Improvement of a Bidirectional Brain-Computer Interface for Neural Engineering Research

1
Improvement of a Bidirectional Brain-Computer Interface for
Neural Engineering Research
Hayley Boyd
Advisor: Eberhard Fetz
June 11, 2019
The Neurochip-3 is an autonomous battery-powered device developed by the Fetz laboratory that
is capable of recording, computing, and stimulating for brain-computer interfaces. When used for
neuroscience experiments involving stimulation and recording of cortical electrodes implanted in
the cortex of a non-human primate, large recovery artifacts are often seen in the recorded signal
following stimulation. These artifacts greatly impede experiments that focus on low-frequency
neural activity immediately following electrical stimulation of the cortex. Here, the artifact is
described and modelled with the goal of eliminating or reducing the problem. The generation of
this artifact is modelled in three parts: the pre-amplifier pathway, saturation of the amplifier in
response to high-voltage inputs, and signal offset recovery. Further descriptions include
observations of inconsistent artifact behavior between amplifier channels, the role of impedance
of cortical electrodes, and a comparison of functionality with another BCI for neural stimulation
and recording. Possible solutions are investigated, including the development of circuit
modifications to prevent stimulation artifacts from entering the amplifier, as well as other
alterations to pre-existing Neurochip circuitry that might improve offset recovery time.

2
1. Introduction
The Neurochip-3 [NC3] is an autonomous battery-powered device developed by the Fetz
laboratory that is capable of recording, computing, and stimulating for open- and closed-loop
brain-computer interfaces (Zanos et al 2011;
https://depts.washington.edu/fetzweb/neurochip3.html). Most neurophysiological experiments in
non-human primates [NHPs] involve recording of neural activity in intermittent sessions, and
brain-computer interface systems that process and respond to neural activity in real time generally
involve rack-mounted equipment. However, the Neurochip is small enough to fit in a chamber on
an NHP’s head, where it interfaces with a neural implant and allows recording and open- or closed-
loop stimulation to occur over many hours in the freely behaving animal. In this way,
neurophysiology experiments are not limited to shorter sessions with constrained animals but may
also be performed in overnight sessions in which the animal may move freely and enter states of
rest.
Previous iterations of the Neurochip have been used in highly successful closed-loop studies.
These have included strengthening of synaptic connections between two sites in the brain (Jackson
et al 2006, Nishimura et al 2013, Zanos et al 2018), use of state-dependent stimulation during sleep
to improve recollection of a task (Rembado et al 2017), and a proof of concept for artificial
recurrent connections such as between the motor cortex and paralyzed muscles (Moritz et al 2008).
These few examples demonstrate the breadth of research that can be conducted with the NC3. Not
only does it allow for research focused on characterizing mechanisms of the brain, but also for
development of post-injury therapy and even methods for enhancing brain function of healthy
individuals.
The NC3 is the third generation of the Neurochip and is considerably more powerful than its
predecessors. While the Neurochip-2 had three recording channels with 8-bit recording resolution
(Zanos et al 2011), the NC3 features an Intan biophysical signal amplifier allowing 16 differential
or 32 unipolar inputs, which can collect 16-bit samples at up to 20kHz sampling rate. The data is
stored on a conveniently removable 64GB micro SD card. Compared to the three stimulation
outputs on the NC2, the NC3 has six, allowing for as many single-ended or three pairwise
differential stimulation channels. It has two powerful computer chips: an ARM processor and

3
FPGA. It also incorporates other new features, including a 3-axis accelerometer to allow
measurement of the animal’s movement and position (Appendix A-C).
The addition of the RHD2216 Intan biophysical signal amplifier contributes some of the largest
and most important improvements to the Neurochip. However, the Intan chip also introduces a
significant problem. Large signal displacements (henceforth referred to as recovery artifacts)
persistently occur in many or all recording channels of the NC3 directly following stimulation.
The cortical response immediately following neural stimulation is often the information that is
sought after in NC3 experiments, and in many cases the artifact is so large as to render the data
extremely difficult to process or unattainable. See Figure 1 for a visualization of the artifact in
comparison to the evoked potentials expected to be seen. The size and shape of the recovery artifact
varies with stimulation parameters such as amplitude, duration, and location of stimulation as well
as amplifier filter settings. The artifact is also known to change over time even when these
parameters are kept constant, demonstrated in Figure 2. This problem is a result of saturation of
the Intan chip, which occurs when the stimulation artifact at the amplifier input exceeds the voltage
input range of +/-5mV, demonstrated in Figure 3.
Figure 1: Stimulation-triggered averages showing the post-stimulation artifact, as well as what the evoked
potential signal should look like without the artifact, acquired from subtraction of an exponential fit. Each
pair is from an experiment in which different locations were stimulated with the same parameters (1250μA
pulses (n=5) at 300Hz).

4
Figure 2: Averages of the artifact across all recording channels over 15-minute intervals taken once every
hour. The stimulation parameters are consistent across time; however, the artifact is observed to change in
magnitude and polarity. The amplitude scale matches that of the Figure 1 Artifact plots; these recovery
artifacts are on a similar time scale as the neurophysiological responses and ten times the amplitude.
Figure 3: Stimulus-triggered averages for 6 separate channels recorded with the NC3, with the y-axis
expanded to allow visualization of the entire stimulus artifact. Horizontal lines indicate the ±5mV allowable
input range for the Intan chip. Channels in which the stimulation artifact breaches this line show a recovery
artifact, while channels with smaller stimulation artifacts do not.

5
The design of the Intan amplifier chip includes some features for artifact cancellation. The
Amplifier Fast Settle is a feature present on the RHS2000 series amplifiers that allows the
amplifiers to be reset by clamping them at baseline. The recommended duration of this setting per
amplifier reset is 2.5/fH, where fH is the upper bandwidth of the amplifier. A typical fH value for
NC3 recordings focused on evoked potentials is around 7.5 kHz, meaning a 350μs fast settle should
be enough to remove the offset. Unfortunately, this feature has not been shown to reliably solve
the artifact problem. It is possible the difference between the clamped state and the signal itself
can result in artifact, and it has also been noted that the fast settle feature can introduce a switch
artifact into the signal.
Because the internal schematic of the Intan chip is unavailable, and because we have no ability or
intent to create design changes for this proprietary device, artifact cancellation designs must be
made external to the amplifier. This means the NC3 artifact problem cannot be removed at its
source, rather solutions must work around the amplifier chip.
To address this recovery artifact, a model of the system that describes the behavior of the artifact
is created through analysis relating stimulation parameters such as intensity and frequency, as well
as filter settings to the size and shape of the resulting artifacts. Then, design changes including
additional circuit components and alterations to the existing Neurochip circuitry are proposed and
iterated to reduce or remove the artifact.

6
2. Development of Design Specifications
The main need and goal is to develop a solution to the problem of recovering biological stimulus-
evoked physiological potentials from the superimposed artifact. Other needs include continued
safety of the device and its operation for both researchers and animals, and no trade-off in
applicability of the device. Further, because the portability of the Neurochip is a key feature, design
changes should not significantly increase the size or weight of the device. The design should also
avoid increasing the complexity or difficulty of operating the NC3. These needs have been ordered
on the following Needs Table, along with a subjective ranking of relative importance.
# Need Importance (1=highest)
1 Design must eliminate or mitigate stimulation artifact 1
2 Device must remain safe for both researchers and animals 1
3 Design changes must not reduce usefulness of device 2
4 Design changes should not increase weight or size of the
Neurochip
2
5 Design should not increase complexity of NC3 set-up 3
The Needs-Metric table below identifies several measurable metrics and their units along with
their correspondence to the previously defined needs. These metrics can be used to quantitatively
determine efficacy of a final design.
# Metric Units Corresponding Needs
1 Artifact Reduction Area under curve reduction
compared to corresponding pre-
solution trials
1
2 Stimulation consistency:
Comparison between
programmed stimulation and
stimulation waveforms
measured with oscilloscope
Error between expected and
measured stimulation amplitude,
duration, and frequency (if
applicable)
2, 3
3 Time to set up experiment Change in the amount of time
required for the researcher to start
an experiment
5
4 User commentary Subjective comments and
observations from researchers
5
5 Weight of hardware changes Grams 4
6 Size of hardware changes cm/cm^3 4

7
Real-World Constraints
Ethical and Social Concerns
Because the Neurochip is designed for minimally invasive experimentation in research animals, it
is subject to ethical concerns regarding the wellbeing of the animals. All NHP protocols are subject
to heavy regulation and must be approved by IACUC to ensure animal wellbeing is maximized.
The relatively small scale of this project within the greater context of the Neurochip combined
with distance from the public eye largely precludes societal constraints aside from those inherent
in NHP research.
This project, being an accessory to the already established NC3, is not expected to introduce any
problems that may result in decreased quality of life or impose other concerns for animal welfare.
Economic Concerns
The Neurochip is designed for use by a specialized and relatively small user group. For this reason,
documentation of any changes created in this project is not expected to require translation or
widespread distribution. Funding currently covering development and support of the Neurochip
covered minor expenses incurred in the acquisition of electronic components.
Legal and Regulatory Concerns
The scope of this project is not expected to change the form or function of the NC3 in any
significant manner, so there is no danger of departing from the device currently approved by
IACUC or causing additional risks for the animals.
Other Factors
Because this project is a relatively minor addition to a pre-existing device with an established but
small user base within a relatively small and isolated field, the author foresees no considerations
needing to be made in relation to public health or other global factors. Further, cultural
considerations need be considered only insofar as the device be appropriate to the culture of
neuroscientists who use it, which in terms of this device only require it to be maximally
understandable and ergonomic for people in this field. At no point will the NC3 be a mass-
produced or widespread product, so there is no cause for concern regarding the environmental
impact of this project beyond the sustainability of methods used during the testing process.

8
3. Methods
3.1 Single-Ended Vs. Differential Neural Recordings
Figure 4: A basic diagram for differential neural recordings. These recording systems utilize two recording
electrodes per channel. The locations of connection to the brain in this diagram are not meaningful.
An important distinction must be made between single-ended and differential recording schemes.
In this paper, differential recording refers to a system in which each recording output is an
amplified difference between two closely-spaced cortical electrodes.

9
Figure 5: A basic diagram for single-ended neural recordings. These recording systems use the same
reference for every channel, meaning only one electrode is utilized per recording channel. The locations of
connection to the brain in this diagram are not meaningful.
The majority of this work focuses on single-ended recording schemes. For these configurations,
the same single reference is used for every amplifier. In all NHP recordings this reference is the
pedestal ground, i.e. the metal base of a connector pedestal that is screwed into the skull and allows
recording and stimulation systems to interface with the cortical implant. When not otherwise
specified, it is also shorted to the tissue ground, which is the reference point for tissue voltages and
return path for electrical stimulation. This configuration is chosen to reduce noise in recordings.

10
3.2 Isolated Amplifier Testing Configuration
Figure 6: Diagram of the circuit used for isolated testing the amplifier. Resistance R between the
stimulator channels and ground is varied in some tests, resulting in altered artifact characteristics.
In order to describe the amplifier’s recovery behavior, a stand-alone RHS2116 Intan amplifier chip
was tested apart from the NC3 recording system. The amplifier was powered directly with 5V
from a power source. The NC3 stimulator was used to generate the input, which along with the
output of the Intan chip was measured using an oscilloscope. The input was plotted along with the
output to ensure that the stimulation was occurring as expected. Biphasic stimulations ranging
20uA, 25uA…. 60uA were tested on this setup, and a selection of these trials is plotted in the
results in Figures 9&10.
3.3 The Potato Configuration
To reduce animal use and improve ease of testing, a Neurochip testing configuration that utilizes
a potato in place of the brain was created. In this configuration, two custom-made electrode pairs
of the same design as those implanted in the animals are directly inserted into a potato. One pair
is used as the anode and cathode for stimulation, and the other pair is recorded on the Neurochip
as two single-ended channels. The reference and tissue ground are inserted into the potato near the
dual electrodes. The advantage of this experimental design over the bench setup that utilized a
fixed resistor as the load is that it more accurately recreates the impedances present in the animal,
allowing easier generation of recovery artifacts that are more like those seen in in vivo recordings.

11
Of course, this setup has the disadvantage that the impedance is not always known or consistent
between tests. Each time the potato configuration is set up, the electrodes are in a slightly different
location in relation to each other. Further, over time the moisture surrounding the electrode will
evaporate, resulting in significant impedance changes over the scale of minutes. The impedance
inherent to the electrodes themselves also cannot be assumed to be constant for all trials, as the
parylene coating on the platinum iridium tips is likely damaged by repeated re-insertion into the
potato. Any plots from separate potato recordings that are compared were taken immediately
following one another and none of the potato interfaces were moved between recordings.
Figure 7: Left: diagram of the potato configuration. Right: structure of the dual electrode pairs used for
stimulation and recording. Each pair includes a deep and shallow cortical electrode: 3mm and 1mm
electrode depths respectively. Each electrode is shielded by a thin coating of Parylene except for at the tips.
All output plots are averages of the recording channels from 100 repeated 2Hz stimulations plotted
from 5ms prior to stimulation to 200ms after stimulation. Gray lines on these plots represent the
individual trials. For some tests, a Tektronix signal generator was used to inject a sine wave into
the potato in order to simulate neurological signals and ensure data is being recorded.

12
4. Results
4.1 The basic model
To characterize the recovery artifact, the model of the system is being separated into three main
parts:
1. How the stimulation parameters become the stimulation voltage seen by the Intan
amplifier. This pathway is defined by the impedance of the stimulation and recording
electrodes as well as the neural tissue surrounding and between them.
2. How the stimulation voltage seen by the amplifier becomes a signal offset following the
stimulation artifact.
3. The shape of the artifact as the signal recovers from the post-stimulation offset.
4.1.1 The pre-amplifier model
Because the stimulation artifact that enters the amplifier is the cause of the recovery artifact, it is
relevant to characterize the circuit between the stimulator and the amplifier. This system can be
described in terms of the pathway of the stimulation pulse. It begins with the NC3 stimulator
injecting current directly into the cortex via two closely spaced electrodes. These electrodes along
with the interface between them and the surrounding tissue are known to have a nonlinear
impedance and are modeled by a resistor and capacitor in parallel. The stimulation pulse then
passes through the neural tissue and is picked up by recording electrodes of the same design as the
stimulation electrodes. The effect of the distance between stimulation and recording electrodes is
modelled by the resistance of the load between them.

13
Figure 8: A circuit diagram modelling the stimulation electrodes, cortical tissue, and recording electrodes
that the stimulation pulse travels through between the stimulator and Intan amplifier chip.
4.1.2 Post-stimulation offset generation
The output behavior of the Intan amplifier directly following a high voltage input is the crux of
the problem. Ideally, a mathematical model would be generated to describe this behavior and allow
accurate prediction of output offset given specific input parameters. Unfortunately, such a
complete and accurate mathematical model is not expected to be possible due to the nature of the
problem. The Intan amplifier is a complex device being used with highly specific parameters that
were not anticipated during its creation or accounted for in its design. Expecting to mathematically
define artifact offset following saturation of such a complex device, and more so without having
knowledge of its interior schematic, is an unrealistic goal, further complicated by the instability of
the artifact offset reported in Figure 2. For these reasons, this section of the model instead focuses
on identifying attributes of the recovery artifact seen in response to various parameters and
explaining them and their differences qualitatively.
The following observations pertain to the stand-alone artifact recovery testing of the Intan
amplifier chip as described in section 3.2. See Figure 6 for the circuit diagram of this testing
configuration.

14
Figure 9: Plots of 2Hz testing stimulations using the circuit from Figure 6 with an R value of 240kΩ. Top:
signal output from the Intan amplifier for three different stimulation trials using amplitudes of 20, 40, and
60μA left to right for 40 separate pulses each. Bottom: Same plots with decreased time and amplitude scales
to allow better visualization of individual recovery artifacts. Note the voltage axes for each condition have
the same range, however the 60μA plot is centered at an offset to 4mV.
In Figure 9, the stimulation and recovery artifacts are both seen to vary in amplitude even within
the same stimulation and setup parameters. Within the same setup parameters at different
stimulation amplitudes, the artifact is also seen to change drastically. At low amplitudes, the post-
stimulation signal starts at a negative offset following the stimulation and slowly returns to zero
between stimulations. At higher amplitudes, the recovery artifact starts at a positive offset and
returns towards baseline. At 60μA, this initial offset is so high that the output never returns to zero,
resulting in a signal offset at about 4mV. Higher stimulations also result in outputs with smaller
peak-to-peak stimulation artifacts. The reason for this comes from the Intan chip’s voltage limit.
At the lower stimulation amplitudes, the stimulation artifact is still railing in the output; however

15
it is not significantly higher than the rail limit at 5mV. As the stimulation increases, the artifact
voltage pushes further past that rail, resulting in longer recovery time. In the output signals, this
makes the negative second part of the biphasic stimulation appear much smaller than the initial
positive phase. This pattern also explains why the recovery slope changes directions at different
amplitudes. At low amplitudes, when the stimulation artifact ends with a significant negative
component, the signal must recover from that negative displacement. At higher amplitudes, when
the duration of the stimulation artifact is spent recovering from the positive rail, the output does
not “see” the negative phase before the stimulation is over, meaning the baseline must recover
from that positive offset after the stimulation.
Figure 10: Plots of 2Hz testing stimulations using the circuit from Figure 6 with an R value of 100kΩ.
Top: 40 separate pulses. Bottom: Same plots with decreased time and amplitude scales to allow better
visualization of individual recovery artifacts. Note that each voltage axis on these plots is across the same
voltage range, although the 60μA stim plot is offset centered at 2mV.

16
To reduce the offset the resistors between the NC3 stimulator and the ground were lowered from
240kΩ to 100kΩ. This decreased the offset, bringing the 60μA test from a 4mV offset to about a
2mV offset. The slope at which the offset artifact returns to baseline also changed drastically, seen
to be near zero at the higher stimulation amplitudes. However, the size of the offset continues to
vary seemingly randomly within a roughly 0.5mV range. Further, it can be noted that the
stimulation amplitude at which the recovery artifact shifts from having a positive slope to a
negative slope has increased in this setup.
4.1.3 Post-stimulus decay function
Although the offset resulting from amplifier saturation is not completely predictable, the decay of
the signal back to baseline is understood to be exponential with a time constant equal to 1/(2*pi*f)
where fL is the lower bandwidth of the high-gain amplifiers. The equation thus looks like this:
𝑁(𝑡) = 𝑁0 𝑒−2𝜋𝑓 𝐿t
Where N0 is the initial offset value following the stimulation. Note that recovery time decreases
exponentially with fL.
Figure 11: Blue: A recovery artifact from an in vivo recording with 700μA stimulations and a high-pass
cutoff of 3Hz. Red: an overlay of the post-stimulus recovery curve as predicted by the post-stimulus decay
function.

17
Effects of high-pass filter on post-stimulus decay
The effect of the high-pass filter on post-stimulus decay was demonstrated in the following 50-
second recordings, using 400μA 2Hz stimulations in a sedated animal. Highpass cutoffs were
tested at 5, 10, 15, 25, 50, 100, 200, and 300 Hz. All recordings had a lowpass of 7.5 kHz.
Figure 12: Left: Two saturated channels recorded with high pass values of 5, 25, 50, 100, 200, and 300 Hz.

18
Figure 13: Left: A non-saturated channel with a cortical-evoked potential recorded with a high pass of 5Hz.
Right: the same channel recorded with a high pass of 300Hz.
Figure 14: Plots of the post-stimulus root mean square of 16 channels for 8 different high pass frequencies.
Plot titles are given as “Recording channel number: electrode number: electrode coordinate”.

19
Figure 15: Each channel shown as recorded with a high pass of 5Hz, seen on the 6mV scale.
Figure 16: Each channel shown as recorded with a high pass of 300Hz, seen on the 6mV scale.

20
4.2 Additional artifact descriptions
4.2.1 Inconsistent channel behavior
To document behavior of different NC3 recording channels in response to stimulation, the
following test was run. Using the potato configuration as described in section 3.3, recording
electrode 1 was connected to NC3 recording channels 1 and 2 while recording electrode 2 was
connected to NC3 recording channels 3 and 4. Stimulation artifacts were then generated using a
50μA symmetric biphasic square wave administered through the stimulation electrodes. In a
second recording, these recording electrode and NC3 channels were reversed such that recording
electrode 1 was connected to NC3 channels 3 and 4 while recording electrode 2 was connected to
NC3 channels 1 and 2. These tests were then immediately repeated with 200μA and then 500μA
stimuli.
Figure 17: The potato configuration for relative channel behavior testing. In this diagram the dual
electrodes pairs are displayed as pairs of colored circles. The pair of stimulation electrodes as well as rec1
and rec2 are paired electrodes of the same design as those implanted in monkey H. In reality they are very
near to each other, with one situated deeper into the potato than the other by about 2-3mm.

21
Figure 18 A&B: Channel behavior with 50μA stimuli delivered at 2Hz. In test A (first row of plots),
recording channels 1 and 2 are connected to rec1 (blue) and recording channels 3 and 4 are connected to
rec2 (orange). In test B (second row of plots), recording channels 3 and 4 are connected to rec1 (blue) and
recording channels 1 and 2 are connected to rec2 (orange).
Figure 18 C&D: Channel behavior with 600μA stimuli delivered at 2Hz. In test C (first row of plots),
recording channels 1 and 2 are connected to rec1 (blue) and recording channels 3 and 4 are connected to
rec2 (orange). In test D (second row of plots), recording channels 3 and 4 are connected to rec1 (blue) and
recording channels 1 and 2 are connected to rec2 (orange).
In the figures above, the pairs of plots within each colored box are connected to the same recording
electrode and thus receive the same electrical input. Plot pairs are thus expected to be identical. In
tests A and B they appear somewhat similar; however there is some notable offset difference
between channels 3 and 4 in test A. Because each recording was taken subsequently and the
channels were mostly behaving consistently between trials, it can also be said that the electrical

22
input seen by channels 1 and 2 in test A is essentially the same as the electrical input seen by
channels 3 and 4 in test B (blue box), and so forth with channels 3 and 4 in test A with channels 1
and 2 in test B. Channel behavior tests C and D are of the exact same format as tests A and B,
except for having a higher stimulation current. Again, when the input for each channel pair is
switched, the waveforms do not appear to simply switch channels as expected. Rather, channel
identity appears to be a major factor in determining behavior for these four channels.
Although the difference in recovery behavior between different channels is extreme, it should be
noted that each amplifier channel within the Intan chip is a separate circuit. It is therefore
conceivable that each channel has different saturation recovery behavior due to minute differences
within manufacturing specifications.
4.2.2 Investigation of electrode impedance effects
To investigate the role of electrode impedance in artifact generation, a White Matter nanoZ was
used to check impedances of Monkey H’s channels over frequencies ranging from 1 to 2000Hz.
Plots for impedance against frequency for channels 49 through 96 are attached in Appendix D.
Most channels exhibit a similar frequency-impedance relationship, although in some cases the
impedance scale is very high, suggesting an electrode is nonfunctional. When compared to a
Neurochip recording using the same channels, there is no correlation noted between overly high
channel impedance and recovery artifact existence or behavior. Example pairs of artifacts and
corresponding channel impedance are included here.
Figure 19A: Impedance plot and unsaturated example waveform for channel 87, with NC3 recording
channel 1. 350μA biphasic stimulations, recording bandpass ranging 0.75Hz to 75kHz. This channel has
very high impedance, and saturation is not occurring under these parameters.

23
Figure 19B: Impedance plot and unsaturated example waveform for channel 90, with NC3 recording
normal to low impedance, and saturation is not occurring under these parameters. With smaller voltage
axes, an evoked response is visible in this signal.
Figure 19C: Impedance plot and unsaturated example waveform for channel 72, with NC3 recording
normal impedance, and saturation is occurring and resulting in a large recovery artifact.
Figure 19 demonstrates a high-impedance channel that is not saturating, a normal-impedance
channel that is not saturating, and a normal-impedance channel that is saturating. By comparing
stimulation average plots from two separate recordings to the impedances of the animal’s
electrodes, it is currently my observation that electrode impedances are not as relevant to artifact
differences as Intan channel identities are.
4.2.3 Investigation of spatial relationships
Distance between stimulation and recording electrodes is another factor thought to contribute to
the size of stimulation artifacts. To check for this, channels from short test recordings were marked
as being or not being saturated. This information was considered along with the mapping of cortical

24
electrodes for this animal, seen in Appendix F. While it appears that channels closer to the
stimulation site might be more likely to be saturated, the pattern is not strong: there will be some
bad channels further from the stimulation site than other good channels and vice versa.
Relative channel position on the Blackrock Cereport connectors was also investigated to ensure
channel crosstalk is not occurring in this hardware. No patterns involving the positions of saturated
and unsaturated channels on this device was observed.
4.3.4 Comparison with an alternative device
To determine what other factors might contribute to the recovery artifact, a comparison was made
with another successful neural stimulation and recording system. Sung Lee from the Electronics
and Telecommunications Research Institute (ETRI) provided this device, which also utilizes the
Intan biophysical amplifier. The ETRI device is differently constructed than the NC3, designed to
allow for Bluetooth wireless data transfer and parameter tuning. It has fewer recording channels,
and its stimulator delivers set amplitudes of voltage rather than current as delivered by the NC3
stimulator. The ETRI device is used successfully in EMG-triggered spinal stimulation in rats
without showing significant recovery artifacts. We tested it alongside the NC3 with Monkey H to
determine whether it would be successful in recording low-frequency neural signals directly
following cortical stimulation.
Both devices were used to record from the same two cortical electrodes in the sedated animal.
Stimulation was applied to the same site in each test, and testing combinations included ETRI
recording with ETRI stimulation, ETRI recording with NC3 stimulation, NC3 recording with
ETRI stimulation, and NC3 recording with NC3 stimulation, each tested at multiple stimulation
intensities. In this way, we can compare two sets of NC3 and ETRI device recordings in which
both are receiving the exact same stimulation conditions. It was determined that the ETRI device
does show amplifier saturation recovery artifacts as the NC3 does when under the same conditions,
although the size, shape, and variance of the signal can vary somewhat between the two devices.
The conclusion drawn from these tests is that there is not a significant improvement in saturation
recovery to be found in the precise differences between these two devices. However, slight
differences in artifact behavior suggest it might still be possible to modify the Neurochip for better
recovery from amplifier saturation.

25
4.3 Circuit Designs
Because the recovery artifact originates within the Intan biophysical amplifier, there is no way to
address the problem at its source. However, one possibility is to develop circuitry that reduces the
amount of stimulation current that enters the amplifier. In theory, a simple switch may be used to
deflect the recording channel to the tissue ground and reference for a short period of time
surrounding the duration of the stimulation. This prevents the large stimulation current from
entering the amplifier directly through the stimulation-to-cortex-to-recording pathway, instead
setting the input of the recording channel to zero for this duration. No physiological data may be
recorded during this period; however, it is already unusable due to the stimulation artifact and is
of short enough duration that it does not encroach on stimulus evoked potentials.
Figure 20: Stimulus switch-out scheme.
This switching method was first tested using the potato configuration. The switching circuit was
implemented using a TS3A5018 10Ω Quad SPDT analog switch powered by a separate battery
and connected between the recording electrodes and NC3 recording inputs using a breadboard. In
order to confirm that signals can be passed through the switching circuit successfully, a Tektronix
signal generator was used to inject a sine wave into the potato through one of the stimulation
electrodes. The waveform was generated at 200Hz, which is similar to the frequency of evoked

26
potentials. However, it should be noted that this waveform is roughly ten times the amplitude of
evoked potentials, due to an amplitude minimum imposed by the signal generator.
Figure 21: Potato configuration with a 200Hz
sinewave injected into potato by signal generator along
with stimulation from Neurochip stimulator. Each plot
is a single sweep. During switching the recording
channel is connected to tissue ground.
A: Channel 1 with switching.
B: Channel 2 with switching.
C: Channel 1 with no switching.
D: Channel 2 with no switching.
The results of this potato configuration sine wave test are shown in Figure 21. Although the
switching mechanism does not prevent the amplifier from reaching its saturation point at -5mV, it
does appear to reduce the size of the post-stimulation artifact offset.
For a closer examination of the effect of the stimulus switch-out scheme, the potato configuration
was used again without the injected sine wave.

27
Figure 22: A demonstration of stimulation artifact waveform changes due to switching. Two vertical lines
across all plots represent the beginning and end of the switching period.
A: Single sweep demonstrating the switch-out to TG
over a sinewave from a signal generator.
B: Channel 1 from potato with stimulation and
switching.
C: Channel 1 from potato with stimulation and no
switching.
D: Channel 1 from potato with switching and no
stimulation.
E: Channel 2 from potato with stimulation and
switching.
F: Channel 2 from potato with stimulation and no
switching.
G: Channel 2 from potato with switching and no
stimulation.
The most notable components of Figure 22 are plots D and G, which are controls with switching
but no stimulation. Because no stimulation is delivered in these trials, the entire waveform is due
to an artifact of the switching circuit itself. Interestingly, channel 1 switching artifacts, from both

28
the stimulation and no-stimulation tests, show considerable variance in stimulation artifact size
and polarity, while the channel 2 waveforms are very consistent. Additionally, it can be noted that
in all cases the artifact extends beyond the switch-out duration. Because this duration is set to be
twice the length of the stimulation waveform, it is seen that the stimulation artifact waveform is
elongated by over a factor of 2. If switching circuits are further investigated, it is important to be
aware of stimulation artifact elongation and ensure the switching period is long enough to fully
encompass the artifact.
To determine whether these patterns are relevant to the true model or simply phenomena of the
potato testing configuration, these experiments were repeated with an implanted NHP.
Figure 23: An in vivo demonstration of
stimulation artifact waveform changes due to
switching. Two vertical lines across all plots
represent the beginning and end of the switch-
out period.
A: Single sweep demonstrating the switch-out
to combined TG/reference over a sinewave
from a signal generator.
B: Channel 1 from animal with stimulation and
switching.
C: Channel 1 from animal with stimulation and
no switching.
D: Channel 1 from animal with switching and
no stimulation.

29
As seen in the potato model, there continues to be a large switching artifact during in vivo tests.
Additionally, the pattern of stimulation artifact elongation is present here as well, although to a
lesser extent.
Although there does appear to be some similarity in artifact geometry for a given channel between
stimulation/switching and no-stimulation/switching experiments, these pairs of waveforms are far
from identical. Rather than the switching artifact simply overtaking the stimulation artifact, the
stimulation is in some way still influencing the artifact.
Figure 24: A circuit design that acts to insert a voltage divider for a short duration surrounding the
stimulus. This diagram is a simplification that does not include the pre-amplifier circuitry on the NC3.
These circuit elements exist between the switching circuit and Intan amplifier.
In hopes of reducing switching noise, we tested a circuit design that keeps the line from recording
electrode to amplifier intact. This circuit was tested using the potato setup, the switch triggered
with the timing as described in Figure 20. Potato recordings with and without the switch being
powered were generated with and without a sinewave being injected into the potato. This design
was tested with both the TS3A5018 (single-pole double-throw) and the 74LVC4066 (single-pole
single-throw) switch. In both cases, the artifact introduced by the switch worsened the recorded
signals.

30
Figure 25: Potato configuration with stimulation from
Neurochip. During switching the recording lead is
connected to ground as in Figure 24.
A: Channel 1 with no switching.
B: Channel 2 with no switching.
C: Channel 1 with switching.
D: Channel 2 with switching.
Figure 26: Potato configuration with stimulation from
Neurochip and a 397.2016Hz 10mVpp sinewave also
inserted into the potato via the stimulation leads. During
switching the recording lead is connected to ground as in
Figure 24.
A: Channel 1 with no switching.
B: Channel 2 with no switching.
C: Channel 1 with switching.
D: Channel 2 with switching.

31
These figures suggest that the voltage-divider switching circuit is not better than the others in terms
of introducing artifact. Currently we are not aware of any switch-based circuit methods that could
reliably reduce stimulation artifact entering the amplifier without also introducing significant
switching noise.
Investigation of other types of circuits that may reduce the size of stimulation artifact entering the
amplifier without the use of switches is warranted. One proposed idea is the use of varistors to
create a circuit that shunts to ground when voltage exceeds a certain level.
Solutions involving circuitry additions to the NC3 do raise concerns regarding redesign of the
device. In its current design, the NC3 does not have much empty board space, and would require
a significant redesign, possibly resulting in increased weight and size to include additional circuit
components. Current draw must also be considered for addition of components requiring power.
4.4 Neurochip alterations
Some alterations to the existing pre-amplifier circuitry were investigated. Larry Shupe modified
an amplifier board with the following changes:
• VESD set to ground (cut trace from R52/C51)
• Channels 1 through 16: pre-amplifier capacitors bypassed
• Channels 1 through 4: 100mOhm resistors replaced with 1MOhm resistors
These alterations were testing via potato stimulation recordings with the modified board and with
an unmodified board. Recordings were made subsequently with the same potato connections and
same neurochip, the amplifier board removed and replaced between recordings. It is notable that
because the amplifier boards are changed, each of the two testing conditions uses a different Intan
chip. Therefore, because of saturation-condition channel inconsistency, it is difficult to compare
artifacts generated on each board. Ultimately, all channels demonstrated similar intensity of
artifacts on both boards. These alterations have not been tested with an animal.
Another NC3 unit was modified to set the tissue ground and VESD (electrostatic discharge
protection voltage) to battery ground. This device was tested with the potato setup, and no notable
reduction in recovery artifact size or propensity for saturation was seen. The fast settle feature was

32
also tested with this modified unit in the potato, and no difference was apparent. This alteration
has not yet been tested with an animal.
Due to saturation inconsistency between amplifier channels, it is difficult to accurately compare
recordings made on separate devices. It is possible that additional testing including neural
recordings would better elucidate the effects of these circuitry alterations.
Discussion
The overarching goal of developing a definitive solution to the Neurochip-3 recovery artifact
problem was not achieved. However, progress was made through several efforts. A model
describing the generation and behavior of the recovery artifact as a function of recording and
stimulation parameters was developed. Unanticipated attributes of artifact behavior were
discovered and described, namely considerable inconsistencies between behavior of different
amplifier channels when under saturation conditions. The effects of the impedances and spatial
relationships of cortical electrodes were investigated and largely discounted as major factors in
channel saturation. Another neural stimulation and recording device that utilizes the Intan
amplifier showed similar recovery artifacts when recording under the same conditions that
generate artifacts in the NC3, showing the problem is not unique to the NC3. Attempts to reduce
the size of artifact were made, including development of circuitry to reduce the amount of
stimulation voltage entering the amplifier and thereby prevent saturation. However, this method
introduces another artifact into the system that largely negates the positive effect. Alterations to
existing NC3 circuitry were also investigated for potential reduction of artifact recovery time but
shown to have no effect in bench testing. This compendium of information will serve the Fetz
group as it moves forward in developing solutions to this problem.
The bottom line of the NC3 recovery artifact problem is that the Intan amplifier is not well suited
for recording low-frequency neural activity directly following stimulation. However, recording
spike activity with a high pass filter of about 300 Hz would encounter the briefest artifacts, on the
order of 1-5ms on the worst of channels. No promising solutions involving circuit alterations have
been uncovered at this time, although circuits custom-designed to minimize the artifact by
subtracting signals at the input could be further explored. Replacement of the Intan chip with an
alternative amplifier should be considered in designing future iterations of the Neurochip.

33
Acknowledgements
The author would like to thank everyone in the Fetz lab group for their assistance, support, and
feedback throughout this project, including and not limited to Eberhard Fetz, Larry Shupe, Irene
Rembado, Lucia Shumaker, Richy Yun, and Hugo (Monkey H). Invaluable assistance was also
provided by Steve Perlmutter, Frank Miles, and Sung Lee. Additional thanks go to members of the
Rudell lab, including John Uehlin, Chris Rudell, and Giorgio Minai, for graciously providing their
electrical engineering expertise.
References
Jackson, A., Mavoori, J., & Fetz, E.E. (2006). Long-term motor cortex plasticity induced by an
electronic neural implant. Nature, 444, 56-60.
Moritz, C. T., Perlmutter, S. I., & Fetz, E. E. (2008). Direct control of paralyzed muscles by cortical
neurons. Nature, 456(7222), 639–642. http://doi.org/10.1038/nature07418
Nishimura, Y., Perlmutter, S., Eaton, R., & Fetz, E. (2013). Spike-Timing-Dependent Plasticity in
Primate Corticospinal Connections Induced during Free Behavior. Neuron, 80(5), 1301-
1309. doi:10.1016/j.neuron.2013.08.028
Rembado, I., Zanos, S., & Fetz, E. E. (2017). Cycle-Triggered Cortical Stimulation during Slow
Wave Sleep Facilitates Learning a BMI Task: A Case Report in a Non-Human
Primate. Frontiers in Behavioral Neuroscience, 11. doi:10.3389/fnbeh.2017.00059
Zanos, S., Richardson, A. G., Shupe, L., Miles, F. P., & Fetz, E. E. (2011). The Neurochip-2: An
Autonomous Head-Fixed Computer for Recording and Stimulating in Freely Behaving
Monkeys. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 19(4),
427-435. doi:10.1109/tnsre.2011.2158007
Zanos S, Rembado I, Chen D-F, Fetz EE, Phase-locked stimulation during cortical beta
oscillations produces bidirectional synaptic plasticity in awake monkeys. Current Biology
28, 2515–2526, 2018.
Additional information describing the NC3 can be found at the following sites:
http://csne-erc.org/research-dissemination/neurochip
https://depts.washington.edu/fetzweb/neurochip3.html

34
Appendix
Figure A: The Neurochip-3 with dimensions.
Figure B: The NC recording board using the RHD2216 Intan biophysical signal amplifier. Two different
configurations of 16 differential inputs or 32 unipolar inputs are possible. Signals are sampled at up to
20kHz with 16-bit precision and 196x fixed gain. Selectable filter settings include 0.1 Hz to 500Hz high
pass and 100 Hz to 20 kHz low pass filters.

35
Figure C: The NC stimulator board. Six outputs allow for either single-ended or pairwise differential
stimulation. Allows asymmetric bi-phasic pulses. Has 60 Volt compliance and can accommodate up to 1ms
phase width.

36
Figure D1: Implant impedances for Monkey H channels 49 through 64, collected on 5/17/2019.
Impedances are given in MΩ, and the x-axis is a base-10 log of the tested frequencies, which are 1, 10, 50,
100, 200, 500, 1000, and 2000Hz.

37
100, 200, 500, 1000, and 2000Hz.

38
100, 200, 500, 1000, and 2000Hz.

39
Samtec TDT ch sites Samtec TDT ch Sites Samtec TDT ch Sites
A01 1 GND3 B01 33 LEOG-Top C01 65 LEOG-side
A02 2 GND2 B02 34 LPFC(31,6)S C02 66 LSM(16,10)S
A03 3 RSM(7,9)S B03 35 NC C03 67 GND5
A04 4 RSM(7,9)D B04 36 LPFC(31,6)D C04 68 LSM(16,10)D
A05 5 RSM(10,12)D B05 37 REOG-Top C05 69 NC
A06 6 RSM(7,3)S B06 38 LPFC(34,6)S C06 70 LSM(13,10)S
A07 7 RSM(10,12)S B07 39 GND4 C07 71 NC
A08 8 RPC(-6,9)S B08 40 LPFC(34,6)D C08 72 LSM(13,10)D
A09 9 RSM(7,3)D B09 41 RPFC(34,6)S C09 73 LSM(16,13)S
A10 10 RSM(16,7)S B10 42 LPFC(34,9)S C10 74 LSM(13,13)D
A11 11 RPC(-6,9)D B11 43 RPFC(34,12)S C11 75 LSM(16,13)D
A12 12 RSM(16,7)D B12 44 LPFC(34,9)D C12 76 LSM(13,13)S
A13 13 RPC(-12,3)S B13 45 RPFC(34,6)D C13 77 LSM(7,6)S
A14 14 RSM(13,7)D B14 46 LPFC(31,9)D C14 78 LPC(-6,9)S
A15 15 RPC(-3,3)D B15 47 RPFC(34,12)D C15 79 LSM(7,6)D
A16 16 RSM(13,10)S B16 48 RPFC(31,3)D C16 80 LPC(-6,9)D
A17 17 RPC(-3,3)S B17 49 RPFC(31,3)S C17 81 LPC(-3,3)S
A18 18 RSM(7,6)D B18 50 LPFC(31,9)S C18 82 LPC(-3,3)D
A19 19 RPC(-12,3)D B19 51 LPFC(28,12)S C19 83 LPC(-6,6)D
A20 20 RSM(13,10)D B20 52 LPFC(28,6)D C20 84 LPC(-6,6)S
A21 21 RSM(13,7)S B21 53 LPFC(28,12)D C21 85 LPC(0,6)D
A22 22 RPC(-3,6)S B22 54 LPFC(28,6)S C22 86 LPC(0,6)S
A23 23 RSM(16,10)S B23 55 *LPFC(31,12)D C23 87 LSM(7,9)S
A24 24 RPC(-3,6)D B24 56 *LPFC(31,12)S C24 88 LSM(7,9)D
A25 25 RSM(16,10)D B25 57 RPFC(34,9)D C25 89 LSM(10,12)S
A26 26 RPC(-6,6)S B26 58 RPFC(34,9)S C26 90 LSM(10,12)D
A27 27 RSM(10,6)S B27 59 *RPFC(34,3)S C27 91 LSM(7,3)S
A28 28 RPC(-6,6)D B28 60 *LPFC(34,3)D C28 92 LSM(7,3)D
A29 29 RSM(10,6)D B29 61 RPFC(31,6)S C29 93 LPC(-3,6)D
A30 30 RPC(-12,6)S B30 62 *LPFC(34,3)S C30 94 LPC(-3,6)S
A31 31 RSM(7,6)S B31 63 RPFC(31,6)D C31 95 LPC(0,3)D
A32 32 RPC(-12,6)D B32 64 *RPFC(34,3)D C32 96 LPC(0,3)S
A33/R1 GND1 B33
GND
(pedestal) C33/R2 GND6
A34
GND
(pedestal) B34
GND
(pedestal) C34
GND
(pedestal)
A35
GND
(pedestal) B35
GND
(pedestal) C35
GND
(pedestal)
A36 NC B36 NC C36 NC

40
Figure E: Cortical channel ordering for monkey H with Cereport connector mapping.

41
Figure F: Cortical electrode mapping for Monkey H. Downward view of head, top of image closest to the
animal’s face.

Improvement of a Bidirectional Brain-Computer Interface for Neural Engineering Research

More Related Content

What's hot (20)

Similar to Improvement of a Bidirectional Brain-Computer Interface for Neural Engineering Research (20)

Recently uploaded (20)

Improvement of a Bidirectional Brain-Computer Interface for Neural Engineering Research