Abstract
Persons with movement and communication impairment due to neurological injury can have function restored through brain–computer interfaces (BCIs) combined with appropriate prosthetic end effectors. Invasive BCIs record cortical activity using penetrating electrodes, or electrodes on the brain surface, in relevant cortical areas. Intended actions are then decoded from these cortical signals to command various end effectors, including virtual typing communication devices, robotic limbs, and paralyzed limbs reanimated through functional electrical stimulation (FES). This chapter discusses components of invasive BCI systems, including the pros and cons of various recording technology, potential cortical areas of implantation, and signals of interest and how they are decoded into functional commands. Finally, the chapter discusses current and future applications of BCI systems, including FES systems for restoration of functional movement, and challenges and opportunities for translating this technology from the laboratory to clinical day-to-day use.
Keywords
Brain–computer interface, Electrocorticography, Functional electrical stimulation, Functional restoration, Local field potential, Microelectrode, Neural decoding, Spike sorting, Stereoelectroencephalography, Tetraplegia
Outline
Introduction 379
Recording Technologies Used in Invasive Brain–Computer Interfaces 380
Stereoencephalography Electrodes 380
Electrocorticography Electrodes 381
Penetrating Microelectrodes 381
Cortical Areas and Signals of Interest for Invasive Brain–Computer Interfaces 381
Neural Decoding 383
Current Applications of Invasive Brain–Computer Interfaces for Motor Restoration 385
Two-Dimensional Cursor Movementsand Virtual Typing 385
Robotic Limb Control 385
Restoration of Paralyzed Arm and Hand Movements 386
Current Challenges and Future Directions of Invasive Brain–Computer Interfaces 386
Electrode Longevity and Robustness 387
Cortical Signal Stability 387
Fully Implantable and Miniaturized Wireless Brain–Computer Interfaces 387
Restoring Natural Motor Function and Sensation 388
References 389
Introduction
Neurological and neuroanatomical injuries and disorders affect a large number of people worldwide, and often result in movement impairment and inability to perform everyday tasks, such as communicating, reaching, and grasping, independently. Persons who have experienced neurological injuries, such as spinal cord injury (SCI), amyotrophic lateral sclerosis, or stroke, can achieve partial restored function through cortical prosthetic systems. A cortical prosthesis is an end effector device that receives an action command to perform a desired function through a brain–computer interface (BCI) that records cortical activity and extracts (i.e., decodes) information related to that intended function. End effectors can range from virtual typing communication systems to robotic arms and hands or a person’s own limb reanimated by functional electrical stimulation (FES). BCI technology can range in levels of invasiveness, temporal and spatial recording resolution, and the types of recorded signals. Noninvasive brain imaging technologies, such as electroencephalography (EEG), magnetoencephalography (MEG), and functional magnetic resonance imaging (fMRI), have been implemented in various simple BCI applications, such as low throughout communication spelling systems ( ). However, these noninvasive BCI options typically are too slow (e.g., fMRI), have low spatial resolution and low signal bandwidth (e.g., EEG), and/or are easily corrupted by external artifacts ( ). Thus they are not ideal for complex real-time applications, such as high-performance communication, control of multidimensional robotic limbs, and reanimation of paralyzed limbs for coordinated reaching and grasping. Invasive BCIs, because of the higher resolution and signal bandwidths available, have the potential to allow people with neurological injury to command more high-dimensional systems naturally and restore more complex functions.
In this chapter we review currently available invasive recording technologies used for BCI neuroprosthetic systems. We examine areas of cortex in which these recording technologies are typically implanted, the types of signals recorded from these areas, and the movement-related information that can be extracted. We review the most relevant algorithms for extracting (i.e., decoding) movement-related information from these cortical signals, as well as practical concerns when developing cortical decoders. Additionally, we discuss current applications of invasive BCIs in conjunction with technology for restoring communication and movement. Finally, we discuss hurdles to clinical translation and daily adoption, and current research aimed at overcoming these hurdles.
Recording Technologies Used in Invasive Brain–Computer Interfaces
Invasive neural recording technologies are generally distinguished based upon the depth of cortical tissue penetration and the resolution of the recorded neural signals. Fig. 27.1 illustrates representative examples of the most common invasive neural recording technologies: stereoencephalography (sEEG) intracranial electrodes, electrocorticography (ECoG) sheet electrodes, and penetrating microelectrode arrays. It should be noted that there are different interpretations of the term “invasiveness.” One interpretation focuses on the depth of penetration into cortical tissue, in which case ECoG electrodes may be viewed as least invasive and sEEG may be viewed as most invasive. However, another interpretation focuses on the level of surgical intervention required, and more specifically the size of a craniotomy required for electrode implantation. In this interpretation, ECoG sheet electrodes may be viewed as the most invasive, due to the usually larger size of the recording sheets (and hence larger craniotomy), in comparison to penetrating microelectrodes and sEEG depth electrodes. Note that resection surgeries for persons with intractable epilepsy routinely require large craniotomies with relatively low rates of minor complication ( ). This section describes each recording technology, including neural features typically recorded, and current BCI implementations of each technology.
Stereoencephalography Electrodes
sEEG intracranial electrodes are long, thin, depth electrodes, similar to those used for deep brain stimulation, with large contacts spaced longitudinally along the shaft. Clinically, sEEG electrodes are routinely and predominantly used for presurgical seizure localization in persons with pharmacologically resistant epilepsy who are candidates for resection surgery. Surgical placement typically involves a small stab incision and burr hole for insertion of the electrode, rather than a larger craniotomy. Consequently, faster postoperative recovery, less tissue healing time, and smaller scars are observed in comparison to other invasive electrode technologies. Retrospective studies comparing complications of placement of sEEG intracranial versus ECoG subdural sheet electrode cases show that sEEG placement resulted in significantly fewer complication rates, and deem placement of sEEGs to be reliable and safe ( ). A major benefit of sEEGs is that they allow recording from potentially every cerebral structure, including difficult-to-access subsurface cortical areas and deep sulci walls where significant movement-related activity is routinely observed ( ). From these areas, sEEGs record field potential activity (0–200 Hz bandwidth), with observed modulation in specific spectral frequency bands (e.g., α: 2–12 Hz, β: 12–30 Hz, γ: 30–60 Hz) potentially conveying relevant movement information. Several recent studies have investigated the potential to use sEEG intracranial electrodes for BCI applications. reported sEEG electrodes recording α modulation from the insular cortex and broadband modulation from deep sulci walls of the motor cortex for prediction of hand grasp force. sEEG electrodes have also been used in standard two-dimensional BCI cursor control center-out tasks with moderate success ( ). Future research and electrode enhancements (such as high-resolution contacts for single neuron recordings) may enhance the performance of sEEG electrodes, potentially making them a very attractive option for invasive BCIs due to their already adopted clinical status.
Electrocorticography Electrodes
ECoG subdural grid electrodes and strips consist of series of flat electrode contacts placed on a thin silicone sheet that either sits on the brain surface directly under the dura or is potentially placed within a sulcus. Like sEEG intracranial electrodes, ECoG grids and strips are used clinically for seizure localization, though in recent years with less frequency due to the required larger craniotomy and higher complication rates. However, ECoG electrodes offer broader coverage of the cortical surface and recording of neuron population-level field potentials, and have been demonstrated for closed-loop control of simple cursor movements ( ) and even control of a three-dimensional robotic arm ( ). Proponents of ECoG-based BCIs argue that the recorded field potential signals may have greater temporal stability, and hence greater clinical viability, than single neuron signals. However, this claim has not been rigorously validated.
Penetrating Microelectrodes
Penetrating microelectrode arrays consist of a number of very small wires or shafts that penetrate into cortical tissue to allow recording of single neuron action potentials and local field potentials (LFPs). Currently the predominant penetrating microelectrode technology is the 96-channel microelectrode array (BlackRock Microsystems, Salt Lake City, UT) ( ). Ninety-six recording electrodes insulated with parylene-C are attached to a silicone platform substrate, which is then connected to a percutaneous pedestal through a wire bundle. The microelectrode shanks typically penetrate 1.0–1.5 mm into cortical tissue. Though only 4 × 4 mm in size, surgical implantation usually requires a larger craniotomy to ensure that the array shanks do not rupture cortical vasculature during placement. Implantation is usually performed using a pneumatic insertion technique ( ). These arrays have demonstrated efficacy in both nonhuman primate (NHP) and human invasive BCI applications, and are the focus of most of this chapter. Other penetrating electrode approaches have been widely used in preclinical studies; specifically, “microwires” and “Michigan probes” have been used successfully, each with unique advantages and disadvantages. Microwires are the lowest-profile penetrating electrodes (i.e., evoke the smallest immune response), and are promising for chronic multisite recordings ( ). Michigan probes have been developed in many formats, but typically include multiple electrode surfaces spaced longitudinally along the shaft and allow incorporation of active electronics ( ). However, these electrodes have some challenges in regards to mechanical stability ( ), and are not yet proven in chronic human applications.
Cortical Areas and Signals of Interest for Invasive Brain–Computer Interfaces
Invasive BCIs have the advantage over their noninvasive counterparts (EEG, MEG, fMRI) in that they allow high temporal and spatial resolution recordings of single neuron action potentials (“spikes”) and LFPs. The spikes and LFPs from recorded small cortical networks can often be directly related to specific aspects of movement. Several motor cortical areas have been investigated as possible recording sites for providing information relevant to decoding natural arm and hand movements, or BCI-commanded target-directed movements. This section presents a brief synopsis of the different types of movement information encoded in various cortical areas of interest ( Fig. 27.2 ), and the various cortical signal features recorded and used in invasive BCIs.
Cortical Areas for Recording and Stimulation
Early investigations of cortical representations of movement focused on demonstrating that single neuronal activities in primary motor cortices (M1) encoded various kinematic (e.g., movement velocity) ( ) and kinetic (e.g., muscle force) ( ) properties of arm and hand movements. These efforts eventually led to the groundbreaking work of Donoghue et al., which demonstrated that NHPs could substitute neural signals from M1 in place of arm movements to command external devices ( ). M1 cortical activities can be considered as the outputs of a multistep motor planning process beginning in the prefrontal cortical areas, and involving posterior parietal, supplementary motor, premotor, and primary sensory cortical areas. fMRI and invasive electrophysiological investigations of the prefrontal cortex show increased cortical activity during mental processing involving working memory and planning tasks, potentially useful for issuing BCI commands for cognitive neuroprostheses ( ). The posterior parietal cortex has also been reported to modulate with respect to planning parameters of visually guided reaching movements, such as the planned time of a targeted reach and the relative path relationships between a desired target, eye-gaze position, hand position, head orientation, and target spatial depth ( ). Similarly, cortical networks in supplemental motor ( ) and premotor areas ( ) have also shown tuning in regard to timing, planning, and sequencing of upcoming reaching movements, as well as relative positioning between gaze direction, hand position, and target position. Premovement planning information derived from each of these cortical areas could be used in conjunction with decoded point-to-point movement commands from M1, or as a qualitatively different command signal that specifies high-level desired movement goals, with the point-to-point movements and trajectories determined by a lower-level automated feedback control system. Modulation of neural activity in the primary sensory cortex (S1) has also been investigated as possibly generating signals for BCI systems ( ), though more recently studies have investigated writing sensory information into the cortex via intracortical microstimulation to enhance efferent control ( ).
Cortical Signals and Features
Most early applications of BCIs focused on recording single-unit action potentials from the aforementioned areas, and then correlating this activity to various parameters of movement. Individual single-unit action potentials were isolated either manually, by assessing how action potential (“spike”) waveforms fit into specific time–amplitude windows, or through automated spike-sorting algorithms ( ). More recent studies have shown that the aggregated multiunit spike trains on a single electrode channel (i.e., unsorted spikes) can carry significant tuning information for various movements and BCI applications ( ). Such a result is important because it eliminates the need for spike sorting, which can be a time-consuming, computationally burdensome, and imprecise process ( ). An alternative to decoding discrete spiking activity is to use the continuous multiunit activity (MUA) on each electrode (300–600 Hz, root-mean-square filtered). In some studies MUA-based predictions of arm reaching movements have been demonstrated to exceed predictions from spiking activities ( ), although this result has not been broadly reproduced. The LFPs are another neural feature that is an alternative to spiking activity. LFPs are thought to represent localized low-frequency synaptic activity of the neuronal population around an electrode tip. Several studies have shown LFPs to have high predictive power for reaching and grasping kinematics ( ), suggesting that LFPs may be a longer-term stable alternative to spiking activity. Finally, a number of these neural signal features may be combined to enhance overall predictive power and robustness in a hybrid neural decoder for BCI applications ( ).
Neural Decoding
Neural decoding is the process by which observed high-dimensional ( k x t ) patterns of cortical activity (typically the firing rate of neuronal action potentials) are mapped to a lower-dimensional ( M x t ) continuous “command vector” for effective command of an end effector, where k is the number of neural features, M is the number of commanded dimensions, and t is time. This M -dimensional space can be the continuous Cartesian velocities ( ẋ , ẏ , ż of a cursor or robotic arm end point), the continuous joint articulations of a robotic limb, or the stimulation values of multiple muscles during restored limb function via FES. While the decoding process can include extraction of discrete-state command signals, in addition to or instead of a continuous command vector, the majority of recent research efforts have focused on decoding continuous command information, and hence the present text focuses on options for continuous decoding.
Applications of continuous neural decoding have generally fallen into the classes of population vector analysis (PVA) and continuous linear filtering. PVA ( ) relies on the observation that M1 neurons often exhibit a “preferred direction” of firing related to movement direction, firing maximally for movements in one direction and declining in a cosine manner for movements in directions other than their preferred direction. PVA essentially maps the direction of desired arm movement to the vector summation of each recorded neural feature’s preferred direction, scaled by the neural feature’s instantaneous activity level. The success of PVA stems from early NHP experiments that characterized the neural activity of individual neurons as having a direction of maximal modulation, known as the preferred direction, in response to natural arm movements ( ). A neuron’s instantaneous firing rate β i could be described by
β i = b 0 + b x m x + b y m y + b z m z = b 0 + k cos Θ C M
Linear filtering decoding methods attempt to regress an output variable Y ( t ) (generally movement velocity) with respect to an input variable X ( t ) (generally multidimensional neural firing rates, with or without neural time history) through
Y ( t ) = b + ∑ u = − m m h ( u ) X ( t − u ) + ε ( t )
Related posts:
Fundamentals of Burst Stimulation of the Spinal Cord and Brain
Gene-Based Neuromodulation
Spinal Cord Stimulation for the Treatment of Low Back Pain
Epilepsy : Anatomy, Physiology, Pathophysiology, and Disorders
Colonic Electrical Stimulation for Constipation
Transcranial Alternating Current Stimulation and Transcranial Random Noise Stimulation
Stay updated, free articles. Join our Telegram channel
Full access? Get Clinical Tree
