- neural prosthetics
- brain–computer interfaces
- neural interface technologies
- microelectrode array
- cortical recording
Technological advances in neural interfaces are providing increasingly more powerful “toolkits” of designs, materials, components, and integrated devices for establishing high-fidelity chronic neural interfaces. For a broad class of neuroscience studies, the primary requirements of these interfaces include recording and/or stimulating from a number of discretely sampled volumes at requisite spatial resolutions for specific periods of time. Translational and clinical applications present additional requirements for safety, usability, reliability, patient acceptance, and cost effectiveness. Innovative solutions result from the constructive tension between ever-increasing application requirements and incorporation of technological advances into usable devices. The purpose of this minireview is to present snapshots of the current state-of-the-art in chronic, microscale neural interfaces by highlighting several leading neuroscience applications and discussing their implications for next-generation interface devices.
Advances in neural interface technologies
Neural interface devices have a large design space in terms of size, shape, and function. Distinct and diverse neural targets prohibit a one-size-fits-all approach. Considering implantable, chronic microelectrode arrays for recording, which typically have the most rigorous requirements for selectivity, sensitivity, precision, and longevity, three of the principal design considerations are (1) spatial and temporal resolution of desired signals (e.g., synaptic currents, single units, and field potentials), (2) number and placement of the electrode sites, and (3) functional lifetime of the device. A growing number of wire and microfabrication electrode technologies intended to meet these requirements are in various stages of development (McNaughton et al., 1983; Nordhausen et al., 1996; Hetke and Anderson, 2002; Nicolelis et al., 2003; Wise et al., 2004; Motta and Judy, 2005; Rennaker et al., 2005; McCreery et al., 2006; Musallam et al., 2007; Neves and Ruther, 2007; Bartels et al., 2008). As a group, these devices have enabled successive advances in highly parallel recording of signals from synaptic currents to single units to local fields through systematic refinements of materials, fabrication processes, packaging, and usage. Wire bundles and arrays are the simpler technology and are the most widespread type of implantable electrode array.
Microfabricated electrode arrays (neural probes) formed by lithographic patterning of thin films of conductors and insulators on silicon or polymer substrates involve more complex technologies, but have the benefit of accessing a larger design space. Microfabricated probes can be readily customized to meet specific experiment requirements. In a silicon technology (Hetke and Anderson, 2002), for example, the probe can have almost any two-dimensional shape with single or multiple shanks (Fig. 1A,B). Electrode sites can be realized from a variety of materials, can have a range of surface areas, and can be placed anywhere along the shank(s) at any spacing; tips can be made very sharp or blunt; and features such as holes and barbs can be included. The high-quality, commercial-grade microfabricated probes that are available today are the result of steady advances in design rules, lithography, substrates, thin-film organic and inorganic dielectric materials, metal deposition, packaging, and quality control.
Proof of concept for obtaining high-quality unit recordings for more than several months has been established through independent studies from several labs that collectively cover diverse types of microelectrode arrays implanted in rat, cat, and nonhuman primate (Rousche and Normann, 1998; Liu et al., 1999; Kipke et al., 2003; Nicolelis et al., 2003; McCreery et al., 2004; Suner et al., 2005; Jackson and Fetz, 2007). However, there is a large degree of variability and unpredictability in chronic performance that results from an incomplete understanding of the failure (and success) modes of implantable microscale devices. At present, the state of the art for implantable arrays is that recording quality typically degrades and uniformly fails over time (lifetimes ranging from several weeks to several months). Scientific advances in understanding the chronic electrode–tissue interface are fueling innovative new approaches for longer-lasting electrode arrays. For example, advanced probe architectures (Fig. 1D) (Seymour and Kipke, 2007), materials (Ludwig et al., 2006; Abidian and Martin, 2008), approaches to mitigate chronic reactive tissue responses (Shain et al., 2003; Spataro et al., 2005), and minimally damaging insertion techniques (Bjornsson et al., 2006; Johnson et al., 2007) are being investigated to create longer-lasting interfaces for high-fidelity recording and stimulation. The engineering strategy is to use these and related advances to develop a design space to systematically result in progressively better implantable devices for chronic neural interfaces (e.g., consistent and measurable performance improvements with successive device versions). This rational design of neural interfaces requires standard protocols and performance assessments that are currently being developed.
Large-scale recording of neuronal and field activity
Studying neurons and neuronal pools requires simultaneous monitoring of the activity of large numbers of individual neurons in multiple brain regions. A major goal therefore is to record from statistically representative samples of identified neurons from several local areas while minimally interfering with brain activity. The technical challenge is to position individual or small sets of recording sites sufficiently close (typically tens to several hundred micrometers) to the target neurons or neuronal segments. Advances in microelectrode technologies increase the numbers and precise positioning of electrode sites, and thus, enable progressively more powerful experimental approaches to understanding the interaction of neurons and neuronal pools.
This type of large-scale recording of neuronal activity involves three critical components: (1) spike sorting methods, (2) identification of neuronal types, and (3) characterization of synaptic inputs. Currently, wires and microfabricated silicon electrode arrays can record from large numbers of neurons and monitor local neural circuits at work (Fig. 1E) (McNaughton et al., 1983; Hoffman and McNaughton, 2002; Csicsvari et al., 2003; Barthó et al., 2004; Buzsáki, 2004).
Isolation and identification of neurons by extracellular signatures
An indispensable step in spike train analysis is the isolation of single neurons on the basis of features in extracellular recordings. Spike sorting methods fall into two broad classes. The first class attempts to separate spikes on the basis of amplitude and waveform variation; the second method separates units on the basis of their spatial location (Wilson and McNaughton, 1993; Buzsáki, 2004). Unfortunately, none of these methods is uniformly satisfactory. When neurons are strongly activated, their amplitude, rise time, decay time and spike width change as a function of the membrane potential before the spike and as a function of the immediate firing history of the neurons. The spike amplitude variation is most substantial during complex spike burst production, with as much as 80% amplitude reduction, primarily because of Na+ channel inactivation (Harris et al., 2001). Therefore, amplitude and waveform-based classification programs will separate action potentials from a single neuron as if they were emitted by many. The amplitude and waveform variability of the extracellularly recorded spike is the major cause of unit isolation errors (Harris et al., 2000; Henze et al., 2000). Additionally, in long-term chronic recordings using implantable microelectrode arrays, unit isolation issues are typically, but not always, compounded by a progressive degradation in signal quality (e.g., ratio of spike amplitude to background noise) for implant lifetimes more than several months.
The spatial triangulation approach is based on the tacit assumption that the extracellularly recorded spikes emanate from point sources (McNaughton et al., 1983). However, every part of the neuronal membrane is capable of generating action potentials and what is observed in the extracellular space is a result of the complex geometry of neurons and the propagation of spike within the neuron. The extent of the somatodendritic backpropagation of the action potential varies as a function of the excitatory and inhibitory inputs impinging on the neuron and the state of various other conductances. The spike recorded in the extracellular field is a summation of the integrated signals from both soma and large proximal dendrites. Thus, the extent of somatodendritic spike backpropagation can affect the estimation of the neuron's virtual “point source” location and may place the same neuron at different locations, resulting in omission errors of unit isolation (Harris et al., 2000). A further problem with the point-source assumption for action potentials is that the somatic origin is not always resolvable with distant recording sites. For example, in the rat neocortex, extracellular spikes can be recorded from the apical shaft of layer V pyramidal neurons as far as 500 μm from the cell body (Buzsáki and Kandel, 1998). As a consequence, electrodes placed, for example, in layer IV, can equally record from layer IV cell bodies or apical dendrites of deeper neurons. Large neurons in the primate cortex are especially prone to such misclassification errors because their spikes can be recorded several hundred micrometers from the cell bodies.
To a large extent, these sources of unit separation errors can be reduced by recording at multiple sites parallel with the axodendritic axis of the neurons. This can be achieved, for example, by using a silicon probe with a shank positioned along the axodendritic axis and with small electrode sites precisely clustered along the shank.
Classification of neuron types
Isolated single units need to be classified into known cell groups of the cortex. This is a very important requirement because cortical networks consist of several neuronal classes, each with a specific computation task (Freund and Buzsáki, 1996; Klausberger and Somogyi, 2008). The potential conclusion from an experiment reporting that all cortical pyramidal cells did one thing and all interneurons something else is qualitatively different from the conclusion that can be drawn from the information that 80% of all (unclassified) cells behaved differently from the rest.
Unfortunately, current knowledge about the extracellular features of neurons in the neocortex severely limits the reliable separation of principal cells of different layers and the numerous interneuron classes. Spike duration criterion has been used to distinguish fast spiking interneurons from pyramidal cells in the rodent cortex (Mitchell et al., 2007). Simultaneous recording of local neuron pairs can provide physiological clues about the excitatory or inhibitory nature of the recorded units because their cross-correlation can display robust, short-time-scale correlations, indicative of monosynaptic excitatory or inhibitory connections (Fig. 1F). Because such synaptically connected pairs are rare (Barthó et al., 2004; Fujisawa et al., 2008), large-scale recording of neuronal populations is needed for their identification.
Characterization of synaptic inputs
A necessary requirement for a more complete understanding of the transformation of inputs by a neuron or neuronal assemblies is information about both their input and output. Unfortunately, no method is available for monitoring all inputs at the resolution of dendrites and spines of single neurons. However, membrane currents generated by neurons pass through the extracellular space. These currents can be measured by electrodes placed outside the neurons. The field potential (i. e., local mean field), recorded at any given site, reflects the linear sum of numerous overlapping fields generated by current sources (current from the intracellular space to the extracellular space) and sinks (current from the extracellular space to the intracellular space) distributed along multiple cells. Local field potential measurements combined with recording of neuronal discharges is the best experimental tool available to study the influence of cytoarchitectural properties, such as cortical lamination, distribution, size, and network connectivity of neural elements on electrogenesis. However, large number of observation points combined with decreased distance between the recording sites are required for high spatial resolution and for making interpretation of the underlying cellular events possible. Recording the voltage gradients by geometrically arranged sites of silicon probes within a target allow current densities to be calculated for an estimation of the mean input to the neuron group in the recorded volume (Nadasdy et al., 1998).
Large-scale recording of neuronal and field activity is continuing to improve through systematic advances in the underlying neural interface technologies that provide the ability to tune electrode arrays to better match neuron morphologies and distributions in the targeted volume and to have optimized signal recording characteristics. This involves engineering details such as smaller feature sizes to support higher site density and more precise site layouts. It also involves high-level aspects such as packaging and quality control to support more component-level integration in a cost-effective manner. This expanded design space is illustrated in a next-generation three-dimensional microelectrode array that provides a modular and customizable way to integrate several one- or two-dimensional planar probes into a three-dimensional assembly to more precisely and fully access a target volume (Fig. 1C). With more device complexity and increased numbers of sites, getting signals from or to the microelectrodes becomes a significant issue, especially in awake and behaving preparations and brain–computer applications. Although physical connectors are the current de facto standard, there is notable progress in developing low-power, low-noise electronic interfaces with wireless interfaces (Neihart and Harrison, 2005; Song et al., 2005, 2007; Oweiss, 2006; Ghovanloo and Najafi, 2007; Sodagar et al., 2007; Lee et al., 2008).
Engaging the brain: recurrent brain–computer interfaces
Brain–computer interface (BCI) studies are providing new insights into accessing cortical circuits through chronic, implantable neural interfaces and engaging them in various types of real-time, closed-loop experimental paradigms. Most BCI systems typically allow neural activity to control computer cursors or external devices (Wolpaw et al., 2002; Carmena et al., 2003; Hochberg et al., 2006; Velliste et al., 2008). Recently, the BCI concept was extended to create a recurrent BCI (R-BCI) system in which neural activity is recorded and processed in real time to control electrical stimulation of particular areas of the brain or muscles through implanted electrodes. Such R-BCIs implement an artificial connection that the adaptive brain could learn to incorporate into normal function. This autonomous recurrent BCI paradigm opens new experimental directions and has promise for several clinical applications. R-BCI systems are intricately related to neural interface technologies because system performance depends not only on neural processing, but also on the performance of the two separate (recording and stimulation) interfaces and the embedded real-time computational unit.
Initial R-BCIs were developed using a “Neurochip” containing autonomously operating circuitry that interacts continuously with the brain of a monkey. The system consisted of a printed circuit board populated with off-the-shelf components and connected to wire electrodes that record the activity of motor cortex cells and/or muscles (Mavoori et al., 2005; Jackson et al., 2006b). The amplified activity was digitized and processed by a programmable computer chip and stored to on-board memory and/or converted in real time to deliver activity-contingent electrical stimuli back to the cortex (Fig. 2A,B). Chronic implantation of the battery-powered Neurochip on the skull allowed continuous operation during free behavior and sleep (Jackson et al., 2007).
Artificial recurrent feedback connections
Clinically, R-BCIs could bridge impaired biological connections and allow the subject to learn to control this prosthetic connection to compensate for lost pathways (Jackson et al., 2006b). Current experiments are testing the ability of nonhuman primates to use motor cortical cell activity to control functional electrical stimulation (FES) of paralyzed muscles (Moritz et al., 2008). In these experiments, recurrent connections, largely via rack-mounted instrumentation, were used to implement direct control of muscle FES during peripheral nerve block. The monkeys' accuracy in using the R-BCI to acquire torque targets improved with practice. Bidirectional movements via independent FES of antagonist muscle groups were controlled by either high and low firing rates of single cells or by modulating pairs of cells with each neuron controlling stimulation of a different muscle group. Surprisingly, neurons could control goal-directed FES equally well regardless of whether they were originally modulated with movement, thus demonstrating an increased source of control signals for brain–computer interfaces. These results demonstrate that direct artificial connections between single cortical cells and muscles can be used to compensate for interrupted physiological pathways and restore volitional goal-directed control of movement to paralyzed limbs. The degree to which prolonged implementation of these artificial connections will allow the brain to integrate them into normal movements remains to be tested.
Long-term modification of synaptic connections through Hebbian mechanisms
By delivering stimuli synchronized with cell activity, continuous operation of the R-BCI has produced significant long-lasting changes in neural connections in motor cortex (Jackson et al., 2006a). The Neurochip converted the action potentials of a single motor cortex neuron to stimuli delivered within a few milliseconds to a neighboring cortical site. Operating continuously during 24 h of free behavior and sleep, this conditioning regimen led to subsequent long-lasting changes in the movement evoked from repetitive microstimulation of the recording site, bringing its output closer to the movements evoked from the stimulation site (Fig. 2C). The simplest mechanism mediating this change is Hebbian strengthening of the physiological connections from the recording site to the synchronously stimulated site. Remarkably, the changes in cortical output remained stable for at least 10 d after the end of conditioning, indicating a potent long-lasting effect. This R-BCI-induced plasticity opens opportunities for many basic experiments into the underlying mechanisms and replication at other sites. The paradigm also has potential therapeutic application to strengthen weak biological connections. For example, activity-dependent cortical stimulation could facilitate recovery from stroke or traumatic brain injury.
The R-BCI paradigm has numerous possible applications, depending on the input signals [neuronal action potentials, field potentials, electrocorticographic (ECoG), or EMG], and the output sites (cortex, spinal cord, cerebellum, muscles, or reinforcement sites), and the transform between them (direct conversion to proportional stimulation, computed functions of detected activity, simulated neural networks). Ultimately, successful incorporation of these R-BCIs into brain function will depend on the brain's ability to exploit and adapt to consistent contingencies. Looking ahead, the possibility of transforming spatiotemporal patterns of multichannel neural recordings to patterns of multichannel stimulation may make it possible to implement recurrent computations in higher-order cognitive areas of the brain, like hippocampus, conceivably producing a “cognitive prosthesis” (Berger and Glanzmann, 2005; Berger et al., 2005).
Further developments in implantable neural interface technologies will enable more sophisticated R-BCIs through larger-scale, higher-fidelity, and longer-lasting microelectrode arrays. Additionally, sophisticated, low-power integrated circuits for signal conditioning and embedded computing and better packaging systems for implantable electronics will drive the development of next-generation implantable integrated microsystems that have increased real-time computing power (Fig. 2D).
Translating microscale neural interface technologies for clinical use
Although there are significant research applications for advanced neural interface technologies, the ultimate application of these technologies is in understanding human nervous system function and in restoring lost function to patients. Neural interface technology has been applied to humans mainly on the input side, with electrical stimulation of the spinal cord and brain being practiced for nearly four decades. Since the first descriptions of implantable neurostimulators in the 1950s and 1960s (Shealy et al., 1967), neural interface systems have steadily become more sophisticated and more robust. However, the translation of research-grade devices to clinical-grade devices takes a significant engineering development effort to make the devices and their components amenable for use in the rugged neurosurgical environment and safe and effective in their intended application.
Significant technical issues can be expected with any device designed to withstand the harsh environment of the human body for years. Failure rates for implantable stimulation devices can exceed 40% (Rosenow et al., 2006). Engineering analysis of mechanical failures of these devices can provide insights to appropriate surgical techniques to minimize mechanical stresses and strains on the system (Henderson et al., 2006).
Recently, cortical neural interfacing with microscale devices has been demonstrated in humans (Hochberg et al., 2006). Although it represents a remarkable achievement, this system is at a stage similar to the first spinal cord stimulator systems, using percutaneous connectors with all of their disadvantages including infection risks, susceptibility to damage, etc. Next-generation microscale neural interface systems will need to approach the safety, durability, and ease of use of current clinical systems such as those used for deep brain stimulation (DBS), including wireless interfaces.
DBS has become a standard treatment for Parkinson's disease, with good results over many years (Krack et al., 2003). The interface to the brain, relatively large compared with microelectrode arrays, consists of a 1.27-mm-diameter probe with four cylindrical contacts 1.5 mm in length, spaced 0.5 mm apart. The electrode contacts are attached via robust coiled wire extensions to a titanium-encased pulse generator implanted in the chest. Another system consists of similar electrodes but with the pulse generator implanted in the skull (Fountas et al., 2005). Both designs have proven to be robust in clinical practice, with the latter system incorporating both sensing and stimulation capabilities.
It is thus technically feasible to create clinically useful interfaces that can sense and stimulate both the cortical surface and deep nuclei. New systems must take into account the challenges of the operating room environment. Implantable devices must be robust enough to allow rough handling by neurosurgeons. Time is of the essence in the operating room, so delivery systems must be simple and quickly deployable. Connectors should be reliable, fluid-proof, and as small as possible, a challenge when the number of connections potentially reaches into the hundreds. Fully developed systems for human use will require not only well engineered devices, but also carefully developed plans for packaging and delivery.
The microscale cortical recording array used in the Cyberkinetics human trial (Hochberg et al., 2006) provides an example of how next-generation packaging and delivery could promote more widespread adoption. This device has been implanted with good success using a wide cortical exposure via craniotomy. By engineering a delivery system that eliminates the need for extensive visualization, these arrays could be implanted in a minimally invasive manner similar to that used for DBS. This could, for example, take the form of a prepackaged “cartridge” with the pneumatic inserter attached, incorporating a pressure sensor that allows for precise positioning on the cortex and even possibly automated insertion. Such a system would allow easy implantation with a high degree of reproducibility.
Recordings from the surface of the brain
Invasive BCIs with implanted intracortical microelectrode arrays use local activity from neurons recorded within the brain (Georgopoulos et al., 1986; Serruya et al., 2002; Taylor et al., 2002; Shenoy et al., 2003; Andersen et al., 2004; Musallam et al., 2004; Lebedev et al., 2005; Hochberg et al., 2006; Santhanam et al., 2006; Lee et al., 2008; Velliste et al., 2008). Signals recorded within cortex have high fidelity, but the stability of intracortical recordings can be variable and decays with time (Santhanam et al., 2007).
Noninvasive BCIs use electroencephalographic (EEG) activity recorded from the scalp (Farwell and Donchin, 1988; Wolpaw et al., 1991; Sutter, 1992; McFarland et al., 1993, 2008; Pfurtscheller et al., 1993, 2000; Wolpaw and McFarland, 1994, 2004; Birbaumer et al., 1999; Millán et al., 2004; Kübler et al., 2005; Blankertz et al., 2006; Vaughan et al., 2006; Müller et al., 2008). Although EEG-based BCIs support higher performance than often assumed, the acquisition of high levels of brain-based control typically requires extensive user training, and BCI performance can also be variable. Despite recent encouraging evidence that BCI technologies using noninvasive or intracortical methods can be useful to people with paralysis [Kübler et al. (2005) and Hochberg et al. (2006), respectively], the difficulties described above currently impede widespread clinical implementation of BCI technologies (Schalk, 2008). ECoG recordings from the surface of the brain may combine the advantages of noninvasive and intracortical BCI approaches, and may provide the basis for powerful but yet practical BCI systems for the disabled.
It has long been known that ECoG signals recorded from the surface of the brain also hold information about movements, and that this information could be used for device control (Brindley and Craggs, 1972; Craggs, 1975). It has dramatically higher spatial resolution than EEG [i.e., tenths of millimeters vs centimeters (Freeman et al., 2003)], broader bandwidth (Staba et al., 2002), higher characteristic amplitude, and far less vulnerability to artifacts such as EMG (Freeman et al., 2003). At the same time, because ECoG does not require penetration of the cortex, it may have greater long-term stability than intracortical recordings (Loeb et al., 1977; Bullara et al., 1979; Yuen et al., 1987; Margalit et al., 2003).
Signal fidelity of ECoG
Initial evidence in animals (Mehring et al., 2003; Rickert et al., 2005) and more recent evidence from humans (Leuthardt et al., 2004; Schalk et al., 2007b) demonstrate that ECoG can give detailed information about kinematic parameters of hand movements (Fig. 3). Moreover, the fidelity of that information, i.e., the correlation between actual and decoded position/velocity of hand movements in humans, was shown to be within the range of the results that were previously reported for intracortical microelectrode recordings in nonhuman primates [see Schalk et al. (2007b), their Table 3]. These human results have since been replicated and further extended for decoding of arm movements (Pistohl et al., 2008; Sanchez et al., 2008), flexion of individual fingers (Kubanek et al., 2007), and even specific aspects (such as the type of phoneme) of actual or imagined speech (Schalk et al., 2007a). Finally, several studies demonstrated that it is possible to combine general understanding of motor-related ECoG responses (Crone et al., 1998a,b; Aoki et al., 1999; Graimann et al., 2002; Sinai et al., 2005; Leuthardt et al., 2007; Miller et al., 2007) with the more specific understanding described above to implement different ECoG-based BCIs (Leuthardt et al., 2004; Wilson et al., 2006; Felton et al., 2007; Schalk et al., 2008) in humans using the BCI2000 software framework (Schalk et al., 2004). These studies showed that ECoG supports accurate nonmuscular one- or two-dimensional movement control in humans with little subject training.
Advances in neural interface technologies intersect with increasingly sophisticated experimental approaches to create innovative approaches for investigating and engaging the brain. The accessible design space for neural interfaces is becoming sufficiently developed to enable enhanced devices that are tuned to meet aggressive application requirements. In the next 3–5 years, there is potential for significant progress along the emerging roadmap for neural interface technologies. Continued application-driven advances will increase the selectivity, sensitivity, precision, bandwidth, reliability, and functional lifetimes of implantable microelectrode arrays for neural recording and stimulation. Focused investigations will increase understanding and control of reactive tissue responses to chronically implanted neural interface devices. Interdisciplinary, engineering science developments will open up ways to integrate implantable microelectrode technologies with complementary types of neural interface technologies, including chemical sensing, drug delivery, optical imaging, and genetic manipulations. Development of standards in the design, assessment, and use of implantable neural interface devices will increase the pace of development and the effectiveness of the technologies and devices. And, there will be additional efforts to translate research-grade microscale neural interface devices to clinical devices for diagnostic and therapeutic applications.
The stage is set for increased interdisciplinary collaboration to drive creative interactions among neuroscience, neural interface technology, and clinical applications. The synergy created by these activities will push progress to new heights and usher in the realization of robust implantable devices for long-term study and treatment of the CNS.
D.R.K. has financial and leadership interests in NeuroNexus Technologies.
This work was supported by National Institutes of Health Grants EB 002030 (D.R.K., W.S.), NS12542 and RR00166 (E.F.), NS044287 (W.S.), NS060269 and NS054346 (J.H.), and EB006356 and EB000856 (G.S.), and by United States Army Research Office Grants W911NF-07-1-0415 and W911NF-08-1-0216 (G.S).
- Correspondence should be addressed to Daryl R. Kipke, Department of Biomedical Engineering, 2212 LBME, 1101 Beal Avenue, University of Michigan, Ann Arbor, MI 48109.