingmovements, the actuator movement is quite repeatable.This movement was tested in free air and in gelatin, andfinally in animal cortex, showing only small variations(approximately 10%) in step size. Activation of eachactuator did not cause any discernible unwanted vibra-tions of the electrodes and did not affect the motion of theother actuators.

2.3 Control Algorithm

Controlling a microdrive to isolate and maintain actionpotentials from active neurons in vivo is a difficult taskeven for experienced experimentalists. An autonomouscontrol algorithm must not only discriminate and opti-mize action potentials in signals with significant noiselevels, but must also deal with eventualities such as thepresence of multiple cells, dying cells, cells with low fir-ing rates and transient noise artifacts due to subjectmovements.

The basic architecture of the autonomous control algo-rithm for one electrode consists of two layers. The firstlayer is a state machine, which performs the initial searchfor action potentials, monitors the cell isolation andmaintenance processes, and commands appropriateactions for the eventualities mentioned above. The sec-ond layer is the stochastic optimization method devel-

oped in Nenadic and Burdick (2004). This methodoptimizes signal quality in the presence of action poten-tials, given that only noisy observations are available.

A simplified diagram of the state machine is shown inFigure 4. At each state, the algorithm samples the neuralsignal for a short length of time (in this case, 20 sec) andsearches for action potentials. Depending on the out-come, the state machine may execute a change of stateand/or send a move command to the microdrive to reposi-tion the electrode.

The system is started in the “Initial” state, once themicrodrive device has been positioned over the desiredrecording region inside the chamber. The purpose of thisinitial state is to advance the electrode without frequentpauses to sample the signal, since initial positioning ofthe microdrive may place the electrodes up to severalmillimeters from the desired recording region. Electrodesare advanced 250 microns (at a velocity of 4 microns persecond) between samples until one of the followingevents occur: either a previously determined target depthis reached, which can be obtained from anatomical datasuch as MRI scans (Scherberger et al., 2003), or untilaction potentials are detected in the neural signal.

Action potentials are detected using the unsupervisedwavelet-based detection method presented in Nenadicand Burdick (2004). Traditional methods such as ampli-tude and power thresholding, window discrimination andmatched filtering require human supervision and experi-ence, as they depend on the amplitude, shape and phase

Initial

Search

Isolate

Maintain

Targetdepth isreached

Signallost

Spikesdetected

Spikesdetected

Signalqualitydrops

Acceptablecell isolated

Figure 4. Simplified diagram of the autonomous algorithm statemachine.

Cham et al. (JN-00369-2004.R1)


of the action potentials recorded, which can change as theelectrode moves relative to the cell body. The wavelet-based method is better suited for unsupervised operationbecause its detection threshold is independent of spikeshape and phase, and can be set beforehand as a trade-offbetween spike omission and false alarms. This methodhas been shown to give consistent performance over awide range of signal-to-noise ratios and firing rates.Action potentials are considered to be present only whenthe number of events detected by the method exceeds aminimum firing rate, in this case 2Hz. Once detected,action potentials are aligned and clustered using the cor-relation method and finite mixture model clusteringmethod described in Nenadic and Burdick (2004).

If the target depth is reached without detection of actionpotentials, the algorithm switches to the “Search” state,which advances the electrode only 50 microns (at 4microns per second) between samples. If action poten-tials are detected while in the “Initial” or “Search” states,the system switches to the “Isolate” state.

The goal of the “Isolate” state is to reposition the elec-trode to maximize signal quality. In the method ofNenadic and Burdick (2004), this goal is mathematicallyformalized by defining a (non-negative) objective func-tion over a segment of a real line in the neighborhood ofthe cell (since the electrode is constrained to linearmotion). The resulting curve is called the “cell isolationcurve”, and the goal is to find the position that maximizesit. The method presented here is independent of the exactchoice of objective function, but it must capture some

measure of signal quality. In this paper, we test the use oftwo different metrics, though others are certainly possi-ble. The first metric tested is the peak-to-peak amplitude(PTPA) of the recorded action potentials. The secondmetric is the distance in principal component space(DPCS) between a useful cell and confounding cells ornoise, which may be useful in the presence of multiplespiking neurons.

Since only noisy observations of the objective functionare available, the objective is defined as a regressionfunction,

(1)

where y is the chosen measure of signal quality, x is theposition of the electrode along its range of motion and

is the expectation operator. Normally, gradientdescent methods may be applied to find the maximum ofthis function. However, estimating the gradient directlyfrom noisy local observations can be numerically unsta-ble. Instead, we estimate the regression function itselfwith a set of basis functions, using all (or a subset of) theprevious observations obtained at the preceding electrodepositions. For simplicity, we choose a polynomial basisfunction of the form:

(2)

where x is electrode position, are

the polynomial coefficients, and n is the order of the

M x( ) E y x( )=

Headstage

ElectrodePositionCommand

A/D DataAcquisition

ActuatorController Box

Amplifiers andFilters

MotorizedMicrodrive isplaced inchamber overregion ofinterest

ActuatorCommands

Neural Signal

1 2 Signals from the electrodes are filtered, amplifiedand read by a data acquisition card

3The control algorithmdetects, aligns, and clusters action potentials, computessignal quality objective function and estimates optimal electrode movement

Workstationand controlalgorithm

4 Motor commandsare sent to the actuatorsand the electrodes arerepositioned for optimalsignal quality

Figure 5. System diagramfor the autonomous sys-tem. The system isolatesand optimizes neural sig-nals via a closed loopposition control of theelectrodes based onmodel-based estimates ofsignal quality.

E( )⋅

M̂ x n B, ,( ) bixi-1( )

i 1=

n

∑=

B b1 b2 … bn, , ,[ ]T=

Cham et al. (JN-00369-2004.R1)


model. The order of the model is adaptively chosenthrough Bayesian theory. For each model order up to amaximum (in our case, 5th order), the posterior probabil-ity of that model given the observed data is computed andthe model order that maximizes this probability is thenchosen. Once the order is determined, the parameters Bare found through the maximum likelihood method(ML), which, in this case, reduces to the linear leastsquare method (LLS).

Once an estimate of the objective function is obtained,the algorithm moves towards the maximum using a formof Newton’s method:

(3)

where xk+1 and xk are the positions of the electrodes atiterations k+1 and k, respectively, C > 0 is an appropri-ately chosen scale factor, and Hk and are estimates of

the Hessian and first derivative of M(x) at xk respectively.The gradient and Hessian are computed explicitly fromformulas of the first and second derivatives of Equation2.

Initially, the “Isolate” state moves the electrode in con-stant increments (in our case, 20 microns) between sam-ples of the neural signal. The basis function

approximation and Newton’s method optimization areused only after enough observations of the regressionfunction have been collected. This is necessary becausethe Bayesian formulation assumes that the number ofobservations is greater than the highest model order con-sidered for the approximation. Otherwise, overfittingmight occur. In our case, the algorithm verifies two con-ditions before the basis function approximation is consid-ered valid. First, it checks that a minimum number ofobservations k0 have been collected (in our case, thisnumber is 5). Second, if enough observations have beencollected then the most likely model order must begreater than 1 (meaning a horizontal straight line). Ifeither of these conditions is not met, the algorithm con-tinues sampling the space in constant increments. If bothconditions are met, the optimization process proceedsaccording to Equation 3.

The state machine remains in the “Isolate” state untileither an upper bound on signal quality is reached, or it isdetermined that the maximum of the regression functionhas been reached. The first criterion prevents the algo-rithm from driving the electrode into the body of a cellthat lies directly in its path. This upper bound must bechosen large enough that spiking information can be dis-cerned, but small enough that the electrode will be kept ata safe distance from the cell. In our case, it is defined as a

xk 1+ xk C Hk-1ξk+= k k0 k0 1+ …, ,=

ξk

Filt

ered

Sig

nal (

V)

Sample band-passed signal from rat cortex

Detected and aligned monkey cortexcell action potentials

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6

-80

-40

0

40

80

Time (ms)

Vol

tage

(m

icro

volts

)

-5

0

5

0 4 8 12 16 20Time (s)

Sample band-passed signal from monkey cortex

Filt

ered

Sig

nal (

V)

Vol

tage

(mic

rovo

lts)

Detected and aligned rat cortexcell action potentials

0 0.2 0.6 1 1.4 1.8 2-80

-40

0

40

80

Time (ms)

0 2 4 6 8 10 12 14-0.8

-0.4

0

0.6

1

Time(s)

Figure 6. Filtered neural signals obtained with the microdrive in rat and monkey cortex.

Cham et al. (JN-00369-2004.R1)


form of signal-to-noise ratio (SNR) as the amplitude ofthe detected spikes divided by the root mean square of the“noise” (recorded signal with the spikes subtracted). Forthe second criterion, the algorithm is determined to havereached the maximum of the regression function whenthe commanded step size from Equation 3 reaches a min-imum value, in this case 1 micron, which will occur whenthe gradient approaches zero. If the maximum value thatis realized is below a lower bound of signal quality, thenthe cell isolation is considered unacceptable, and thealgorithm switches back to the “Search” state. In theexperiments presented in this paper, we tested variousvalues of the upper and lower bound SNR, finding bestresults with values of 12 and 8 respectively.

If the cell is considered isolated, the state machinechanges to the “Maintain” state. In this state, the algo-rithm samples the neural signal while keeping the elec-trode stationary. This continues until the measured signalquality falls below the acceptable SNR level. Once thisoccurs, the cell is no longer considered isolated, and thealgorithm switches back to the “Search” state in order tore-acquire the signal.

A significant problem occurs when the firing rate of acell being isolated is intermittent, or the cell stops firingaltogether, especially in the presence of other nearbycells. Such situations can create extreme outliers in sam-pling that can confound the isolation algorithm. Thiseventuality was remedied by increasing the recordingtime at each sample. This time was set to 20 seconds,which allowed activity from intermittent firing cells to becaptured consistently. The threshold for determining thepresence of an active cell was an average firing rate of2Hz over the 20 second sample period. If, while trackinga cell in the “Isolate” or “Maintain” state, the firing ratedrops below this threshold, the signal is sampled at thatposition one more time before determining that the cell isno longer present and resetting to the “Search” state onceagain.

2.4 Experiment setup and procedure

Initial experiments were performed on anesthetized ratsto test the basic operation of the microdrive and controlalgorithm. The rats were anesthetized with isoflurane,administered ketamine (1ml/Kg, IP) and atrophine(0.05ml, subcutaneously) and prepared for surgery. Onceanesthetized, the rats were held in a sterotaxic rig via earbars and mouth piece, and administered isofluraneaccordingly to maintain sedation while heart and temper-ature were monitored. A small craniectomy was per-formed over the barrel cortex region, and the microdrive

was suspended over the craniectomy using a stereotaxicarm. The electrodes were then lowered under operatorcontrol through dura and into cortical tissue using thepiezoelectric motors in both multiple and single-electrodedrive configurations.

Single-electrode experiments were conducted in theposterior parietal cortex (Andersen and Buneo, 2002) ofan awake, behaving adult macaque monkey. The micro-drive was installed in the cranial chamber at the begin-ning of each recording session. Using the verticalpositioning knob, the device was lowered by hand to atarget depth. The algorithm was then activated and thesystem operated autonomously, using the PTPA metric.The monkey performed simple saccade (eye movement)tasks in a darkened room during the experiments. Theanimal care and handling in these experiments were inaccord with the guidelines of the National Institutes ofHealth and have been reviewed and approved by the localInstitutional Animal Care and Use Committee.

Between recording sessions, the electrodes wereremoved from the microdrive and sterilized, and themicrodrive was disassembled and cleaned with alcohol.Typical microdrive assembly and disassembly times were20-25 minutes by an experienced operator with goodmanual dexterity. Alcohol was also sometimes applied tothe junction between the piezoelectric element and thecarrier tube of the actuators before each recording session

3870 3890 3910 3930 39501.2

1.4

1.6

1.8

2

2.2

2.4

2.6

Position (microns)

Dis

tanc

e in

PC

Spa

ce

Isolation curve in rat cortex

Figure 7. Isolation curve recorded from rat cortex. The dashedline shows the measured signal quality as a function of elec-trode position with sample spike forms (solid blue lines) at dif-ferent positions indicated by the dotted lines. The solid red lineshows the fitted basis function approximation.

Cham et al. (JN-00369-2004.R1)


to clear out any small debris or viscous biological liquidthat impeded motion.

A diagram of the autonomous system is shown in Fig-ure 5. The output of the electrode from the microdrivewas connected to a DAM-80 (World Precision Instru-ments Inc., USA) headstage and amplifier in the ratexperiments, and to a Plexon (Plexon Inc., USA) head-stage and amplifier in the monkey experiments. The sig-

nal output from the amplifiers in both setups wasrecorded by a data acquisition card (National InstrumentsInc., USA). Filter and gain settings varied with experi-mental conditions and objectives. A faraday cage wasused to shield the set up from ambient noise in the ratexperiments. The carrier tubes of the actuator were con-nected to ground (they are electrically insulated from theactuator signal) to provide additional shielding to theelectrode. The piezoelectric motors were powered and

0

5

10

x 105

-5

0

5

-5

0

5

0

5

10

x 105

-5

0

5

-5

0

5

0

5

10

x 105

-5

0

5

-5

0

5

0

5

10

x 105

-5

0

5

-5

0

5

0

5

10

x 105

-5

0

5

-5

0

5

0

5

10

x 105

-5

0

5

-5

0

5

0

5

10

x 105

-5

0

5

-5

0

5

2750 2800 2850 29000

5

10

0 5 10 15 20-5

0

5

0 10 20 30-5

0

5

Position (um) Time (s) Time (ms)

Obj

ectiv

e F

unct

ion

Sam

ple

Dat

a S

tream

Ave

rage

d W

avef

orm

2720 2740 2760 2780 2800 2820 2840 2860 2880

-2

0

2

4

6

8

Position (microns)

PT

P A

mpl

itude

(V

olts

)

a)

b)

Figure 8. Cell Isolated in monkey cortex using the autonomous semi-chronic microdrive system. a) Progression of algorithm inpresence of action potentials. The left column shows snapshots of the sampled objective function and the basis function approxi-mation; the middle column shows the signal spike train; the right column shows the averaged waveform of the detected spikes. b)Final isolation curve and average spike waveforms at each position. In this case, the signal was optimized until the maximum ofthe isolation curve was found.

Cham et al. (JN-00369-2004.R1)


activated by a controller purchased with the actuators(Klocke Nanotechnik, Germany, part NWC). The posi-tion sensor was powered by a standard 5V power supply,and its output also read by the data acquisition card. Thecontrol algorithm was implemented in Matlab (Math-works Inc., USA).

3. Results

The motorized microdrive in single-electrode configu-ration and the control algorithm were used to record fromneurons in rat and monkey cortex. Sample filtered neuralsignals obtained with the microdrive in rat and monkeycortex are shown in Figure 6. The top plots show rawdata streams bandpass filtered between 300Hz and 10kHzfor the rat and monkey. The bottom plots show thedetected, aligned and clustered spikes from their respec-tive data streams using the detection and alignment meth-ods previously discussed. Figure 7 shows sample resultsfrom rat cortex. Shown in the figure is signal quality (inthis case measured by the DPCS metric) as a function ofelectrode position, with the corresponding averaged spikeshapes at each position. These results demonstrate thepresence of our conceptual isolation curves in rat cortex,which are the basis of our autonomous algorithm.

Figure 8 shows a sample result of autonomous cell iso-lation in monkey cortex. In this case, the algorithm wasinitiated after the microdrive was installed in the chamberand allowed to operate autonomously without humanintervention. The algorithm first advanced the electrodein the “Initial” and “Search” states for over 1.5mm untilfaint spike activity was detected. Shown in Figure 8a isthe sequence of steps taken by the algorithm once thestate machine transitioned to the “Isolate” state. The firstcolumn shows the positions of the electrode and the mea-sured objective function (in this case, PTPA). The secondcolumn shows the corresponding data stream at that posi-tion, and the third column shows the average spike wave-form. As shown, the algorithm first advanced theelectrode in constant increments. After enough observa-tions were made to allow an adequate model to be fitted,the polynomial fit was made, shown as a solid line. Oncea peak in the objective function was detected, the controlalgorithm repositioned the electrode toward the optimallocation. The sequence ended when the cell was deter-mined to be isolated by the algorithm, as per the mini-mum step size criteria previously discussed. Figure 8b isa concatenated plot of the sequence, showing the finalisolation curve approximation (solid line), the progres-sion of the algorithm (dotted line) and the average wave-forms.

The results from Figure 8 illustrate the potential risk ofunconstrained signal maximization. As shown, the firingrate of the cell increased as the electrode moved forward,indicating that the electrode was affecting cell behaviorpossibly due to very close proximity. Figure 9 showsfinal results of the algorithm in monkey cortex in whichthe upper bound on SNR was implemented, in order toprevent potential damage to the cell. In this case, the

1810 1830 1850 1870

0.2

0.4

0.6

Position(microns)

Dis

tanc

e in

PC

Spa

ce

Cell isolation in monkey cortex

0 0.5 1 1.5 2 2.5-50

0

50

0 0.5 1 1.5 2 2.5-50

0

50

0 0.5 1 1.5 2 2.5-50

0

50

0 0.5 1 1.5 2 2.5-50

0

50

Time (s)

Sam

ple

Dat

a S

tream

(m

V)

Figure 9. Cell isolated in monkey cortex using the autonomoussemi-chronic microdrive system. In this case, the cell was con-sidered to be isolated when the signal quality reached a maxi-mum value, in order to avoid potential damage to the cell. Thetop plot shows the sampled signal quality (in black) and the fit-ted regression function (in red) as a function of electrode posi-tion, with average action potentials shown for four differentpositions. The data streams that correspond to the four posi-tions are shown below (in blue), with correspondence indicatedby the dashed lines.

Cham et al. (JN-00369-2004.R1)


algorithm advanced the electrode to maximize the regres-sion function until the upper bound was reached, at whichpoint the cell was considered isolated.

To date, ten cells have been isolated with the autono-mous system in monkey cortex. Cells have been isolatedand maintained for up to 3 hours. Figure 10 shows a sam-ple time history of signal quality after the cell has beendetermined isolated by the algorithm and the statemachine entered the “Maintain” state. As shown, signalquality (measured by the SNR) degrades over time, pos-sibly due to tissue migration. Once the signal droppedbelow the lower bound SNR threshold, the algorithmautomatically re-initiated the search and isolation pro-cesses and reacquired the signal. The continuous mea-surements of the PTPA metric shown in the figure andthe consistency of the spike shape provide evidence thatthe same cell was being tracked throughout the entireexperimental session.

4. Discussion

4.1 Motorized microdrive design

The novel miniature motorized microdrive was showncapable of advancing and retracting electrodes in corticaltissue with micron precision and recording high-qualityneural signals. The electro-mechanical design of themicrodrive addresses several issues in recording implantdesign. First, the overall device is of minimal size andweight so that it does not significantly affect behavior inawake non-human primates and can be implanted semi-chronically (for several days or weeks at a time). Many ofthe commercially available motorized microdrives suchas the ones mentioned in the introduction use relativelylarge actuators and are meant only for acute experiments.

0 50 100 150

4920

4960

5000

5040

Time (minutes)

Ele

ctro

de D

epth

(mic

rons

) 8

9

10

11

12

13

Sig

nal t

o N

oise

Rat

io

InitialIsolation

Isolation Threshold

Cell Re-isolated

Algorithm repositions electrodeto re-isolate signal quality

Figure 10. Time history of signal quality over a period of 2.5 hours, after the cell was isolated by the autonomous algo-rithm. The figure shows how signal quality slowly degraded, perhaps due to tissue migration. Once signal qualitydegraded below a minimum threshold, the autonomous algorithm repositioned the electrode to reacquire an acceptablesignal quality level.

Cham et al. (JN-00369-2004.R1)


Second, miniature actuators often have very small forceoutput, and require special attention to minimize losses inpower from, for example, friction due to misalignment.The piezo-electric linear actuators used in our design arenovel to this application and were shown to be capable ofsubmicron precision as well as sufficient force output for

advancement in cortical tissue (but not for advancementthrough monkey dura, requiring a sharpened guide tubeto pierce it). Such precision is necessary to obtain optimalsignal quality, given that action potentials from a typicalcell can be lost by movements as small as 50-100 microns(Rall, 1962; Lemon, 1984; Gray et al., 1995). The long

a b

cd

e f

g h0-80 threaded 1-64threaded

1.52

A A

B

B

24 threads per inchoutside thread

2.38

A-A

12

1B-B

K

K 24 threads per inchinside thread

F

F F-F

M

M

J

J

9.52

5.25

3.27

J-J

M-M

E

E

E-E

holehole

20

22.23

10

K-K

35.23

25.23

1.506.35

Figure 11. Engineering drawings of the individual parts of the prototype motorized microdrive. a, b, c) Bottom view andcross sectional views of the positioner. d) The two parts making up the motor assembly cap. e) Turning knob for gross verti-cal positioning. f) Piezo-electric actuator and brass manifold subassembly. g) Chamber adapter. h) Chamber. All dimen-sions are in millimeters.

Cham et al. (JN-00369-2004.R1)


stroke of the actuators also makes them well suited to thisapplication, since a range of motion of several millime-ters, if not centimeters, is often required depending on thedepth of the target structure, and the accuracy of theimplantation procedure. In addition, the inherent linearmotion of the actuators reduced the amount of carefulassembly necessary and did not require the design of alead screw transmission. Gears and lead screws can oftenintroduce a significant amount of imprecision in the drivedue to gearing backlash. Such hysteresis can result ininstability or inaccuracy in the control algorithm. In gen-eral, the actuators were tolerant to significant handlingand electrode loading and unloading. The drive mecha-nism of the actuators can be sensitive to friction due tomisalignment or viscous liquids, reducing the force out-put and requiring careful attention in assembly and main-tenance.

The microdrive was tested in both multiple and singleelectrode configurations. Typical signal-to-noise ratios ofspike amplitude were 8-10, and as high as 15. Withproper shielding and grounding, the electronic noise gen-erated by the idle actuators was on the same order ofmagnitude as ambient and 60Hz noise from surroundingequipment, and did not significantly impact signal qual-ity. During motion of the electrodes, the actuators gener-ated noise that saturated the electrode signal. However,since the control algorithm requires that the signal besampled at discrete positions, simultaneous motion andrecording was not necessary.

The microdrive performed well over several dozenrecording sessions without sign of performance degrada-tion. The device is rugged in construction, safe and rela-tively easy to install in the head chamber and to reloadand maintain electrodes. Careful attention must be paidwhen front-loading the electrodes into the device, as thefragile electrode tips can be easily damaged.

4.2 Control algorithm

The algorithm presented in this paper was shown toautonomously command the microdrive to seek and iso-late action potentials from cells. All of the different meth-ods used in the isolation algorithm, from spike detection,alignment and clustering to regression function modelselection and estimation, require no supervision andaccount for the stochastic nature of the task. The resultsshown for monkey cortex were obtained with no humanintervention once the microdrive was placed inside thechamber.

The algorithm and isolation results presented were forone electrode. Interference in the form of tissue move-

ments between multiple movable electrodes is possible,however, and strategies for coordinating the movementof multiple electrodes will have to be developed.

While the stochastic optimization method of Nenadicand Burdick (2004) is deeply rooted in theory, the heuris-tic-based state machine architecture allows for flexibilityin dealing with unpredictable eventualities common inrecording environments in behaving animals. Many ofthe state transitions for dealing with eventualities arebased on common practices by experienced humanexperimentalists. Although the basic implementation pre-sented here was shown to work in the monkey, the exactheuristics and values used by the state machine may needto be fine-tuned for particular experimental conditionssuch as subject species, brain region and type of cell tar-geted in order for the autonomous algorithm to performoptimally.

4.3 Autonomous system

The novel microdrive and the algorithm presented heredo not necessarily need to be implemented together. Themicrodrive design provides a working device for acute orsemi-chronic recordings that can be controlled by ahuman operator, or by an alternate control algorithm.Similarly, the algorithm can be used to control othermicrodrives for autonomous operation. The successfulintegration of the two systems, however, is presented as afirst step towards future “smart” neural implants that arefully autonomous. Such autonomous implants could con-tribute to the efficiency and flexibility of neurophysiolog-ical studies by freeing the experimentalist from time-consuming tasks such as frequent implantation surgeriesand finding and maintaining high-quality neural signals.

5. Acknowledgements

We thank the members of the Andersen lab at Caltech,especially Daniella Meeker, Bijan Pesaran, BorisBreznen, Sam Musallam, Kelsie Pejsa and Lea Martel.Thanks also to Aimee Eddins and Rodney Rojas for helpin fabricating the prototype. This work was funded byNIH, DARPA, the Boswell Foundation, ONR and NSF.

6. References

Andersen RA and Buneo CA. Intentional maps in posteriorparietal cortex. Ann.Rev. Neurosci 25: 189-220, 2002.

Adrian E. The impulses produced by sensory nerve endings.J Physiol 61: 49-72, 1926.

Baker SN, Philbina N, Spinksa R, Pinchesa EM, WolpertaDM, MacManusb DG, Pauluisc Q and Lemon RN. Multi-

Cham et al. (JN-00369-2004.R1)


ple single unit recording in the cortex of monkeys usingindependently moveable microelectrodes. J. NeuroscienceMethods 94: 5-17, 1999.

deCharms RC, Blake DT and Merzenich MM. A multielec-trode implant device for the cerebral cortex. J NeurosciMethods, 93: 27-35, 1999.

Fee MS and Leonardo A. Miniature motorized microdriveand commutator system for chronic neural recording insmall animals. J Neurosci Methods 112: 83-94, 2001.

Fee MS. Active stabilization of electrodes for intracellularrecording in awake behaving animals. Neuron 27: 461-468, 2000.

Gray CM, Maldonado PE, Wilson M and McNaughton B.Tetrodes markedly improve the reliability and yield ofmultiple single-unit isolation from multi-unit recordings incat striate cortex. J Neuroscience Methods 63: 43-54.

Green J. A simple microelectrode for recording from thecentral nervous system. Nature 182: 962, 1958.

Hoffman KL and McNaughton BL. Coordinated Reactiva-tion of Distributed Memory Traces in Primate Neocortex.Science 297: 2070-2073, 2002.

Hubel D. Tungsten microelectrode for recording from singleunits. Science 125: 549-550, 1957.

Keating JG and Gerstein GL. A chronic multi-electrodemicrodrive for small animals. J Neurosci Methods 117:201-206, 2002.

Korshunov VA. Miniature microdrive for extracellularrecording of neuronal activity in freely moving animals. JNeurosci Methods 57:77-80, 1995.

Kralik JD, Dimitrov DF, Krupa DJ, Katz DB, Cohen D andNicolelis MAL. Techniques for long-term multisite neu-ronal ensemble recordings in behaving animals. Methods25: 121-50, 2001.

Kubie JL. A Driveable bundle of microwires for collectingsingle-unit data from freely-moving rats. Physiology &Behavior 32: 115-118, 1984.

Lemon R. Methods for neuronal recording in conscious ani-mals. Pages 18-37, 1984.

Nenadic Z, Burdick JW. Autonomous electrode positioningfor optimization of extracellular recordings. IEEE TransBiomedical Eng, submitted, 2004.

Nenadic Z, Burdick JW. Spike detection using the continu-ous wavelet transform. IEEE Trans Biomedical Eng, inpress, 2004.

Normann RA, Maynard EM, Rousche PJ and Warren DJ. Aneural interface for a cortical vision prosthesis. VisionResearch 39: 2577-2587, 1999.

Porada I, Bondar I, Spatz WB and Kruger J. Rabbit andmonkey visual cortex: more than a year of recording withup to 64 microelectrodes. J Neurosci Methods 95: 13-28,2000.

Rall W. Electrophysiology of a dendritic neuron model. Bio-phys. J. 2: 145-167, 1962.

Rousche PJ and Normann RA. Chronic Recording Capabil-ity of the Utah Intracortical Electrode Array in Cat Sen-sory Cortex. J. Neurosci. Meth., 82:1-15 (1998).

Scherberger H, Fineman I, Musallam S, Dubowitz DJ, Bern-heim KA, Pesaran B, Corneil BD, Gilliken B, and Ander-sen RA. Magnetic resonance image-guided implantationof chronic recording electrodes in the macaque intrapari-etal sulcus. J Neurosci Methods 130:1-8, 2003

Strumwasser F. Long-term recording from single neurons inbrain of unrestrained mammals. Science 127: 469-470,1957.

Tolias AS, Siapas AG, Smirnakis SM, Logothetis NK. Cod-ing visual information at the level of populations of neu-rons. Program No. 557.5. 2002 Abstract Viewer/ItineraryPlanner. Washington, DC: Society for Neuroscience,2002.

Venkatachalam S, Fee MS and Kleinfeld D. Ultra-miniatureheadstage with 6-channel drive and vacuum-assistedmicro-wire implantation for chronic recording from theneocortex. J Neurosci Methods 90: 37-46, 1999.

Vos BP, Wijnants M, Taeymans S and De Schutter E. Min-iature carrier with six independently moveable electrodesfor recording of multiple single-units in the cerebellar cor-tex of awake rats. J Neurosci Methods 94: 19-26, 1999.

Wall PD, Freeman J, Major D. Dorsal horn cells in spinaland in freely moving rats. Exp Neurol 19: 519-529, 1967.

Williams JC, Rennaker RL and Kipke DR. Long-term neu-ral recording characteristics of wire microelectrode arraysimplanted in cerebral cortex. Brain Research Protocols 4:303-313, 1999.

a semi-chronic motorized microdrive and control algorithm ... · cham et al. (jn-00369-2004.r1)...

Documents