Brain-Controlled Interfaces: Movement Restoration with Neural ProstheticsRecording TechnologyEEGECoGLFPsSingle UnitsElectrode-Tissue InterfaceBrain Tissue Response to Implanted ElectrodesExtraction AlgorithmsSomatosensationSensory Recording InterfacesSensory Input InterfacesEffectorsBCI ProgressIntracranial Human StudiesConclusionAcknowledgmentsReferencesNeuron 52, 205–220, October 5, 2006 ª2006 Elsevier Inc. DOI 10.1016/j.neuron.2006.09.019ReviewBrain-Controlled Interfaces:Movement Restorationwith Neural ProstheticsAndrew B. Schwartz,1,*X. Tracy Cui,2Douglas J. Weber,3and Daniel W. Moran41Departments of Neurobiology and BioengineeringCenter for the Neural Basis of CognitionMcGowan Institute for Regenerative Medicine2Department of BioengineeringMcGowan Institute for Regenerative Medicine3Departments of Physical Medicine and Rehabilitationand BioengineeringThe University of PittsburghPittsburgh, Pennsylvania 152134Departments of Biomechanical Engineeringand NeurobiologyWashington UniversitySt. Louis, Missouri 63130Brain-controlled interfaces are devices that capturebrain transmissions involved in a subject’s intentionto act, with the potential to restore communicationand movement to those who are immobilized. Currentdevices record electrical activity from the scalp, on thesurface of the brain, and within the cerebral cortex.These signals are being translated to command sig-nals driving prosthetic limbs and computer displays.Somatosensory feedback is being added to this con-trol as generated behaviors become more complex.New technology to engineer the tissue-electrode inter-face, electrode design, and extraction algorithms totransform the recorded signal to movement will helptranslate exciting laboratory demonstrations to pa-tient practice in the near future.From the rapid growth in biotechnology, neural engi-neering has emerged as a new field. The merger of sys-tems neurophysiology and engineering has resulted inapproaches to link brain activity with man-made devicesto replace lost sensory and motor function. The excite-ment in this field is based not only on the prospect ofhelping a wide range of patients with neural disorders,but also on the certainty that this new technology willmake it possible to gain scientific insight into the waypopulations of neurons interact in the complex, distrib-uted systems that generate behavior. This review willaddress recent progress in cortical motor prosthetics.Related reviews are also available (Schwartz, 2004;Wilson et al., 2006; Lebedev and Nicolelis, 2006;Leuthardt et al., 2006).Neural prosthetics are devices that link machines tothe nervous system for the purpose of restoring lostfunction. Two broad approaches are used in this field:neurons are stimulated or inhibited by applied current,or their activity is recorded to intercept motor intention.Stimulation can be used for its therapeutic efficacy, as indeep brain stimulation to ameliorate the symptoms ofParkinson’s disease or to communicate input to thenervous system (for example by transforming sound toneural input with cochlear prosthetics). In contrast,recordings are used to decode ongoing activity for useas a command or input signal to an external device.Capturing motor intention and executing the desiredmovement form the basis of brain-controlled interfaces(BCI), a subset of neural prosthetics used to decodeintention in order to restore motor ability or communica-tion to impaired individuals.Every BCI has four broad components: recording ofneural activity; extraction of the intended action fromthat activity; generation of the desired action with aprosthetic effector; and feedback, either through intactsensation, such as vision, or generated and applied bythe prosthetic device (Figure 1).Recording TechnologyThe first step in the BCI process is to capture signalscontaining information about the subject’s intendedmovement. While researchers have envisioned usingmethods based on either magnetic (Georgopouloset al., 2005) or electromagnetic (Weiskopf et al., 2004;Yoo et al., 2004) signals from the brain, these devicesare not yet practical for BCI use. Currently, the fourprimary recording modalities are electroencephalog-raphy (EEG), electrocorticography (ECoG), local fieldpotentials (LFPs), and single-neuron action potentialrecordings (single units). All of these methods recordmicrovolt-level extracellular potentials generated byneurons in the cortical layers. The methods are classi-fied by whether the electrodes are placed on the scalp,dura, cortical surface, or in the parenchyma, and bythe spatial and spectral frequency of their recordedsignals. Generally, there is a tradeoff between these pa-rameters; the more invasive the recording technique, thehigher the spatial/spectral frequency content of the re-corded signal which, in turn, depends on the currentdensities conducted through the volume of the head.The primary current sources and sinks, i.e., where cur-rent enters the cell and leaves the cell, respectively,are synapses (both excitatory and inhibitory) and thevoltage-sensitive gates underlying neuronal action po-tentials. Because most nonspherical neurons are ori-ented radially, these currents approximate a dipolesource, which contains both equal and opposite polari-ties, oriented perpendicular to the cortical surface.Taken as a whole, the cortex can be modeled as athin, convoluted sheet of aligned dipoles whose individ-ual magnitudes vary continuously in time. BCI recordingaims to sample this dipole sheet and extract the desiredcontrol signal.From a purely engineering point of view, the optimalmethod of recording this electrical information wouldbe to place a series of small electrodes directly intothe dipole sheet to intercept signals from individual neu-rons (single-unit BCI designs). The ability of a microelec-trode to record single-unit action potentials depends onmany factors, such as electrode impedance, tip size andshape, whether the target cell has an open or closed ex-tracellular field, and the size and orientation of the targetneuron. Layer V cells in the motor cortex have the largest*Correspondence: [email protected] bodies in the cerebrum (>100 mm) and generate largeelectrical fields, making them an ideal source for extra-cellular recording. Multisite silicon probes can recorddistinguishable spikes from layer V neurons in rat senso-rimotor cortex located more than 300 mm away in the ax-ial direction (Buzsaki and Kandel, 1998), although thisdistance is likely