Uploaded by l.rohan

A CMOS-Based Bidirectional Brain Machine Interface System With Integrated fdNIRS and tDCS for Closed-Loop Brain Stimulation

A CMOS-Based Bidirectional Brain Machine
Interface System With Integrated fdNIRS and tDCS
for Closed-Loop Brain Stimulation
Yun Miao
, Member, IEEE and Valencia Joyner Koomson, Member, IEEE
Abstract—A CMOS-based bidirectional brain machine interface
system with on-chip frequency-domain near infrared spectroscopy
(fdNIRS) and transcranial direct-current stimulation (tDCS) is
designed to enable noninvasive closed-loop brain stimulation for
neural disorders treatment and cognitive performance enhancement. The dual channel fdNIRS can continuously monitor absolute
cerebral oxygenation during the entire tDCS process by measuring NIR light’s attenuation and phase shift across brain tissue.
Each fdNIRS channel provides 120 dBΩ transimpedance gain at
80 MHz with a power consumption of 30 mW while tolerating up
to 8 pF input capacitance. A photocurrent between 10 and 450 nA
can be detected with a phase resolution down to 0.2°. A lensless system with subnanowatt sensitivity is realized by using an avalanche
photodiode. The on-chip programmable voltage-controlled resistor
stimulator can support a stimulation current from 0.6 to 2.2 mA
with less than 1% variation, which covers the required current
range of tDCS. The chip is fabricated in a standard 130-nm CMOS
process and occupies an area of 2.25 mm2 .
Index Terms—Brain machine interface, CMOS, frequency
domain NIRS, optical sensing, tDCS.
RANSCRANIAL direct-current stimulation (tDCS) is a
non-invasive neuro-modulation technique, which has a
great potential in treating neurological diseases and enhancing motor and cognitive performance [1]. During tDCS a weak
current, typically between 0.5 to 2 mA, is delivered to brain by
electrodes on the scalp to facilitate or inhibit neural activities.
Compared with commonly used brain stimulation methods such
as deep brain stimulation (DBS) and transcranial magnetic stimulation (TMS), which requires brain surgery or large interface
coils, tDCS provides a non-invasive interface with relatively
small footprint. Previously reported experimental data and clinical studies have shown the benefits of tDCS for a variety of
neurological diseases and disorders such as depression, stroke,
aphasia, chronic pain, Alzheimer’s, and Parkinson’s [2]–[7]. The
potential to enhance the performance of healthy subjects in motor and cognitive domains has also been demonstrated [8]. To
Manuscript received September 25, 2017; revised November 28, 2017 and
January 16, 2018; accepted January 18, 2018. Date of publication March 1,
2018; date of current version June 5, 2018. This paper was recommended by
Associate Editor M. Kalofonou. (Corresponding author: Yun Miao.)
The authors are with the Department of Electrical and Computer Engineering, Tufts University, Medford, MA 02155 USA (e-mail: yun.miao@tufts.edu;
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TBCAS.2018.2798924
date most tDCS experiments and studies follow an open-loop
manner, where a predefined current is applied to subjects in
the same group for a predefined period. The open-loop manner
does not compensate for user variabilities, such as tissue resistance, skull defects, and baseline cortical excitability [9]. As
a result, a significant inter-subject variability and intra-subject
variability is observed [10]–[12], which indicates a real-time,
programmable stimulation strategy and dosage is desired for
each user. A variety of devices utilizing closed-loop stimulation are developed to enhance the performance of tDCS; and
the majority employ electroencephalogram (EEG) techniques
[13], [14]. EEG devices have several advantages, including low
cost and low power, however, several engineering challenges are
presented. Firstly, as both EEG acquisition and tDCS operate in
electrical domain the cross-coupling interference between EEG
and tDCS makes real-time brain monitoring very challenging.
Furthermore, for safety concerns large sized sponge electrodes
are widely used in tDCS to control current density, which are
not suitable for EEG acquisition and can block EEG caps at
stimulation sites. In addition, tDCS sponge electrodes introduce
large artifacts that prevent EEG recoding [15].
NIRS techniques enable monitoring of brain oxygenation by
measuring the optical absorption and scattering properties of
brain tissue. Generally, it has a better spatial resolution than
EEG, a lower cost than MRI, and has been used in functional
brain studies for decades [16]. As an optical sensing method
NIRS can solve the incompatibility between EEG and tDCS in
both time and spatial domains. As shown in Fig. 1. NIRS can
continuously monitor brain oxygenation without interference
with tDCS through the entire brain stimulation process. Optical fibers can reach the scalp directly through small holes on
sponge electrodes giving us more flexibility in the placement
of stimulation and recording sites. NIRS does have a limited
penetration depth and inferior time resolution than EEG, but in
our applications, they are not regarded as major disadvantages
since tDCS is supposed to mainly affect the cerebral cortex in a
time scale of seconds to minutes [17], [18]. Several continuous
wave NIRS (cwNIRS) based devices were already reported and
experimental results showed a great potential for NIRS based
closed-loop brain stimulation [19], [20].
In this paper, a CMOS based bidirectional brain machine
interface system with on chip fdNIRS and tDCS is described.
Unlike cwNIRS, from which only the relative change of oxygenation is extracted from the amplitude of received signal,
1932-4545 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.
Authorized licensed use limited to: WASHINGTON UNIVERSITY LIBRARIES. Downloaded on October 09,2023 at 03:15:43 UTC from IEEE Xplore. Restrictions apply.
Fig. 3.
Fig. 1. (a) Optical fibers enable NIRS recording at stimulation sites. (b) NIRS
can continuously monitor brain oxygenation during the entire tDCS process
without cross-coupling interference.
System architecture.
frequency domain NIRS measurement. After that brain hemodynamics are estimated and a customized tDCS dosage can be
applied to address the variability among different users. The
operation of the proposed loop can be divided into 3 domains:
optical domain, analog domain, and digital domain. Modern
computing systems can readily provide the computation power
required for digital operations, however, most commercial devices for optical and analog operations in fdNIRS and tDCS are
still based on expensive and bulky instruments or discrete electronics. The proposed system focus on the optical and analog
design of fdNIRS and tDCS toward a miniaturized bidirectional
brain machine interface, which can adjust tDCS dosage according to brain hemodynamics.
B. System Design
Fig. 2.
Closed-loop brain stimulation based on fdNIRS and tDCS.
fdNIRS can capture the absolute value of oxy-hemoglobin
(HbO2 ), deoxy-hemoglobin (HHB), total hemoglobin (tHB),
and brain tissue oxygenation (SO2 ) by measuring both amplitude and phase of the received signal [21]. The quantitative
measurement will not only give us more information to address
the inter-subject and intra-subject variability but also reduce the
crosstalk between light absorption and scattering [22].
The rest of the paper is organized as follows. The system architecture is introduced in Section II. Section III describes the design consideration and implementation of each
building blocks. Section IV presents the characterization and
measurement results. Conclusion and design remarks are made
in section V.
A. A CMOS-Based Bidirectional Brain Machine Interface
Fig. 2 presents the structure of the fdNIRS and tDCS based
brain stimulation loop. Firstly, the absorption coefficient (μa )
and reduced scattering coefficient (μ s ) of brain tissue are extracted from the amplitude and phase of the acquired signal in a
Two different sensing method are widely used in fdNIRS
measurement: broadband method and single tone method. The
mechanism of a broadband system is very similar with a vector
network analyzer. The modulation frequency is swept across a
wide frequency range between tens of MHz to multi GHz [23],
[24]. The absorption coefficient (μa )and reduced scattering
coefficient (μs ) can be acquired in a single sweep at a fixed
source-detector distance. For a single tone system, the light
is modulated at a fixed frequency but measurements are performed at multiple source-detector distance [25]. The broadband
method is very challenging for a low cost and compact design as
it requires light sources and detectors that can work up to multi
GHz; the required high performance frequency synthesizer and
signal processing circuits are very complex and power hungry
as well. In this paper, the single tone multi-distance method is
used for cost and power concerns.
The system architecture is shown in Fig. 3. NIR lights modulated at 80.001 MHz are applied to tissue with a linear spaced
distance from sensor. μa and μs are extracted by feeding the
slope of amplitude-distance curve (SA C ) and phase-distance
curve (Sφ ) to (1) and (2) [25], where ω is the angular frequency
of modulation and v is speed of light in the media.
μa =
Authorized licensed use limited to: WASHINGTON UNIVERSITY LIBRARIES. Downloaded on October 09,2023 at 03:15:43 UTC from IEEE Xplore. Restrictions apply.
μs =
SA2 C − Sφ2
The quantitative oxygenation of brain tissue can be estimated
by applying the above frequency domain multi-distance measurement at different wavelengths in NIR range [26]. In our
design an APD with 1.77 mm2 active area (Hamamatsu S925115) is adopted as the detector. The advantage of a large area APD
is that it can collect sufficient diffusive light from tissue without using any lens or other expensive optics. Both light sources
and the APD are coupled to tissue with optical fibers. A fiber
coupled setup will not only reduce the footprint of interface but
also provide a solid galvanic isolation between tissue and APD,
which is reverse biased at 200 V.
The fdNIRS circuits are composed of two identical channels.
One is connected to the APD for sensing; the other one is driven
by a signal with arbitrary amplitude and phase as reference.
The phase information is acquired by comparing the phase of
sensing channel and reference channel in order to reduce the
influence of temperature, process variation and other less controllable factors. In each channel the photocurrent from APD
is first amplified by a wideband TIA and then filtered by a 4th
order Gm-C bandpass filter to suppress out of band noises. The
output of the filter is down converted to 1 kHz by a modified
Gilbert mixer to alleviate the timing requirement in phase detection and limit noise bandwidth. A computer or a mobile device
can be used to analyze the resulting fdNIRS signal and control
the on-chip tDCS through an integrated 5-bit DAC realizing a
closed-loop stimulation.’
A. Transimpedance Amplifier
The TIA is designed to amplify an 80 MHz low level photocurrent with 100 dBΩ gain, while tolerating the APD’s 3.6 pF
capacitance and other parasitics from test PCB, packages, and
other off chip components. Although some designs utilizing
regulated cascode input stage and inverter cascode based output
stage showed a good noise and power performance [27], a resistive feedback structure is still a more practical choice to interface
the large area APD for bandwidth and input capacitance concerns. For a first order approximation, a TIA’s transimpedance
gain is set by the feedback resistor Rf and the bandwidth is
given by (3), where A is the gain of core voltage amplifier and
Ctotal is the total input capacitance. To achieve the desired gain
and bandwidth the TIA is implemented with a 60 dB gain core
voltage amplifier and 120 kΩ feedback resistors.
f3 dB =
2πRf Ctotal
Besides the gain requirement, the core amplifier needs to satisfy bandwidth requirement for stability concerns. A single stage
amplifier can hardly provide enough gain and bandwidth simultaneously. As shown in Fig. 4 the core amplifier is realized by
cascading 3 gain stages. Each stage provides approximate 20 dB
gain with a bandwidth of around 1GHz, which ensures enough
Fig. 4.
Schematic of the TIA and core voltage amplifier
phase margin and stability. At 2nd and 3rd stages cascode transistors M3 and M4 are inserted to reduce the Miller-capacitance
for bandwidth enhancement. The 1st stage follows a simple
common source topology as its Miller-capacitance can be absorbed by the TIA’s input capacitance and cascode transistors
will introduce excessive noises.
The main noise contributions of the proposed TIA at frequency of interest can be summarized by (4), where k is
Boltzmann’s constant, T is the temperature, gm is the transconductance of input transistors, γ is the process dependent noise
coefficient, Cin is the capacitance of the input transistors, CAPD
is the capacitance of APD, and f represents the bandwidth.
In2 =
(2π (Cin + CAPD ))2 2
+ 4kT γ
By substituting gm with 2πfT Cin we can see the equation get
a minimum when Cin = CAPD [28], [29], however, this design
strategy can lead to many practical problems. Our TIA needs to
interface an APD with 3.6 pF capacitance. If the input transistors are sized with large width toward the 3.6 pF capacitance a
very large biasing current is required to maintain a desired inversion coefficient and avoid the punishment in fT . In addition,
the large input capacitance from input transistors will affect the
TIA bandwidth as well. So, in our design the input transistors are
sized toward a good balance between noise, power, and bandwidth by following the techniques described in [30], [31]. In 1st
stage M1 and M2 are sized with W/L of 240 um/0.24 um and biased in moderate inversion region to provide optimal gm/I, while
Authorized licensed use limited to: WASHINGTON UNIVERSITY LIBRARIES. Downloaded on October 09,2023 at 03:15:43 UTC from IEEE Xplore. Restrictions apply.
Fig. 6.
Fig. 5. (a) Schematic of bandpass Gm-C biquad. (b) Schematic of the
transconductance cell.
maintaining enough fT , bandwidth, linearity, and a reasonable
capacitance matching. 2nd and 3rd stages are designed toward
low power and high bandwidth as their noise contribution is not
as significant as the 1st stage amplifier. The 3-stage core voltage
amplifier’s 3 dB bandwidth is 680 MHz and consumes 7 mA
current from 1.5 V supply. The TIA can provide over 100 dBΩ
transimpedance gain and 180 MHz 3 dB bandwidth with up to
8 pF input
√ capacitance. The simulated input referred noise is
3.2 pA/ Hz at 80 MHz.
B. Filter and Mixer
The TIA is followed by a bandpass filter to suppress out
of band noise from the APD and wideband TIA. The filter
also serves as a gain stage to further amplify the signal against
the mixer’s high noise figure. A 4th order bandpass Gm-C filter is implemented by cascading 2 bandpass biquad shown in
Fig. 5(a). The Gm cell is the most critical building block for a
Gm-C filter. It needs to maintain a good linearity across a wide
input range. A high DC gain is also desired to minimize phase
errors. To achieve high DC gain and good linearity a transconductance cell with tunable negative resistance load and dynamic
source degeneration is designed. As shown in Fig. 5(b) the cross
coupled transistors M7 and M8 realize a negative resistance to
boost the gain. M3 ∼ M6 serve as dynamic source degeneration to improve the linearity. The bandpass filter is centered at
80 MHz with a 3 dB bandwidth of around 40 MHz. The center
frequency can be fine-tuned by an off-chip reference current.
The simulated SFDR is greater than 50 dB with a 60 mV input
Schematic of the folded mixer.
signal. The filter provides 10 dB gain and consumes 10 mA
current from a 1.5 V power supply.
In a frequency domain multi-distance measurement, the
signal level changes significantly from a small to large sourcedetector distance. After passing through TIA and filter the signal’s amplitude is amplified by a factor of over 110 dB so the
linearity and swing become important concerns for mixer design. To ensure good linearity and wide dynamic range a modified Gilbert mixer with folded input stage is implemented with
3.3 V I/O devices. As shown in Fig. 6 the folded input stage
with source degeneration can handle the relative low common
mode voltage from filter with good linearity and dynamic range.
M9 ∼ M12 serve as current sink to limit the DC current through
switches and load resistors. For a Gilbert mixer, the source degeneration resistor RS should be large enough to linearize the
input transconductance under large input signals, while its conversion gain is proportional to RL /RS , where RL is the load
resistor. In a conventional Gilbert mixer RS and RL share the
same DC path. Good linearity and conversion gain can be hardly
achieved simultaneously due to limited supply voltage. In this
design, as the current through RL is limited by the current sink
10 kΩ load resistors can be used to achieve large conversion
gain without compromise in linearity. The current sink can also
reduce the flicker noise form switches, which is related to the
DC current. The mixer achieved a conversion gain of 10 dB with
less than 5 mW power consumption.
C. Stimulator
The stimulator needs to tolerate the huge variation of electrode impedance, skin impedance, electrode skin contacts and
other less controllable factors across different users and conditions. An up to 2 mA full scale current is also desired to
ensure the effectiveness of stimulation. The output impedance
and compliance voltage are of the most important parameters
to ensure an accurate dosage of stimulation current. On the
contrary, the benefit of resolution and linearity can be easily
diluted by the huge variation across users and interfaces, which
makes them less important parameters. In addition, for most
tDCS applications the stimulation current is varied in the order
Authorized licensed use limited to: WASHINGTON UNIVERSITY LIBRARIES. Downloaded on October 09,2023 at 03:15:43 UTC from IEEE Xplore. Restrictions apply.
Fig. 7.
Schematic of the VCR current source.
of 0.1 mA to 1 mA. Chasing unnecessary resolution and linearity will not bring us much practical benefits. In our design
a voltage controlled resistor (VCR) current source is adapted
from [32] toward an optimized output impedance, compliance
voltage, and full scale current. Although its linearity is inferior to conventional current steering stimulators, a high output
impedance and large compliance voltage range can be achieved.
The schematic is shown Fig. 7 M1 works in deep linear region as a voltage controlled resistor. M2 and an error amplifier
realize a feedback loop to regulate V1 to 0.15 V and boost
the output impedance. Its output resistance and output current
can be described by (5), (6), and (7), where Ae is the gain of
the error amplifier, Gds1 is the equivalent drain-source conductance of M1, gm 2 is the transconductance of M2, and rds2 is the
drain-source resistance of M2.
Iout = V 1 · Gds1
Rout =
Ae gm 2 rds2
Gds1 = μn Cox
(Vtune − Vth )
As M1’s Vds is well regulated the output current is linearly
modulated by Vtune . M3 ∼ M5 are used to compensate the
reduction of Gds1 due to velocity saturation and other second
order effects at high Vg s . They are biased with a resistive voltage
divider connected to the 5-bit current steering DAC and turn
on one by one with the increased Vtune . The error amplifier is
implemented by a folded-cascode structure with 65 dB DC gain,
which boosts the output impedance to MΩ level. Compared with
conventional current steering stimulator with cascode output
stage this stimulator can achieve a comparable output impedance
with larger compliance voltage. In addition, in this topology only
M1 and M2 are expected to tolerate the high voltage of tDCS. If
a high voltage process is available a fully integrated solution can
be realized by using only 2 high voltage LDMOS, which will
result in a huge reduction in layout area and power consumption
Fig. 8.
Chip micrograph.
compared with conventional current steering topology. In this
design, the stimulator is implemented with 3.3 V I/O devices in
a standard 130 nm CMOS process. An off-chip buffer FET is
used to protect the chip from the high voltage of tDCS.
A. fdNIRS Characterization and Measurements
The chip micrograph is shown in Fig. 8. The fdNIRS system is
highly mixed mode (optical and analog) and is supposed to sense
a weak diffusive light with large dynamic range. To avoid using
complex and expensive optical instruments in measurements the
chip’s dynamic range and phase resolution are characterized in
electrical domain. Optical measurements are operated on both a
solid tissue phantom and liquid tissue phantoms to further verify
its functionality and evaluate system level errors.
1) fdNIRS Dynamic Range Characterization: In the electrical test setup, an RF voltage source (Rigol DG4000) and a
passive APD model composed of a 100 kΩ series resistor and a
6 pF shunt capacitor is used to mimic the behavior of the large
area APD. As the TIA has a low input impedance around 100
Ω the passive APD model can reliably inject high frequency
current into the chip with low errors. The output of the chip is
first buffered by off-chip buffers with 3 dB voltage gain (Texas
Instrument INA333) then digitized by a DAQ card (National
Instrument USB6259). The acquired signal is transferred and
stored in a computer for further analysis and processing.
The dynamic range of the fdNIRS system was characterized
by an input current versus output voltage measurement. During the measurement current modulated at 80.001 MHz was
applied to the fdNIRS’ input. Mixer’s LO was set to 80 MHz
resulting in a 1 kHz signal at the output. The signal was then
sampled at 200 kHz with an oversampleing rate of 100 to reduce the quantization noise and minimize noise folding. Output
voltage at different input current amplitude is demonstrated in
Authorized licensed use limited to: WASHINGTON UNIVERSITY LIBRARIES. Downloaded on October 09,2023 at 03:15:43 UTC from IEEE Xplore. Restrictions apply.
Fig. 9.
fdNIRS output voltage versus input current.
Fig. 9. For each input current the output voltage is averaged for
0.1 second. We can see that the fdNIRS system can provide a
linear transimpedacne gain over 120 dBΩ between 10 nA to
450 nA. For an input current between 5 nA and 10 nA the gain
error was still below 1 dB but the degraded SNR would lead to
larger phase errors, which is discussed in the following section.
2) fdNIRS Phase Characterization: To characterize the
fdNIRS system’s phase resolution an input phase difference
versus output phase difference measurement was performed.
In the measurement, the signal’s phase in sensing channel was
changed by certain steps while keeping the reference channel’s
phase constant. To improve the system’s noise tolerance and
alleviate the timing requirement the digitized signal was first
smoothed with least-squares fit in MATLAB [24]. The phase
difference was then extracted from the correlation function
between sensing and reference channel.
The phase measurement results with 20 nA input current are
shown in Fig. 10. At each step the phase difference is averaged
for 0.1 second. It shows that the chip can support 360° phase
detection and a phase difference of 0.2° can be detected with a
standard deviation of less than 0.15° and a DNL of less than 0.4
LSB (0.08°). The phase measurement was repeated at different
input current level as well. The 0.2° phase resolution can be
maintained between an input range of 20 nA to 450 nA and
decreased to around 0.5° at 10 nA. For current below 10 nA the
phase resolution degrades significantly due to the limited SNR.
The key parameters of reported silicon based fdNIRS systems
and this work are summarized in Table I. This chip demonstrated a significant improvement in sensitivity and dynamic
range compared with the system reported in [29]. Although
the power consumption increased due to the use of high order active filters, system level power consumption can be well
compensated by the reduction of required laser power. This
chip matched the sensitivity of the system based on Silicon
Germanium (SiGe) Heterojunction Bipolar Transistor (HBT)
[33]. A significant improvement in phase resolution was also
demonstrated. Although SiGe HBTs have a superior performance in bandwidth, noise, and power consumptions, designs
based on standard CMOS process will largely reduce the cost in
mass productions. The system in [33] achieved a dynamic range
Fig. 10. (a) Input phase difference versus detected phase difference, 0° ∼
360°, 10°/step. (b) Input phase difference versus detected phase difference, 0°
∼ 2°, 0.2°/step.
Dynamic Range
Sthalekar [29]
Sthalekar [33]
This Work
180 nm CMOS
130 nm BiCMOS
130 nm CMOS
27 dB
33 dB
60 dB
20 nW
<1 nW
<1 nW
18 mW
18 mW
30 mW
of 60 dB by using a TIA with programmable loads of 100 kΩ,
10 kΩ, and 1 kΩ. Each gain mode demonstrated a dynamic range
of around 23 dB. A programmable load is a very practical way
to extend the dynamic range, however, as fdNIRS systems are
phase sensitive very complex calibrations are required in real
applications to compensate the different phase response across
3 gain modes. Furthermore, in most fdNIRS applications the
power level of received diffusive light is very low. Improving
Authorized licensed use limited to: WASHINGTON UNIVERSITY LIBRARIES. Downloaded on October 09,2023 at 03:15:43 UTC from IEEE Xplore. Restrictions apply.
μ s
μ s
μ s
Fig. 11.
Optical measurement setup.
the dynamic range while maintaining a high gain is desired in
most of the cases. The 10 kΩ and 1 kΩ modes in [33] may only
be beneficial in extreme cases.
3) Solid Tissue Phantom Measurement: To verify the fdNIRS’ functionality and characterize its system level errors optical measurements were performed on a commercial solid tissue
phantom (INO BIOMIMICTM optical phantom) with known
μa and μs . Some primary results at a single wavelength of
785 nm was already ready reported in our previous work [34].
To further evaluate the fdNIRS’ performance we conducted additional measurements at 685 nm and 830 nm, which covered
the most widely used wavelengths in NIR range. The system’s
dynamic performance was also verified with liquid phantom
The optical measurement setup is shown in Fig. 11. NIR light
emitted from commercial laser diodes (Thorlabs, HL6750MG,
L785P090, HL8338MG) was applied to the solid phantom with
optical fiber (1000 um core) and coupled to a large area APD
(Hamamatsu S9251-15) with another optical fiber (1000 um
core). The sensor and source distance was adjusted from 2.0 cm
to 2.5 cm with 0.1 cm/step by a commercial translation stage
(Thorlabs, MTS50A-Z8) for a frequency domain multi-distance
measurement. The optical power was set below 2.5 mW to fit
the laser safety requirement and avoid errors from significant
temperature changes in tissue [35]. At each distance, signals
were recorded for a 0.1 second. The frequency arrangement
for modulation, LO, and sampling rate were same with the
ones used in electrical measurements. The results are shown in
Table II. and compared with the optical parameters characterized
by the manufacture with high accuracy time-resolved transmittance technique [36]. We can see that for all three wavelengths
the error in μa stays below 20.35% and error in μs is less
than 23%.
4) Liquid Tissue Phantom Measurement: An important
merit of fdNIRS system is to capture both the baseline
685 nm
685 nm
785 nm
785 nm
830 nm
830 nm
Expected (cm−1 )
Measured (cm−1 )
Error (%)
σ (cm−1 )
information and dynamic change of tissue’s optical parameters.
To further verify the system’s functionality optical measurements were performed on 2 liquid phantoms with controlled
absorption and reduced scattering coefficient.
In the first measurement 10L of liquid phantom composed of
4:6 parts of milk to water was prepared and put in a container
with dark walls. India ink (Dr. Ph. Martin’s Black Star) was
added to the phantom step by step to change its absorption
coefficient. India ink is a very strong absorption media so it
was pre-diluted to 10% by water to make the dosage more
controllable. It was also ultrasoniced for 20 minutes to reduce
the influence of large sized particles in India Ink [37]. At each
step μa and μs at 830 nm was measured. Laser power was set
to 2.5 mW as in the solid phantom setup but the modulation
depth was reduced to avoid saturation at low ink concentration.
Optical fibers were put inside the liquid phantom directly. An
infinite-media approximation was used in μa and μs extraction.
The rest setup was all the same as solid phantom measurement.
The average results of 5 measurements at each step is shown in
Fig. 12(a). Without ink the measured μa is 0.35 cm−1 , which
is very close to the reported water’s μa of 0.32 cm−1 [38], and
0.29 cm−1 [39]. The measured μa increases linearly with a fitted
slope of 496 cm−1 . The slope is very consistent with previous
reported data of 372 cm−1 , the difference can be a result of ink
variations across different brands and patches [40]. μs fluctuates
within a small range, which is as expected.
To make sure the fdNIRS system can work with different
absorbers another liquid phantom was prepared. Fountain pen
ink (Parker Super Quink Ink) was used as the absorber. The ratio
of milk to water was adjusted to 3:6 to examine the system’s
functionality in μs detection as well. The rest setup was all
the same as previous measurement. As shown in Fig. 12(b)
without ink the measured μa is around 0.33 cm−1 , which is
very close to the first measurement. A linear increment in μa is
observed with addition fountain pen ink and the slope is in the
same order of previous reported data [33]. A slightly lower μs
is also successfully detected, which is the result of lower milk
5) Error Analysis and Comparison: A comparative analysis
of the presented work and other published fdNIRS systems can
be found in Table III. The broadband system reported in [23],
[24] have a very good accuracy of 5% and 14.48% but require
either vector network analyzer or expensive discrete electronics
for the frequency sweeping. High performance APD modules
are also used to detect NIR lights modulated up to 1GHz.The
single tone system reported in [26] achieved a 15% accuracy
Authorized licensed use limited to: WASHINGTON UNIVERSITY LIBRARIES. Downloaded on October 09,2023 at 03:15:43 UTC from IEEE Xplore. Restrictions apply.
Technology Node
Sensing Method
Light Coupling
Maximum Error
Fantini [26]
Pham [23]
No [24]
Sthalekar [29]
Sthalekar [33]
This Work
Instrument Based
Single Tone
120 MHz
3 mm Fiber
APD Module
Instrument Based
10 ∼ 1000 MHz
1 mm Fiber
APD Module
Discrete Electronics
10 ∼ 1000 MHz
Direct Coupling
130 nm biCMOS
Single Tone
80 MHz
1 mm Fiber and
130 nm CMOS
Single Tone
80 MHz
1 mm Fiber
180 nm CMOS
Single Tone
100 MHz
1 mm Fiber and
Objective Lens
reduced. The solid phantom and liquid phantom measurements
showed improvements in error and linearity compared with the
system in [33] as well. It should be noted that unlike the results
reported in [33], where the error increases with the μa of the
solid phantom, this chip exhibits less than 23% error on a solid
phantom with μa of 0.12 cm−1 . The solid phantom used for
experimental tests is fabricated with an absorption coefficient
that is 2 to 9 times larger than previously used solid phantoms in
[33] and is more consistent with the optical property of human
forehead [41], [42]. A lower than expected μa and higher than
expected μs is observed at all three wavelengths in solid phantom measurements. The measurement error for both coefficients
is lower for the liquid phantoms compared to the solid phantom.
Part of the error in solid phantom measurement is likely to be
an offset from the fiber-phantom interface. In the solid phantom
measurement, the fibers were attached to the phantom surface
through fiber holder arms (Thorlabs, PRA-SMA), while the
fibers were put into the liquid directly during liquid phantom
measurements. The holder arms were supposed to keep a good
contact and a constant 90° incident angle against frictions between the fiber tip and phantom surface during source-detector
distance adjustment. The holder arms have a considerable foot
print of around 4 cm by 7 cm, which may introduce undesired
surface reflection and affect the semi-infinity media assumption. The error can be potentially reduced with the calibration
technique mentioned in [24] or using a customized probe with
smaller footprint and absorbent materials [26].
B. Stimulator Measurement
Fig. 12. (a) Measured absorption coefficient and reduced scattering coefficient
at different India ink concentration. (b) Measured absorption coefficient and
reduced scattering coefficient at different fountain pen ink concentration.
by utilizing a photomultiplier tube (PMT). The PMT requires a
very high bias voltage and is sensitive to mechanical shock and
external magnetic fields, which limited the system’s mobility
and robustness. Our chip can work with a standard APD and has
a comparable accuracy with systems employing bulky instruments or discrete electronics. An improvement in gain, dynamic
range, phase resolution, and sensitivity is achieved compared
with other integrated solutions [29], [33]. In addition, as no objective lens or collimator is used, the system’s cost is largely
The I-V curve of the on-chip stimulator at different input
codes was measured with a high accuracy source meter (Keithley 6430). From Fig. 13. we can see that from 0.4 V to 2.5 V
for current below 1.5 mA the output impedance stays above 1
MΩ and for current between 1.5 mA and 2.2 mA the minimum
impedance is still greater than 781 kΩ. For each input code
the current variation form 0.4 V to 3.0V is less than 1%. Its
current transfer curve with an off-chip buffer FET (Diodes Incorporated BSS138) was also measured under 20V supply and
8 kΩ resistive load to mimic the voltage and load condition of
tDCS. As shown in Fig. 14. the 5-bit programmable stimulator
can generate a current form 0.6 mA to 2.2 mA with a DNL
less than 0.4LSB, which covers the required range of tDCS. Its
linearity and resolution is inferior to the stimulator in [13] due
to the mechanism of VCR current source, but the large full scale
Authorized licensed use limited to: WASHINGTON UNIVERSITY LIBRARIES. Downloaded on October 09,2023 at 03:15:43 UTC from IEEE Xplore. Restrictions apply.
dosage and strategy can be applied to users with the help of simultaneous brain monitoring and stimulation.
The on-chip fdNIRS achieved a sub nW sensitivity with a
large area APD. The improvement in gain, dynamic range, and
phase resolution make a lensless setup possible. The solid phantom measurement showed an error less than 23%, which is comparable with instrument based prototypes. Both baseline information and dynamic change in 2 liquid phantoms were successfully captured as well. The accuracy can be potentially enhanced
with calibration techniques and customized optical probes. The
integrated tDCS can supply a current between 0.6 mA to 2.2 mA
with a voltage down to 0.4 V. The system highly reduced the
cost and size of fdNIRS and tDCS. The proposed system will
benefit the use of closed-loop brain stimulation in both research
and clinic applications.
Fig. 13.
Fig. 14.
Stimulator I-V curve at different input codes.
Stimulator output current versus input codes.
current, large compliance voltage, and high output resistance
will highly improve the power efficiency and reliability of tDCS.
Electrodes wrapped in saline soaked sponge are widely used
in tDCS instruments. A recent study on a commercial tDCS
system showed that the electrode impedance can hardly exceed 40 kΩ before stimulation and will drop to around 5 kΩ
soon after the start of stimulation [43]. As a compact and
robust stimulator, our VCR current source is well suited for
standard tDCS applications. However, with growing interest in
high-definition tDCS (HD-tDCS) miniaturized electrodes with
higher impedance may become more popular in the future. More
advanced feedback structures [44], [45] can be adopted to boost
the output impedance in future designs.
This paper presents a CMOS based bidirectional brain machine interface system with integrated fdNIRS and tDCS for
closed-loop brain stimulation. The combination of fdNIRS and
tDCS solves the incompatibility between EEG and fdNIRS in
both spatial and time domains. A real-time, programmable tDCS
[1] M. A. Nitsche et al., “Transcranial direct current stimulation: State of the
art 2008,” Brain Stimulation, vol. 1, no. 3, pp. 206–223, 2008.
[2] M. A. Nitsche, P. S. Boggio, F. Fregni, and A. Pascual-Leone, “Treatment of depression with transcranial direct current stimulation (tDCS):
A review,” Experimental Neurology, vol. 219, no. 1, pp. 14–19,
[3] J. M. Baker, C. Rorden, and J. Fridriksson, “Using transcranial directcurrent stimulation to treat stroke patients with aphasia,” Stroke, vol. 41,
no. 6, pp. 1229–1236, 2010.
[4] J. Fridriksson, J. D. Richardson, J. M. Baker, and C. Rorden, “Transcranial
direct current stimulation improves naming reaction time in fluent aphasia:
A double-blind, sham-controlled study,” Stroke, vol. 42, no. 3, pp. 819–
821, 2011.
[5] F. Fregni, S. Freedman, and A. Pascual-Leone, “Recent advances in the
treatment of chronic pain with non-invasive brain stimulation techniques,”
Lancet Neurology, vol. 6, no. 2, pp. 188–191, 2007.
[6] R. Ferrucci et al., “Transcranial direct current stimulation improves recognition memory in Alzheimer disease,” Neurology, vol. 71, no. 7, pp. 493–
498, Apr. 2008.
[7] P. S. Boggio et al., “Effects of transcranial direct current stimulation on
working memory in patients with Parkinsons disease,” J. Neurological
Sci., vol. 249, no. 1, pp. 31–38, 2006.
[8] L. Jacobson, M. Koslowsky, and M. Lavidor, “tDCS polarity effects in
motor and cognitive domains: a meta-analytical review,” Experimental
Brain Res., vol. 216, no. 1, pp. 1–10, 2012.
[9] A. R. Brunoni et al., “Clinical research with transcranial direct current
stimulation (tDCS): Challenges and future directions,” Brain Stimulation,
vol. 5, no. 3, pp. 175–195, 2012.
[10] S. Wiethoff, M. Hamada, and J. C. Rothwell, “Variability in response to
transcranial direct current stimulation of the motor cortex,” Brain Stimulation, vol. 7, no. 3, pp. 468–475, 2014.
[11] J. C. Horvath, O. Carter, and J. D. Forte, “Transcranial direct current stimulation: Five important issues we aren’t discussing (but probably should
be),” Frontiers Syst. Neurosci., vol. 8, pp. 2–9, 2014.
[12] F. Fregni et al., “Regulatory considerations for the clinical and research
use of transcranial direct current stimulation (tDCS): Review and recommendations from an expert panel,” Clinical Res. Regulatory Affairs,
vol. 32, no. 1, pp. 22–35, Feb. 2015.
[13] T. Roh, K. Song, H. Cho, D. Shin, and H.-J. Yoo, “A wearable neurofeedback system with EEG-based mental status monitoring and transcranial electrical stimulation,” IEEE Trans. Biomed. Circuits Syst., vol. 8,
no. 6, pp. 755–764, Dec. 2014.
[14] M. A. B. Altaf, C. Zhang, and J. Yoo, “A 16-channel patient-specific
seizure onset and termination detection SoC with impedance-adaptive
transcranial electrical stimulator,” IEEE J. Solid-State Circuits, vol. 50,
no. 11, pp. 2728–2740, Nov. 2015.
[15] M. D. Johnson et al., “Neuromodulation for brain disorders: Challenges
and opportunities,” IEEE Trans. Biomed. Eng., vol. 60, no. 3, pp. 610–624,
Mar. 2013.
[16] M. Ferrari and V. Quaresima, “A brief review on the history of human
functional near-infrared spectroscopy (fNIRS) development and fields of
application,” NeuroImage, vol. 63, no. 2, pp. 921–935, 2012.
Authorized licensed use limited to: WASHINGTON UNIVERSITY LIBRARIES. Downloaded on October 09,2023 at 03:15:43 UTC from IEEE Xplore. Restrictions apply.
[17] N. Lang et al., “How does transcranial DC stimulation of the primary
motor cortex alter regional neuronal activity in the human brain?,” Eur. J.
Neurosci., vol. 22, no. 2, pp. 495–504, 2005.
[18] P. C. Miranda, M. Lomarev, and M. Hallett, “Modeling the current distribution during transcranial direct current stimulation,” Clinical Neurophysiology, vol. 117, no. 7, pp. 1623–1629, 2006.
[19] R. Mckendrick, R. Parasuraman, and H. Ayaz, “Wearable functional near
infrared spectroscopy (fNIRS) and transcranial direct current stimulation
(tDCS): Expanding vistas for neurocognitive augmentation,” Frontiers
Syst. Neurosci., vol. 9, Sep. 2015..
[20] U. Ha, Y. Lee, H. Kim, T. Roh, J. Bae, C. Kim, and H.-J. Yoo, “A wearable EEG-HEG-HRV multimodal system with simultaneous monitoring
of tES for mental health management,” IEEE Trans. Biomed. Circuits
Syst., pp. 758–766, Dec. 2015.
[21] M. Calderon-Arnulphi et al., “Detection of cerebral ischemia in neurovascular surgery using quantitative frequency-domain near-infrared spectroscopy,” J. Neurosurgery, vol. 106, no. 2, pp. 283–290, 2007.
[22] V. Toronov et al., “The roles of changes in deoxyhemoglobin concentration and regional cerebral blood volume in the fMRI BOLD signal,”
NeuroImage, vol. 19, no. 4, pp. 1521–1531, 2003.
[23] T. H. Pham, O. Coquoz, J. B. Fishkin, E. Anderson, and B. J. Tromberg,
“Broad bandwidth frequency domain instrument for quantitative tissue
optical spectroscopy,” Rev. Scientific Instrum., vol. 71, no. 6, pp. 2500–
2513, 2000.
[24] K. S. No and P. Chou, “Mini-FDPM and heterodyne mini-FDPM: Handheld non-invasive breast cancer detectors based on frequency-domain photon migration,” IEEE Trans. Circuits Syst. I: Regular Papers, vol. 52,
no. 12, pp. 2672–2685, Dec. 2005.
[25] S. Fantini, M. A. Franceschini, and E. Gratton, “Semi-infinite-geometry
boundary problem for light migration in highly scattering media: A
frequency-domain study in the diffusion approximation,” J Opt. Soc. Amer.
B, vol. 11, no. 10, p. 2128, Jan. 1994.
[26] S. Fantini, “Frequency-domain multichannel optical detector for noninvasive tissue spectroscopy and oximetry,” Opt. Eng., vol. 34, no. 1, pp. 32–42,
Jan. 1995.
[27] M. Atef, A. Atef, and M. Abbas, “Low-power transimpedance amplifier
for near infrared spectroscopy,” in Proc. IEEE Int. Symp. Circuits Syst.,
2016, pp. 2423–2426.
[28] E. Sackinger, Transimpedance Amplifiers. New York, NY, USA: Wiley,
2005, pp. 105–158.
[29] C. C. Sthalekar and V. J. Koomson, “A CMOS sensor for measurement
of cerebral optical coefficients using non-invasive frequency domain near
infrared spectroscopy,” IEEE Sensors J., vol. 13, no. 9, pp. 3166–3174,
Sep. 2013.
[30] M. Ingels and M. Steyaert, “A 1-Gb/s, 0.7-μm CMOS optical receiver
with full rail-to-rail output swing,” IEEE J. Solid-State Circuits, vol. 34,
no. 7, pp. 971–977, Jul. 1999.
[31] D. M. Binkley, B. J. Blalock, and J. M. Rochelle, “Optimizing drain
current, inversion level, and channel length in analog CMOS design,”
Analog Integr. Circuits Signal Process., vol. 47, no. 2, pp. 137–163, Oct.
[32] M. Ghovanloo and K. Najafi, “A compact large voltage-compliance high
output-impedance programmable current source for implantable microstimulators,” IEEE Trans. Biomed. Eng., vol. 52, no. 1, pp. 97–105, Jan.
[33] C. C. Sthalekar, Y. Miao, and V. J. Koomson, “Optical characterization of
tissue phantoms using a silicon integrated fdNIRS system on chip,” IEEE
Trans. Biomed. Circuits Syst., vol. 11, no. 2, pp. 279–286, Apr. 2017.
[34] Y. Miao and V.J. Koomson, “A silicon based fdNIRS system with integrated tDCS on chip for non-invasive closed-loop neuro stimulation,” in
Proc. IEEE Int. Symp. Circuits Syst., 2017, pp. 1–4.
[35] S. A. Carp, P. Farzam, N. Redes, D. M. Hueber, and M. A. Franceschini, “Combined multi-distance frequency domain and diffuse correlation spectroscopy system with simultaneous data acquisition and real-time
analysis,” Biomed. Opt. Express, vol. 8, no. 9, p. 3993, Jul. 2017.
[36] J.-P. Bouchard et al., “Uncertainty analysis of time resolved transmittance characterization of solid tissue phantoms,” paper presented at the
Proceedings of Design and Performance Validation of Phantoms Used in
Conjunction With Optical Measurement of Tissue II, San Francisco, CA,
USA, Nov. 2010.
[37] S. J. Madsen, M. S. Patterson, and B. C. Wilson, “The use of India ink
as an optical absorber in tissue-simulating phantoms,” Phys. Med. Biol.,
vol. 37, no. 4, pp. 985–993, Jan. 1992.
[38] D. J. Segelstein, “The complex refractive index of water,” M.S. thesis,
Dept. Phys., Univ. Missouri-Kansas, Kansas City, MO, USA, 1981.
[39] G. M. Hale and M. R. Querry, “Optical constants of water in the 200-nm
to 200-μm wavelength region,” Appl. Opt., vol. 12, no. 3, pp. 555–563,
Mar. 1973
[40] P. D. Ninni, F. Martelli, and G. Zaccanti, “The use of India ink in tissuesimulating phantoms,” Opt. Express, vol. 18, no. 26, p. 26854, Jul. 2010.
[41] J. Choi et al., “Noninvasive determination of the optical properties of adult
brain: Near-infrared spectroscopy approach,” J. Biomed. Opt., vol. 9, no. 1,
pp. 221–229, 2004.
[42] R. Zimmermann, F. Braun, T. Achtnich, O. Lambercy, R. Gassert, and
M. Wolf, “Silicon photomultipliers for improved detection of low light
levels in miniature near-infrared spectroscopy instruments,” Biomed. Opt.
Express, vol. 4, no. 5, p. 659, Mar. 2013.
[43] C. Hahn, J. Rice, S. Macuff, P. Minhas, A. Rahman, and M. Bikson,
“Methods for extra-low voltage transcranial direct current stimulation:
Current and time dependent impedance decreases,” Clinical Neurophysiology, vol. 124, no. 3, pp. 551–556, 2013.
[44] M. Sivaprakasam, W. Liu, M. Humayun, and J. Weiland, “A variable
range bi-phasic current stimulus driver circuitry for an implantable retinal
prosthetic device,” IEEE J. Solid-State Circuits, vol. 40, no. 3, pp. 763–
771, Mar. 2005.
[45] E. Noorsal, K. Sooksood, H. Xu, R. Hornig, J. Becker, and M. Ortmanns, “A neural stimulator frontend with high-voltage compliance and
programmable pulse shape for epiretinal implants,” IEEE J. Solid-State
Circuits, vol. 47, no. 1, pp. 244–256, Jan. 2012.
Yun Miao (M’15) received the B.E. degree from
Harbin Institute of Technology, Harbin, China, in
2011, and the M.S. degree from Columbia University, New York, NY, USA, in 2013, both in electrical
He is currently working toward the Ph.D. degree in
electrical engineering at Tufts University, Medford,
MA, USA. His research interests include analog and
RF circuits for optical sensing and communication.
Valencia Joyner Koomson (M’99) received the B.S.
and M.Eng. degrees in electrical engineering and
computer science from Massachusetts Institute of
Technology, Cambridge, MA, USA, in 1998 and
1999, respectively, and the Ph.D. degree in electrical engineering from the University of Cambridge,
Cambridge, U.K., in 2003.
She is currently an Associate Professor with the
Department of Electrical and Computer Engineering,
Tufts University, Medford, MA, USA. Her research
interests include the design of optical and radio frequency integrated circuits for high-speed wireless communication and biomedical imaging.
Dr. Koomson was the recipient of the NSF Faculty Early Career Development
(CAREER) Award in 2010. She is a member of the IEEE Circuits and Systems
Society and Solid-State Circuits Society.
Authorized licensed use limited to: WASHINGTON UNIVERSITY LIBRARIES. Downloaded on October 09,2023 at 03:15:43 UTC from IEEE Xplore. Restrictions apply.