Microelectrode array recordings of neuronal activity present significant opportunities for studying the brain with single-cell and spike-time precision. However, challenges in device manufacturing constrain dense multisite recordings to two spatial dimensions, whereas access to the three-dimensional (3D) structure of many brain regions appears to remain a challenge. To overcome this limitation, we present two novel recording modalities of silicon-based devices aimed at establishing 3D functionality. First, we fabricated a dual-side electrode array by patterning recording sites on both the front and back of an implantable microstructure. We found that the majority of single-unit spikes could not be simultaneously detected from both sides, suggesting that in addition to providing higher spatial resolution measurements than that of single-side devices, dual-side arrays also lead to increased recording yield. Second, we obtained recordings along three principal directions with a multilayer array and demonstrated 3D spike source localization within the enclosed measurement space. The large-scale integration of such dual-side and multilayer arrays is expected to provide massively parallel recording capabilities in the brain.
Monitoring the interplay of neuronal ensembles in the brain is important for understanding mechanisms underlying learning, memory, and behavior. Using advances in multisite microelectrode fabrication techniques, it is possible to measure the activity of tens to hundreds of neurons in parallel (Buzsáki 2004). Integrated microelectronic circuits are poised to facilitate a transition to even higher recording capacities (Olsson et al. 2005). However, significant challenges remain in the development of implantable devices for sampling extracellular neuronal activity with high spatial resolution across a region of interest. Most existing electrode microelectrode arrays are distributed along two spatial dimensions, inherently restricting their ability to capture the dynamics of large neuronal ensembles exhibiting complex connectivity patterns. A method for extending extracellular measurements into the three-dimensional (3D) domain would provide a unique high-resolution functional mapping tool for the brain. Significant effort has thus been made to create implantable microstructures supporting a dense 3D array of recording electrodes (Wise et al. 2008).
Microwire arrays (Nicolelis et al. 1997) and analogous silicon-based microstructures (Campbell et al. 1991) provide limited scope for dense 3D measurements because their sampling volume is constrained by the availability of only one recording site per wire. An alternative architecture, the planar microelectrode array (Najafi et al. 1985), uses electrodes arranged on one or more implantable silicon shafts. The principal advantage of this design lies in the ability to simultaneously obtain depth and axial information with a high level of spatial precision. Tests with such devices have demonstrated their effectiveness in densely sampling neuronal activity within a single plane (Blanche et al. 2005; Csicsvari et al. 2003). Moreover, an extension of the traditional planar architecture has been proposed in the form of double-sided electrodes (Perlin and Wise 2004). These devices contain electrodes on two parallel planes, separated by the thickness of the implantable shaft, and manifest a simple, localized 3D recording geometry. Most of the large-scale 3D arrays involve modular assembly of planar arrays using multiple structural layers (Bai et al. 2000; Hoogerwerf and Wise 1994; Neves et al. 2007; Pang et al. 2005; Yao et al. 2007). Research on such devices has primarily focused on important aspects of the fabrication process, insertion into the cortex, and chronic response of tissue to the implant. However, despite achieving impressive levels of device complexity, high spatial resolution extracellular measurements do not appear to have been carried out in 3D with existing neural probe microstructures.
Here, we demonstrate 3D recording in devices created by a silicon fabrication method involving dual-side processing and assembly techniques. The design introduces two new recording functionalities, which were explored in vivo with two separate devices. Attributes of both devices enable 3D extracellular measurements with a high level of spatial and temporal precision. First, we developed a dual-side electrode array by placing separately addressable recording sites on each planar surface of the implant. This allowed us to increase the density of recording sites in the region of interest by twofold. Moreover, our results indicate that conventional single-side arrays may shield spike activity of some neurons facing the other side of the array. Second, we assembled a double-layer probe using a pair of planar microstructures placed 100 μm apart. A unique feature of this device was that the planar recording sites on both layers faced inward, toward the space between the two layers. This enabled us to record and triangulate the source of extracellular action potential signals from several sites within the enclosed region.
The arrays were built using surface and bulk micromachining techniques. The starting material was a silicon substrate with a thickness of 25 ± 5 μm or 50 ± 5 μm, which determined the final device thickness (Du et al., unpublished observations). Gold recording sites (100 μm2) were patterned on both sides of the substrate, by first defining the features on the front side, then turning the substrate over and repeating the metallization process on the back. The top and bottom insulation layers on each side consisted of 2-μm films of parylene C and thermal oxide, respectively. The silicon device substrate was patterned into its final shape by means of a deep reactive ion etch process. A representative device is shown in Fig. 1A. The shaft dimensions are 4 mm × 70 μm (l × w) and its silicon layer thickness is 50 μm, although 25-μm-thick devices were also used in the course of these experiments. The end of each shaft contains eight electrodes on the front as well as the back side (Fig. 1B) and is tapered to facilitate tissue penetration.
Assembly proceeded by flip-chip bonding onto a flexible printed circuit board (PCB) using an anisotropic conductive adhesive film, which permitted current to flow out of plane but remained highly resistive in plane. The flip-chip bonding process entirely eliminated the need for wire bonds in all of our devices. In addition, it resulted in an essentially flat contact profile between the silicon device layer and PCB; this attribute was exploited to connect a PCB to each side of the dual-side array (Fig. 1C), by first flip-chip bonding the contacts on the front side, turning the structure over, and finally bonding with another PCB to the back-side contacts. This novel assembly method raises the prospect for creating modular, 3D microstructures. A simple modular device was realized by bonding two carefully aligned silicon layers on each side of a PCB to create a double-layer structure (Fig. 1, D and E). After assembly, the contact regions were sealed with epoxy. Finally, to maintain a high signal-to-noise ratio (SNR) and minimize electrical cross talk, electrode impedance was reduced from roughly 2.1 to 0.2–0.25 MΩ with electrodeposited gold. A detailed analysis of cross talk is presented in the appendix. The finished devices contained 32 connected recording channels: 16 per side and 16 per layer for the dual-side and double-layer arrays, respectively.
Single-unit and local field potential (LFP) recordings
Acute measurements were carried out in locusts (Schistocerca americana), whose antennae and body were fixed and brain desheathed and perfused with saline (Laurent and Davidowitz 1994). Signals were sampled at 15 kHz during ten consecutive 20-s trials, in which a 1-s cis-3-hexanol or apple-odor pulse was applied to the antenna. Single-unit measurements were made with devices inserted in the antennal lobe, which contains projection neurons that are known to respond to odor stimuli (Laurent and Davidowitz 1994). The probe was inserted at a maximum speed of roughly 10 μm/s and lowered to a final depth of 250–300 μm with respect to the tip, corresponding to approximately the length of the antennal lobe and the full span of the electrode array. Signals were fed to two custom-built 16-channel head stages, followed by main-stage amplifiers and band-pass filters, before being stored for off-line analysis. Figure 2 shows representative multisite recordings captured with one of our devices. The high recording site density facilitates measurement of the same unit at many sites on the array (Fig. 2, A and B). Single-unit clusters were identified using a spike-sorting algorithm (Mazor and Laurent 2005; Perez-Orive et al. 2002), which involved analyzing signals from four electrodes (“tetrodes”; Gray et al. 1995). The putative units obtained from several tetrode sets were then combined, while taking care to avoid double-counting units observed on multiple tetrodes. Raster plots of two projection neurons measured in parallel are shown in Fig. 2C. As a further demonstration of the probe's recording functionality, LFPs were measured in the mushroom body (Fig. 2D), displaying characteristic odor-evoked oscillations. The odor-evoked responses of single units and LFPs were found to agree with previous studies using extracellular measurements in the locust (Perez-Orive et al. 2002).
Dual-side electrode array
Figure 3, A and B shows the noise-averaged spike waveforms of two putative projection neurons that were measured in parallel on both sides of a dual-side array. The device was 25 μm thick. Signals from the first unit were strongest on the front, but some appreciable signals were also measured on the back. In contrast, the second unit is measured almost exclusively with back-facing sites. Figure 3 leads to two salient observations: 1) that extracellular fields decay on the scale of tens of microns beyond the presumed vicinity of the neuron (Henze et al. 2000) and 2) that ordinary spatial attenuation of fields may not fully explain the signal amplitude asymmetry between the two sides.
To gain further insight into the significance of signal attenuation effects in dual-side array recordings, we compared spike amplitude decay profiles that were measured from the same side with those from opposite side of the array. Figure 4A shows the normalized spike amplitude as a function of separation from the site of maximum measured amplitude (V0). The results display a high degree of variability, but suggest some important trends. In particular, signals from the V0 side (blue points) decayed over longer lengthy scales than those from the opposite side (red points). Moreover, the amplitude was found to drop by close to 40% at a distance of 25 μm on the V0 side, whereas it was about 80% attenuated at the same relative separation on the opposite side. This asymmetry is likely due to extracellular current shielding at the substrate–fluid interface (Moffitt and McIntyre 2005; Perlin and Wise 2004). A consequence of this effect is that dual-side arrays may routinely detect more units in the implanted region than their single-side counterparts.
Figure 4A also suggests that, on average, extracellular fields on the same side of the array decay in roughly inverse proportion to distance. We approximate the normalized peak-to-peak signal measured by an electrode as (1) where xr is the radial distance from the electrode and λ is a characteristic decay length. The best fit of Eq. 1 to the points in Fig. 4A gave values of λ = 33 ± 2 μm and n = 1.4 ± 0.2. Similar characteristic length scales have been reported elsewhere (Gray et al. 1995; Segev et al. 2004). Furthermore, the exponent closely matched the value that was obtained by Gold et al. (2007).
Noise in extracellular measurements
Recording yield—the number of well-isolated single units—is affected by the presence of voltage noise. Electric potential fluctuations primarily arise from three sources: 1) in the amplifier electronics and external electromagnetic interference, 2) at the electrode–fluid interface, and 3) in the brain in the form of unsorted background activity. The combination of these processes is manifested as the total measured noise and is given by (2) For the measurements used to produce Fig. 4A, the total root mean square noise was found to be 15 ± 5 μV in the 300- to 5,000-Hz band. By comparing values with the amplifier initially shorted, and subsequently connected to electrodes immersed in locust saline, we determined that δVamp = 4 ± 1 μV, δVelectrode = 5 ± 1 μV, and by process of elimination, δVbrain = 14 ± 6 μV. Next we assume a reliable detection event requires the peak-to-peak amplitude of a spike to exceed approximately 5δVtot, which corresponds to about 75 μVpp on at least one channel. This estimate is in close agreement with the actual lowest sortable spike amplitude (Fig. 4B) of 68 μVpp. We found that only 16 out of a total of 73 single units recorded with the dual-side array satisfied the above-cited minimum amplitude criterion at least one channel on each side of the array. Thus the likelihood of simultaneously detecting the same unit on both sides was found to be 22%.
As noise in the system increases, the effective spike detection volume around an electrode is diminished. Using the best-fit parameters to Eq. 1, and 〈V0〉 = 204 μV, we estimate in Fig. 4C the average detection range per electrode (r) as a function of measurement noise. At the noise level of the antennal lobe, we find r to be about 40 μm. However, we note that differences in both δVbrain—which may reflect local inhomogeneities in firing activity—and extracellular fluid resistivity could give rise to significant variations in detection range. For example, noise has been found to differ by a factor of about 4 throughout the cat cortex (Buchwald and Grover 1970).
The most pronounced extracelluler current shielding effect between the front and back sides of dual-side arrays is likely to be manifested at the oxide and parylene insulation films; these highly resistive layers provide necessary electrical isolation for the recording channels. It is therefore unclear whether further reduction in the shaft's cross section would ameliorate the suggested shielding effect. An alternative approach may be, where possible, to target regions that manifest less measurement noise. For example, in the absence of noise processes contributing to δVbrain, we estimate the detection radius to be close to 140 μm. This upper bound is similar to an estimate obtained by a different method using multisite probe measurements in the rat hippocampus (Henze et al. 2000; Shoham et al. 2006). Under these highly optimal conditions the likelihood of simultaneously detecting spike activity from the same neuron on both sides of our dual-side array would increase to about 58%.
Double-layer electrode array
The 2 × 2 shaft microstructure, depicted in Fig. 1D, was designed to measure extracellular signals from an enclosed volume of tissue. This probe was tested successfully, as shown by the single-unit spike activity in Fig. 5A. The waveforms reveal an inhomogeneous spatial dependence of potential along the x–y and x–z sampling planes. A qualitative comparison of this observation, with models of extracellular potential around a spiking neuron (Gold et al. 2006), suggests extensive branching of neuronal processes. This appears to be consistent with the radial secondary dendritic geometry of antennal lobe projection neurons (Laurent and Naraghi 1994).
The 3D geometry of the double-layer array raises the prospect for localizing the source of extracellular fields within the enclosed space. Figure 5B displays the calculated center-of-field positions of two simultaneously recorded single units. A more detailed description of the analysis method is provided in the appendix. The majority of points in each cluster are concentrated in roughly spherical regions of about 10-μm diameter. This suggests that single units can be localized with a spatial resolution comparable to the cell body size. Moreover, the use of multilayer arrays in combination with more sophisticated analysis techniques could eventually enable tracking of signal propagation in the brain and facilitate the study of neuronal interactions in cell assemblies (Barthó et al. 2004; Blanche et al. 2005).
Spatial resolution of 3D recordings
Our results, although acquired from relatively simple structures, establish the feasibility of achieving dense extracellular measurements with three geometric degrees of freedom, using dual-side and multilayer electrode arrays. A notable advantage of our dual-side array design is that the electrodes on each side are connected to separately accessible recording channels. A different design (Perlin and Wise 2005) can accommodate a single channel per unit area on the shaft, which can be converted into either a front, back, or double-sided electrode via the use of through-hole etch techniques. Double-sided electrodes that simultaneously sample both sides of the array average the signal between the front- and back-facing sites; this makes signals susceptible to attenuation in the presence of asymmetric fields. On the other hand, although our approach requires twice as many channels per unit area, it is not susceptible to attenuation from averaging and enables up to a twofold higher sampling density. The fine-scale resolution provided by our dual-side arrays is viewed as an important feature for maximizing the recording yield from sparsely firing thin laminar structures in the brain. It is also intriguing to consider the possibility of using multilayer structures to perform dual stimulation and recording trials with a high degree of spatial precision. Such investigations may provide additional insight into single-cell level effects of electrical stimulation on brain function (McIntyre et al. 2004).
We envision merging the dual-side and multilayer device modalities and extending them toward more complex 3D microstructures aimed at large-scale recording applications. However, the limited number of input channels on the electronics imposes a trade-off between spatial sampling density and sampling volume, which can be tailored by adjusting the electrode and shaft spacing. The devices presented here were intended for fine-scale spatial resolution and small volume measurements, which may be most useful for probing single or few cell phenomena in great detail. A lower sampling density and larger volume may be ideal for exploring structure–function relationships in moderately sized structural units of neuronal ensembles, such as cortical columns (Mountcastle 1997). Such devices could also be used to map extracellular current source density in 3D to characterize information flow within and between multiple structures (Bragin et al. 1995). Finally, large, coarse-scale studies may be the most practical method for functional mapping of cubic millimeter level volumes and may present a complementary approach to functional magnetic resonance imaging, whose spatial resolution is limited to about 1 mm3 and temporal resolution does not enable single-spike detection (Logothetis et al. 2001). They may also hold promise as an alternative method to voltage-sensitive dye imaging (Mann et al. 2005), by enabling access to regions deep inside the brain without the need for using slices.
Figure 6A represents a cross section of the basic structural component of a proposed large scale 3D microelectrode array. The device consists of 8 × 8 shafts and features both dual-side and multilayer modalities. The width and thickness of the shafts are assumed to be equal (w) and the nearest-shaft spacing (d) is the same along the vertical and horizontal axes. The detection volumes surrounding the electrodes are approximated as hemispheres, whose radius r is equivalent to the spike-detection range estimated in Fig. 4C. Each shaft contains only one recording site on the front as well as the back side of each shaft and is therefore best suited for moderate to coarse spatial scale measurement applications. To increase the depth of tissue probed by the 3D structure, the unit module (length 2r) can be replicated along the longitudinal axis of the shafts. However, since all components are assumed to be identical, it is sufficient to consider a single module in the treatment that follows.
Ultimate limitations of 3D extracellular recordings
As the scale and complexity of the implanted structure increase, the risk of disrupting physiological brain activity is likely to rise. This suggests that two fundamental system design constraints are the amount of tissue damaged by the silicon microstructure and the volume accessed in the measurement. The former constraint requires small shafts and large shaft spacing, whereas the latter requires a large SNR and densely packed electrodes.
Using an intermediate value of 100 μm as the estimate for r, we predict in Fig. 6B the unique (i.e., nonoverlapping) fractional volume in the entire array that should lie within range of measurable spikes. The fractional recording volume increases with smaller shaft separation until the condition d ≤ r is satisfied, at which point significant overlap occurs between adjacent patches. Figure 6C shows the percentage of tissue displaced by the implant as a function of shaft spacing. The actual level of functional disruption may be considerably more complex than what is described by a simple volumetric estimate (Claverol-Tinture and Nádasdy 2004; Polikov et al. 2005); however, the fractional displacement volume provides a lower bound on the proportion of damaged cells expected to occur in the implanted region.
The preceding discussion on design limitations can help estimate the recording yield of the next generation of 3D microstructures. As an example, we consider an 8 × 8 shaft device in which the structural dimensions are reduced to w = 25 μm (Najafi et al. 1990) through the use of nanofabrication techniques. A shaft spacing of d = 200 μm would displace 2% of the structure's volume, which is similar to the value of some other implanted systems (Bai et al. 2000; Campbell et al. 1991). Assuming r = 100 μm, a 1-mm-long device would use five modular sections (since each module is 200 μm long), bringing the total number of required data channels to 640. About 40% of the 1.8 × 1.8 × 1-mm3 enclosed volume would lie within the detection radius of an electrode. Based on previous investigations, one typically detects only 1 to 10% of total available neurons in a selected region (Shoham et al. 2006). Thus given a neuronal density of 50,000/mm3 we would expect to measure between 650 and 6,500 units in 3.2 mm3 of cortex. It thus appears that devices using dual-side and multilayer arrays could enable an unprecedented scale of recordings in a spatially confined region. Future work needs to address the challenges of ensuring a minimally disruptive interface of 3D silicon microstructures with the brain.
Electrical cross talk and signal-attenuation effects
Stray capacitive coupling is one of the factors affecting the performance of extracellular recordings. To characterize these effects on the fabricated microelectrode arrays, a model was developed using a general purpose analog electronic circuit simulator (SPICE). The equivalent circuit is depicted in Fig. A1. The model is qualitatively similar to that used by Najafi et al. (1990), but here we distinguish between two probe configurations: 1) cross talk between adjacent traces on the same side of a shaft and 2) cross talk between opposing traces on a dual-side or double-layer structure. It is assumed that the conductive Si device substrate is grounded when inserted in the extracellular milieu. In Fig. A1 we show the percentage of cross talk (Vout,2/Vout,1) and signal transmission (Vout,1/Vin) as a function of equivalent electrode impedance. The results predict that with an electrode impedance of 0.25 MΩ, cross talk should be ≤0.6% in both configurations. Importantly, this value is insufficient to produce spurious spike-detection events. Moreover, capacitive shunting of the signal transmission path to the extracellular fluid gives rise to an attenuation of roughly 3%, the majority of which is due to the input capacitance at the headstage. Note that signal attenuation in configurations 1 and 2 is the same to within about 0.5% for electrode impedance <1 MΩ, and we present only the worst case values, which correspond to configuration 2.
The relevant circuit parameters were defined as follows:
Vin is the simulated extracellular signal, a 1-kHz sine wave.
Vout is the simulated signal as measured after the preamplifier input.
Rs, the spreading resistance encountered by current spreading out into fluid in the electrode vicinity (Franks et al. 2005), was calculated to be about 10 kΩ.
Relectrode and Celectrode are the equivalent resistive and capacitive components of the electrode impedance, respectively. These parameters were varied such that the net parallel electrode impedance ranged from 10 kΩ to 5 MΩ.
Rma, the resistance of the metallic traces between the recording site and the input to the preamplifier, was measured to be about 500 Ω.
Csub, the capacitive coupling to the extracellular fluid through the parylene C and thermal oxide layers, was estimated to be ≤2.7 pF.
Ccp,ij is the coupling on the silicon probe between channels i and j. In the case of adjacent traces on the same side of a shaft (Configuration 1), Ccp,ij = 0.1 pF. In the case of traces on either opposite sides of a dual-side shaft, or opposite layers of a double-layer shaft (Configuration 2), Ccp,ij ≤2.7 pF.
Ccf.,ij is the coupling on the flexible cable between channels i and j. This was estimated to be ≤1 pF.
Camp, the input capacitance at the headstage, equals 12 pF.
3D spike source localization
To show spatial separation of spikes with the 3D probe, we first carried out tetrode-based spike sorting to identify the single-unit clusters. Treating each cluster separately, we then determined the center of field position for all spike events via the expression (A1) where u⃗i are the coordinates of the ith electrode and Aji is the peak-to-peak amplitude observed for spike j at electrode i (Nádasdy et al. 1998).
This work was supported by a Broad Fellows Program in Brain Circuitry grant to J. Du and S. C. Masmanidis, Della Martin Foundation and Beckman Foundation grants to I. H. Riedel-Kruse, and a National Institute on Deafness and Other Communication Disorders grant to G. Laurent.
We thank C. Koch for advice and support; S. Cassenaer, Z. Nádasdy, U. Rutishauser, T. Siapas, M. Vähäsöyrinki, and H. Xu for discussions and technical assistance; and M. Walsh and T. Heitzman for electronic hardware. Device fabrication was carried out at the Kavli Nanoscience Institute, Caltech, and the National Nanotechnology Infrastructure Network, University of California, Santa Barbara.
The costs of publication of this article were defrayed in part by the payment of page charges. The article must therefore be hereby marked “advertisement” in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
- Copyright © 2009 the American Physiological Society
- Bai et al. 2000.↵
- Barthó et al. 2004.↵
- Blanche et al. 2005.↵
- Bragin et al. 1995.↵
- Buchwald and Grover 1970.↵
- Buzsáki 2004.↵
- Campbell et al. 1991.↵
- Claverol-Tinture and Nádasdy 2004.↵
- Csicsvari et al. 2003.↵
- Franks et al. 2005.↵
- Gray et al. 1995.↵
- Gold et al. 2007.↵
- Gold et al. 2006.↵
- Henze et al. 2000.↵
- Hoogerwerf and Wise 1994.↵
- Laurent and Davidowitz 1994.↵
- Laurent and Naraghi 1994.↵
- Logothetis et al. 2001.↵
- Mann et al. 2005.↵
- Mazor and Laurent 2005.↵
- McIntyre et al. 2004.↵
- Moffitt and McIntyre 2005.↵
- Mountcastle 1997.↵
- Nádasdy et al. 1998.↵
- Najafi et al. 1990.↵
- Najafi et al. 1985.↵
- Neves et al. 2007.↵
- Nicolelis et al. 1997.↵
- Olsson et al. 2005.↵
- Pang et al. 2005.↵
- Perez-Orive et al. 2002.↵
- Perlin and Wise 2004.↵
- Polikov et al. 2005.↵
- Segev et al. 2004.↵
- Shoham et al. 2006.↵
- Wise et al. 2008.↵
- Yao et al. 2007.↵