Microelectrodes

Microelectrodes are penetrating electrodes designed to record from individual neurons, ­commonly referred to as single units . Most modern microelectrodes are configured as arrays of needlelike probes with electrodes along the shafts of each probe. The electrodes generally have surface areas and spacings on the order of 1 to 10 µm and are fabricated using silicon substrates. Individual microelectrode arrays are typically designed for sampling tens of hundreds, and potentially thousands, of neurons simultaneously.

Ideally, each microelectrode contact will measure action potentials from a single neuron. In practice, only a subset of contacts will yield clean recordings of action potentials from individual neurons. Otherwise, contacts can also measure local field potentials, representing a superposition of electrical activity from surrounding neurons.

The acquisition of action potentials from a population of neurons in the motor cortex can precisely determine movement intentions based on a phenomenon known as population coding . Put simply, individual neurons in the motor cortex have been shown to have firing rates that follow a cosine tuning curve, in which the firing rate of each neuron maximally corresponds to a preferred direction of movement and minimally corresponds to movements in the opposite direction (i.e., 180 degrees), with intermediate directions following a cosine curve (as shown in Fig. 10.3A–B ). By sampling a sufficient number of neurons in the motor cortex, shown to be as few as 10 to 100, continuous motor movements can be accurately reconstructed by taking the vector sum of the predicted movement directions for each neuron in the population as prescribed by the predetermined tuning curves ( Fig. 10.3C ). Because the resulting control signal represents a fairly direct mapping of the measured neuronal function, this approach yields naturalistic control of motor prostheses.

Population coding of cortical motor neurons.

(A) Hand reaching directions and associated spike raster plots from a single neuron. (B) Corresponding cosine tuning curve for the neuron. (C) Predicted direction vectors (red) based on the individual tuning curves of a population of neurons.

Modified from Georgopoulos AP, Schwartz A, Kettner R. Neuronal population coding of movement direction. Science . 1986;233(4771):1416–1419.

Electrocorticography and Stereotactic Electroencephalography

Electrocorticography (ECoG) uses comparatively larger disk electrodes in rectangular arrays or strips placed on the cortical surface via a craniotomy. The standard clinical electrodes are 5 mm in diameter with 1 cm spacing between contacts, although micro-ECoG arrays also have been used with diameters and spacings on the order of tens of microns. Although ECoG does not penetrate the brain or record single unit activity, it can provide detailed information for decoding sensory, motor, and cognitive processes for a BCI. ECoG recording provide signals that are morphologically similar to local field potentials with spectral bandwidth of roughly 0 to 250 Hz. ECoG is used clinically for localizing and mapping as part of surgical planning for intractable epilepsy and intraoperatively for glioma removal. To minimize the extent of the required craniotomy, ECoG grid placement is typically localized to particular cortical areas associated with the patient’s pathology, typically on a single hemisphere. For clinical purposes, the number of implanted ECoG electrodes typically ranges from 16 to 128, depending on the nature and location of the pathology.

A related technology employed for epileptic seizure localization is stereotactic electroencephalography (sEEG). sEEG electrodes are positioned along a cylindrical shaft (∼1 mm in diameter), which is inserted deeper into the brain through small burr holes in the skull using stereotactic guidance. The surface area, number of electrodes implanted, and nature of the measured signals of sEEG for clinical purposes are roughly comparable to ECoG, although higher-density contacts and microwires can be used. Although, in contrast to ECoG, sEEG electrodes penetrate brain tissue, there is evidence that the procedures are as effective as ECoG and have improved surgical recovery outcomes compared with the craniotomy required for ECoG. Also, unlike the single, localized craniotomy for ECoG, the comparatively minute burr holes for sEEG electrodes allows for a broader but sparse sampling of both superficial and deeper brain structures. This creates a trade-off for investigating and utilizing specific brain regions and networks compared with the generally superior local cortical resolution provided by ECoG.

Although the traditional lower-frequency rhythms can be observed in intracranial recordings, one unique and commonly used feature of ECoG/sEEG signals for BCI control is broadband gamma activity. This activity manifests as roughly uniform fluctuations in the signal spectrum over the approximate range of 70 to 250 Hz, which resembles a broadband noise process and is not practically observable in scalp EEG, as illustrated in Fig. 10.4 . This activity has been shown to be highly correlated with a wide variety of cognitive and behavioral function. It is important to note that although high-frequency oscillations exist, broadband gamma is generally not considered a brain oscillation akin to lower-frequency gamma (∼40–70 Hz) or the traditional lower-frequency EEG bands below 40 Hz (i.e., delta, theta, alpha, beta).

Electroencephalogram power spectral bands for hand movement versus rest. LFB, Low frequency band; HFB, high frequency band.

Modified from Miller KJ, Sorensen LB, Ojemann JG, Nijs M. Power-law scaling in the brain surface electric potential. PLoS Comput Biol . 2009;5.

Other Implanted Electrodes

One of the earliest demonstrations of invasive BCI control in humans was using neurotrophic electrodes. Neurotrophic electrodes consist of a 1- to 2-mm hollow glass cone attached to several gold conductive wires. The electrode is filled with trophic factors to encourage the growth of axons and dendrites into the cone. Unique features of neurotrophic electrodes are their ability to isolate and record from single or small groups of neurons and their biocompatibility for long-term implantation. However, compared with modern microelectrode arrays, neurotrophic electrodes have not been designed to conveniently sample larger numbers of neurons and have therefore not been widely adopted for BCI research.

More recently, an endovascular thin-film stent-electrode array was implanted in the superior sagittal sinus adjacent to the primary motor cortex of two participants with ALS. This sensor array, known as a stentrode , consists of 17 circumferential sensors on an 8 mm × 40 mm monolithic, self-expanding nitinol scaffold designed for minimally invasive intracranial delivery using catheter venography. The array is connected to a 50-cm flexible transvascular lead and inserted into an inductively powered wireless telemetry unit. For this feasibility study, the participants were able to successfully perform a typing task using attempted movements in combination with an eye tracker for cursor navigation.

Active Control Signals

The most commonly used active control signals for robotic control applications are event-related desynchronization/synchronization (ERD/ERS) and sensorimotor rhythms (SMRs). These are more naturalistic control signals since they are generated from the motor cortex during actual and imagined movements.

SMRs are idling rhythms occurring in the alpha band of the motor cortex during rest, which can be observed in an estimated 80% of the population. To distinguish from unrelated alpha-band activity, these are often referred to as mu rhythms. A defining characteristic of these rhythms is that the amplitude is attenuated over the corresponding area of the motor cortex during continued actual or imagined movements. Related modulations can also be observed in the beta and gamma bands, although it is not clear whether these are contributed by distinct neural processes or are simply, in part, a by-product of the unique temporal morphology of mu rhythms. Users can learn to modulate SMR amplitudes continuously to achieve continuous, dimensional BCI control (e.g., imagine right-hand movement to move a cursor to the right). The temporal, spatial, and spectral characteristics and how they can be mapped to dimensional BCI control are illustrated in Fig. 10.5A .

Active control signals.

(A) Topographical and spectral representations of sensorimotor rhythms for dimensional control using left- and right-hand movement/imagery. (B) ERS/ERD of the alpha band corresponding to an isolated movement/imagery. ERD, Event-related desynchronization; ERS, event-related synchronization.

ERD refers to the transient decrease in alpha-band power over the motor cortex associated with the onset of isolated movement or imagery, followed by the transient ERS increase as the motor activity returns to the resting state, as shown in Fig. 10.5B . Thus, this transient activity can be used as a discrete switch, for instance to activate/deactivate a hand grasp orthotic.

Other active control signals exist, such as alpha-band modulation during cognitive processes such as mental math or object rotation. However, although these modulations can be reliably detected in the EEG, they require the user to perform a mental task that is not directly related to the primary robotic control task. This can be unnatural, unintuitive, and distracting and is ultimately not an ideal control signal for a practical assistive device.

Although ERD/ERS and SMR represent the most natural and intuitive scalp EEG activity for dimensional control of robotic or assistive devices, these signals tend to be nonstationary, and significant task training may be required for users to reliably modulate these signals to achieve practical device control.

Reactive Control Signals

Reactive control signals result from predictable changes in brain activity generated from the user’s attention to specific sensory stimuli that are mapped to aspects of the control task. These predictable changes in brain activity are known as stimulus evoked potentials (SEPs). SEPs can either be transient, which are generally time-locked with the sensory stimulus and resolve on the order of a second, or steady state, which continue as long as attention to the stimulus is maintained.

One of the earliest scalp EEG BCIs is the P300 speller. It operates based on the brain’s response to a rare or novel sensory stimulus. This brain response is known as an event-related potential (ERP), which is a specific type of SEP that includes a cognitive component (opposed to a purely sensory reflex). In this case, the cognitive component is the recognition that the stimulus is rare or novel.

In a simple P300 speller paradigm, the user is presented with a matrix of flashing symbols on a computer monitor, for instance, similar to a visual keyboard. The user attends to the desired symbol as each symbol is flashed in a random sequence. The general concept is illustrated in Fig. 10.6 . Because the user does not know when the desired symbol will be flashed, an ERP is generated when the symbol is flashed while the user maintains fixed attention. Multiple ERPs must be accumulated over multiple stimulus sequences to reliably represent the characteristic brain response and determine the desired symbol. This process can be made more efficient via row/column flashing or a variety of other enhancements to the original paradigm. Although the original P300 speller paradigm was based on visual attention, auditory or tactile attention can also be utilized with decreased performance compared with visual paradigms.

Reactive control signals for a visual paradigm. Six targets presented on a computer monitor flash independently according to (1) a steady-state pattern with a unique stimulus frequency for each target or (2) transient pulses with a unique stimulus onsets for each target. The user focuses visual attention on a single target and (1) the electroencephalogram (EEG) over the occipital cortex oscillates at the target stimulus frequency (plus harmonics) for the steady-state stimulus scenario or (2) an EEG-evoked potential can be detected over the visual cortex (and potentially other areas) with respect to the stimulus onset for the transient stimulus scenario.

Effectively, such paradigms provide discrete command selection, equivalent to a computer keyboard—albeit at a much slower rate (∼4–5 symbols per minute). These commands can be mapped to achieve discrete dimensional or goal-oriented control of a robot, for example.

To achieve continuous dimensional control, steady-state sensory evoked potentials can be used, the most practical being the steady-state visual evoked potential (SSVEP). Unlike transient evoked potentials resulting from a single stimulus event, the steady-state evoked potentials result from a repeating stimulus pattern, such as a blinking light. As such, rather than generating a transient EEG response, predictable, continuous repeating EEG patterns over the occipital cortex are generated with attention to the repeating stimulus patterns. These EEG patterns manifest as distinct spectral peaks in the frequency domain that can be reliably identified and tracked, in contrast to the time-domain techniques used to identify transient evoked potentials.

A typical SSVEP interface consists of a spatial array of individual lights or symbols on a screen, each concurrently repeating with a unique temporal flashing pattern and mapped to a control command, as shown in Fig. 10.6 . When the user focuses attention to a specific light/symbol, the brain activity unique to that symbol is detected and the associated command is executed. In contrast to transient evoked potentials, the command can be maintained with continued attention to the stimulus and can be readily changed by shifting attention to a different stimulus to execute the associated control command in a continuous fashion. This can enable joystick-like dimensional control of a robotic device with latency on the order of 0.5 to 1 seconds. It is also possible to achieve high communication rates with discrete selection via keyboard-like interfaces using SSVEP and related variants. As with transient responses, BCIs that utilize steady-state auditory and tactile (somatosensory) potentials, have also been developed.

Control paradigms that rely on visual evoked potentials have several limitations. Foremost, performance can drastically decrease with inability to control eye movements, which is the case for many patients with locked-in syndrome. Also, these interfaces often create a visual cacophony, can be unnatural and fatiguing, and can be distracting from the primary control task, which can make them inconvenient and impractical for long-term use.