IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 4, NO. 3, JUNE 2010 181
HermesD: A High-Rate Long-Range Wireless
Transmission System for Simultaneous
Multichannel Neural Recording Applications
Henrique Miranda, Vikash Gilja, Cindy A. Chestek, Student Member, IEEE,
Krishna V. Shenoy, Senior Member, IEEE, and Teresa H. Meng, Fellow, IEEE
Abstract—HermesD is a high-rate, low-power wireless transmis-
sion system to aid research in neural prosthetic systems for motor
disabilities and basic motor neuroscience. It is the third generation
of our “Hermes systems” aimed at recording and transmitting
neural activity from brain-implanted electrode arrays. This system
supports the simultaneous transmission of 32 channels of broad-
band data sampled at 30 ks/s, 12 b/sample, using frequency-shift
keying modulation on a carrier frequency adjustable from 3.7 to
4.1 GHz, with a link range extending over 20 m. The channel rate
is 24 Mb/s and the bit stream includes synchronization and error
detection mechanisms. The power consumption, approximately
142 mW, is low enough to allow the system to operate continuously
for 33 h, using two 3.6-V/1200-mAh Li-
batteries. The
transmitter was designed using off-the-shelf components and is
assembled in a stack of three 28 mm
28-mm boards that fit in
a 38 mm
38 mm 51-mm aluminum enclosure, a significant
size reduction over the initial version of HermesD. A 7-dBi circu-
larly polarized patch antenna is used as the transmitter antenna,
while on the receiver side, a 13-dBi circular horn antenna is
employed. The advantages of using circularly polarized waves are
analyzed and confirmed by indoor measurements. The receiver
is a stand-alone device composed of several submodules and is
interfaced to a computer for data acquisition and processing. It is
based on the superheterodyne architecture and includes automatic
frequency control that keeps it optimally tuned to the transmitter
frequency. The HermesD communications performance is shown
through bit-error rate measurements and eye-diagram plots. The
Manuscript received September 07, 2009; revised November 19, 2009. Cur-
rent version published May 26, 2010. The work of H. Miranda was supported
in part by the Fundação para a Ciência e Tecnologia Fellowship, in part by the
Fulbright Ph.D. Scholarship, in part by the Focus Center for Circuit and System
Solutions (C2S2), in part by the Rethinking Analog Design (RAD) initiative,
and in part by the Stanford Center for Integrated Systems. The work of V. Gilja
was supported in part by the National Science Foundation (NSF) Graduate Re-
search Fellowship and in part by NDSEG Fellowship. The work of C. A. Chestek
was supported in part by the National Science Foundation (NSF) Graduate Re-
search Fellowships and in part by the William R. Hewlett Stanford Graduate
Fellowship. The work of K. V. Shenoy was supported in part by a McKnight En-
dowment Fund for Neuroscience Technological Innovations in Neurosciences
Award, in part by an NIH Director’s Pioneer Award 1DP1OD006409, and in
part by the Stanford Center for Integrated Systems. The work of T. H. Meng was
supported in part by the Focus Center for Circuit & System Solutions (C2S2),
in part by the Rethinking Analog Design (RAD) initiative, and in part by the
Stanford Center for Integrated Systems.
H. Miranda, C. A. Chestek, and T. H. Meng are with the Department of Elec-
trical Engineering, Stanford University, Stanford, CA 94305 USA (e-mail: hmi-
randa@stanford.edu;thm@stanford.edu ).
V. Gilja is with the Department of Computer Science, Stanford University,
Stanford, CA 94305 USA.
K. V. Shenoy is with the Departments of Electrical Engineering and Bioengi-
neering, and the Neurosciences Program, Stanford University, CA 94305 USA
(e-mail: shenoy@stanford.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.2010.2044573
sensitivity of the receiver is
83 dBm for a bit-error probability of
. Experimental recordings from a rhesus monkey conducting
multiple tasks show a signal quality comparable to commercial
acquisition systems, both in the low-frequency (local field poten-
tials) and upper-frequency bands (action potentials) of the neural
signals. This system can be easily scaled up in terms of the number
of channels and data rate to accommodate future generations of
Hermes systems.
Index Terms—High-rate frequency-shift keying (FSK) trans-
mitter, in-vivo neural recording, neural prosthetics, wireless
high-rate multichannel biotelemetry.
I. INTRODUCTION
N
EURAL recordings from freely-moving animals are an
emerging area of neuroscience research. The approach
presented in this paper enables the study of complex behav-
iors, such as social behavior, locomotion, or navigation [1], [2].
Wired systems are not feasible for studying freely-moving ani-
mals since cables significantly restrain movements and the an-
imals tend to remove any cable connection attached to them.
Another recording challenge is that the neural population under
analysis by implanted multielectrode arrays changes. Therefore,
overnight recording is necessary to track neurons over time for
learning studies, or for combining experimental trials across
days [3]. Wireless systems can be used in preclinical safety and
efficacy trials to test the effect of various medical treatments on
brain activity and behavior over long time periods. They also
find applications in the field of neural prosthesis, in which brain
activity is measured through various means and transformed
into command signals for an external device. Using a wireless
approach, algorithms can be tested in a less constrained setting.
However, regardless of the neural prosthetic application, the use
of many simultaneous neural channels is required, which can
pose a substantial challenge for wireless design.
For human clinical systems, neural implants with transcuta-
neous connections should ideally be avoided since they are a
potential source of infections, besides being aesthetically dis-
pleasing. Optimally, the wireless telemetry device would also be
implanted along with electrode array and power source [4], [5].
Several research groups are developing single-chip systems that
may eventually be small enough to enable a fully implantable
solution. However, single-chip solutions are not optimized for
neuroscience research. For example, several systems require the
external use of a power coil [6], [7] that cannot be worn by a
freely-moving animal. Other systems are limited to short-range
1932-4545/$26.00 © 2010 IEEE
Authorized licensed use limited to: Stanford University. Downloaded on May 29,2010 at 15:18:49 UTC from IEEE Xplore. Restrictions apply.
182 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 4, NO. 3, JUNE 2010
wireless links [8], [9]. Due to power and bandwidth limitations,
single-chip systems often compress data, transmit lower reso-
lution signals [6], or use threshold crossings to represent indi-
vidual action potentials [10], [11].
In contrast to clinical systems, large animals used for neu-
roscience research, such as monkeys, can carry a substantial
power supply that can be changed every 24 h as part of an ex-
perimental routine. Since the full bandwidth of neural signals is
of interest for neuroscience experiments, the acquisition of full
action potential waveforms as well as lower frequency neural
signals, such as electrocorticograms and local field potentials
(LFP), is very desirable in recording systems. These different
constraints underlie the need for a system optimized for neu-
roscience research using inexpensive commercial off-the-shelf
(COTS) hardware.
Several implementations of high-rate multichannel wireless
transmitters for neurological acquisition systems using off-the-
shelf components were reported earlier. Rizk and colleagues
propose a 96-channel, 1-Mb/s system capable of transmitting
one channel of broadband data over a range of 2 m using an
amplitude-shift keying (ASK) RF data transceiver; it requires
150 mW of power for the transmitter and an additional 300 mW
for each of the 32-channel analog interface circuits [6]. Obeid
and colleagues took a rather different approach that employed
an embedded computer and 802.11b wireless local-area network
(WLAN) equipment to implement a wearable telemetry system.
It is capable of transmitting 12 channels of broadband neural
data and requires 4 W of power, which is far too high for mul-
tiday recordings [12]. Other multichannel analog-based trans-
mitter systems [13]–[15] can be economical in terms of power
consumption since no analog-to-digital conversion stage is re-
quired. However, it is very difficult to ensure a controlled signal
fidelity under various channel impairments, such as multipath or
shadowing. Systems encoding data in the pulse duration or in the
pulse position can also be impaired by the delay spread of the
wireless channel. Moreover, multiplexing and demultiplexing
waveforms in the analog domain are not a straightforward oper-
ation, requiring additional circuitry to generate synchronization
markers inside the multiplexed signal.
Several other systems employ onboard spike detection and/or
spike sorting algorithms [6], [7], but the hardware-processing
resources are usually too scarce to provide high-quality spike
classification on a large number of channels simultaneously,
within a reasonably low-power budget. The approach followed
in the system reported here is to move the computational com-
plexity of the neural signal processing away from the transmitter
into the receiver processor. The availability of high-speed com-
puting resources in the receiver is plentiful, thus enabling the
experimentation of a wide range of processing algorithms.
To meet the required specifications at sufficiently low power
and small size for use with a rhesus macaque, we have devel-
oped HermesD, which provides simultaneous transmission of
32 channels of neural data sampled at 30 ks/s. This a miniatur-
ization of the system introduced in [16] to fit a smaller form
factor enclosure (detailed in [10]). While the prototype in [16]
occupied five boards of approximately 70 mm
70 mm in size,
the system presented in this paper is assembled in a stack of
three 28 mm
28-mm boards, whose details are presented in
TABLE I
SPECIFICATIONS OF
SEVERAL
HERMES GENERATIONS
the following sections. Neural data are collected and stored by
a custom-designed bench-top stand-alone receiver, improved
with tunability and automatic carrier frequency tracking char-
acteristics over the initial design presented in [16]. HermesD
is mainly targeted for neuroscience research applications that
involve multiday freely behaving experiments. It can measure
action potentials and local field potentials from all 32 electrodes
by virtue of transmitting broadband signals. It also provides
the unique capability to transmit neural data beyond 20 m,
allowing the study of macaques embedded in social colonies
in very large housing rooms. It can be further scaled up to
96 channels by expanding the analog amplifier array and the
number of analog-to-digital converters (ADCs). This is the
third generation of the Hermes systems that brings the capa-
bility of a wireless recording of a large number of wideband
neural channels. Table I compares the main characteristics of
HermesD with the previous generations.
In the following sections, we describe in detail the HermesD
wireless acquisition hardware, its communications perfor-
mance (antenna pattern, bit-error rate (BER), eye diagrams,
link budget), and we present and discuss in-vivo recordings
from a a rhesus macaque resting or engaged in a reaching task
while comfortably seated and head posted in an experimental
rig.
II. H
ERMESDSYSTEM
The block diagram of the HermesD system is shown in Fig. 1.
The transmitter is implemented using only COTS components
assembled in a stack of three printed-circuit boards (PCBs).
They are housed in a 38 mm
38 mm 51-mm aluminum en-
closure secured to the subject’s head. This enclosure is also used
for HermesCnano, the previous generation system [10]. It in-
cludes the batteries and an external microstrip patch antenna that
was specifically designed for the frequency range of HermesD,
as shown in Fig. 2. The receiver is built out of a combination
of off-the-self radio-frequency (RF) modules and custom-made
circuits. The recovered data packets are acquired by a digital
input/output (I/O) board interfaced to a PC for storage and pro-
cessing. The receiver is described in detail in Section II-B.
The neurological signals are captured by a 400-
m pitch
10
10 microelectrode array developed originally at the Uni-
versity of Utah [17] and now commercially available from
BlackRock Microsystems, Inc., Salt Lake City, UT. HermesD
was designed to process 32 of the 96 available channels,
Authorized licensed use limited to: Stanford University. Downloaded on May 29,2010 at 15:18:49 UTC from IEEE Xplore. Restrictions apply.
MIRANDA et al.: HERMESD: A HI GH-RATE LONG-RANGE WIRELESS TRANSMISSION SYSTEM 183
Fig. 1. HermesD system block diagram.
Fig. 2. Left: HermesD enclosure with a transmitter patch antenna; Right:
Transmitter board stack with batteries attached. Bottom: Physical design of
HermesD. 100-electrode array is implanted in macaque motor cortex. Preclude,
duragen, a silicone elastomer, and methyl methacrylate protect the brain, skull,
and array. AZIF connector attaches to the skull (CKI). A custom connector
provides 32 of 96 channels to the PCB stack. The aluminum housing embedded
in methyl methacrylate protects electronics and batteries.
with possible future expansion to support all of them. After
signal amplification and multiplexing, the neural data from
each channel are digitized at 30 ks/s with a resolution of 12
b/sample, followed by packetization and modulation. The
frequency-shift keying (FSK) modulation method is used due
to its low power consumption, the simplicity of its modulator
(a VCO is the only required device), and its relative robustness
against multipath channels. This modulation type also enables
the use of noncoherent techniques at the receiver.
The HermesD transmitter can be tuned to any frequency be-
tween 3.7 GHz and 4.1 GHz. Since the transmission bandwidth
is 60 MHz, it is possible to have six transmitters operating
simultaneously and to have their signals collected by indepen-
dent receivers. This is an interesting option if social interaction
between several animals within a housing room is to be studied.
There are a few reasons for the choice of this frequency band.
First, the allocated services for the band are commercial satellite
broadcast and fixed point-to-point microwave links, which pose
virtually no risk of interference to this system, and vice versa.
The UWB devices also operate in this frequency range, but since
the maximum allowed power density is
41.3[18] dBm/MHz,
corresponding to a power level 19 dB below the HermesD
transmitted power (in the 60-MHz bandwidth), these devices
represent a low interference risk. UWB interference probability
is further minimized if the UWB devices are not located in the
same room where the recordings take place due to the wall
attenuation. Second, man-made electromagnetic noise, such
as that originated by electric motors in some appliances, has a
low-power density in this microwave region and is also unlikely
to cause interference. Third, the frequency is high enough
to enable the design of small- and high-efficiency antennas
such as the one built for HermesD, which will be described
in Section II-A.IV. Finaly, a high carrier frequency enables
the use of high bandwidth signals to accommodate high bit
rates — important for scalability purposes. This system can
easily achieve more than 100 MHz of bandwidth for a larger
number of channels. The standard animal cages used during
the recording experiments have a small mesh aperture, on the
order of 2.5 cm
2.5 cm. If low carrier frequencies are used
(below 1 GHz), the cage attenuation can be very significant.
Transmission tests showed that frequencies around 4 GHz
suffer very little attenuation.
The RF output power of the transmitter, approximately
100
W, is enhanced by an antenna gain of 7 dB, providing
enough power to cover a range over 20 m with a comfortable
link margin of about 22 dB.
A. Transmitter
The detailed HermesD transmitter block diagram is shown in
Fig. 3.
1) Amplifier Array and ADC: The 32-channel signals are
amplified and filtered by two 16-channel biopotential array am-
plifiers from Intan Tech, LLC (RHA1016) [19]. These provide
46 dB of voltage gain and a configurable upper cutoff frequency,
which was set to 5 kHz in our experiments, while the lower
cutoff frequency was fixed at 0.05 Hz. Channels are multiplexed
before being digitized by the dual channel ADC, the Linear
Technology LTC1407A-1. Both multiplexed outputs from the
amplifier arrays are sampled simultaneously and converted into
14-b samples. Only 12 b/sample are transmitted since the useful
dynamic range is not improved by increasing the number of bits
beyond this resolution. This dynamic range is limited by the
noise level of the amplifier array and the maximum expected
neural spike amplitude. The measured input-referred noise level
Authorized licensed use limited to: Stanford University. Downloaded on May 29,2010 at 15:18:49 UTC from IEEE Xplore. Restrictions apply.
184 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 4, NO. 3, JUNE 2010
Fig. 3. HermesD transmitter block diagram.
Fig. 4. HermesD frame format.
of each channel is 3.2 Vrms. The value of each least-signif-
icant bit (LSB) was set to 1.5
V and spike amplitudes can
be as high as 6.3 mV before clipping occurs. This is a reason-
able limit since both action potentials and LFPs are expected
to be within this range. The sampling rate of the multiplexed
channels is 480 ks/s which corresponds to a data sample rate
of 480/16
30 ks/s per individual channel. This rate is com-
patible with the frequency content of neural signals, which is
negligible above 8 kHz.
2) Stream Packetizer and Frame Structure: The stream pack-
etizer collects samples from the ADC and organizes them in
structured frames that contain additional information for syn-
chronization and error detection. The packetizer is implemented
in a Xilinx XC2C64A CPLD (Complex Programable Logic De-
vice), and utilizes 43 of its 64 available macrocell blocks. The
maximum clock frequency at which the CPLD could run is
58 MHz, a limit considerably higher than the value needed by
HermesD (24 MHz). This limit is primarily imposed by the
check sum generator. Since this frequency also sets the system
bit rate, this same CPLD device can be used for future Hermes
versions with a higher channel count or higher channel sam-
pling rate. The frame format is depicted in Fig. 4. These frames
contain 50 bits and are continuously transmitted back-to-back.
There are four 7-bit fields that contain neural data information
(payload) as well as auxiliary fields that have relevant framing
tasks. Their purpose and justification are described as follows.
• FSS: Frame Sync Sequence—a 9-b synchronization pat-
tern of alternating zeros and ones (010101010) that marks
the start of a frame and aids the task of the clock recovery
circuit at the receiver (this pattern acts as a short clock
burst). The frame construction mechanism prevents this se-
quence from appearing anywhere else in the body of the
frame.
• Ch: Channel—a 4-b number that corresponds to the
channel being scanned by the amplifier arrays; this field
also acts as a frame-sequence number, useful to detect
missed frames during transmission errors.
• D1..D4: a pair of 12-b samples from the selected channel
of each amplifier array in two’s complement format.
• FCS: a 7-b checksum that is used to validate the frame
integrity; this checksum is obtained by simply summing
the four 7-b data fields (Ch and D1 to D4) and by dis-
carding any overflow bits (sum modulo 128) so that
.
Each frame also contains stuffing bits (
fields) that are added
at the end of every group of 7 b, starting after the FSS. The value
of each
bit is the complement of the value of the second bit
that precedes it. This not only ensures that the FSS pattern is not
created but also guarantees that the maximum sequence of equal
bits is limited to 8, so that the receiver bit recovery circuit can
operate flawlessly. The last stuffing bit has a different purpose:
it allows extra time for the FCS field to be computed in real
time, which is an important consideration if more complex error
checking methods or if higher speeds are used.
With this bit stuffing technique, in addition to a small frame
size, frame synchronization can be very fast, minimizing the
number of rejected frames during a resynchronization event.
This comes at the expense of 44% of frame overhead, with 12%
contributed by stuffing bits.
The sampling frequency of each neural channel is determined
by
ks/s, where
is the system clock and is the number of bits in the frame.
3) Modulator: The FSK modulator is built around a minia-
ture VCO module, the SMV3895A from Z-Communications,
Inc. The output frequency is set by a simple resistor divider and
the VCO is left in a free-running mode. No frequency stabilizing
mechanism is used, such as a phase-locked loop (PLL), due to
power saving reasons: current low-power PLL integrated cir-
cuits (ICs) consume more than 30 mW [20]. In fact, as the occu-
pied bandwidth of the modulated signal is fairly large, approx-
imately 60 MHz, the VCO frequency stability or phase noise
is not a major concern. The typical temperature drift of this
VCO is about 0.44 MHz
C. This drift has a low system impact
since the room where the experiments take place is temperature
controlled, causing an offset smaller than
2 MHz (the animal
temperature self-control reduces the effects of ambient tempera-
ture variation even more). To further enhance the frequency sta-
bility, a 10-dB attenuator was inserted at the output of the VCO.
This allows the output frequency to be fairly independent of any
impedance variation of the antenna due to its occasional prox-
imity to external objects (frequency pulling). The worst-case
frequency pulling was
MHz, measured with a moving short
connected to the transmitter output. On the receiver side, the
maximum frequency instability is typically less than
MHz,
caused by the first local oscillator (LO). The combined effect
of all frequency error sources is limited to
7 MHz, an offset
easily corrected by the receiver frequency tracking mechanism
described in Section II-B.III.
Authorized licensed use limited to: Stanford University. Downloaded on May 29,2010 at 15:18:49 UTC from IEEE Xplore. Restrictions apply.
MIRANDA et al.: HERMESD: A HI GH-RATE LONG-RANGE WIRELESS TRANSMISSION SYSTEM 185
Fig. 5. Measured return loss of the HermesD RHCP patch antenna mounted
on the electronics enclosure. Its 10-dB impedance bandwidth is 420 MHz and
covers the 3.7–4.1-GHz frequency range.
Using a wide signal bandwidth has clear advantages in giga-
hertz wireless telemetry systems that do not employ frequency
control or phase locking. A previously reported system using a
free-running VCO at 3.2 GHz and a signal bandwidth of only
10 kHz [21] becomes impractical without frequency stability
control mechanisms or fine frequency tracking at the receiver
side.
The chosen VCO has a low modulation input capacitance
(
), making it suitable for large signal band-
widths. In HermesD, the source impedance driving the VCO is
100 , which sets the maximum modulation bandwidth
to
MHz
thus enabling a maximum transmitter bit rate in excess
of 100 Mb/s. In HermesD, the pulse-shaping filter is a
third-order Bessel filter with a cutoff frequency of 30 MHz.
The frequency deviation for the FSK signal is set to
20 MHz, which corresponds to a modulation index of
.
4) Antennas: The transmitter antenna sits on top of the alu-
minum housing and is shown in Fig. 2. Its microstrip patch de-
sign allows a very low profile construction: the patch area is
24 mm
24 mm and the substrate used is 3.2 mm thick made
out of Teflon (RO5880 laminate from Rogers Corp.). The an-
tenna has an impedance bandwidth of 420 MHz for 10 dB of
return loss, as shown by the
plot of Fig. 5, and its oper-
ation frequency is matched to the transmitter frequency range.
Basic microstrip patch antennas have a relatively low fractional
impedance bandwidth, usually in the 1%–10% range. The use of
a thick and low relative permittivity substrate (2.2 in the present
case) in conjunction with the circular polarization (CP) charac-
teristic were key in obtaining a fractional bandwidth of 11%.
CP was obtained by truncating two opposite corners at 45
[22]
and its truncation extension was optimized to minimize the axial
ratio across the frequency range, using Agilent’s Momentum
EM simulator. The measured gain is 7.0 dB and its 88% of
efficiency was enabled by the very low loss properties of the
substrate (
0.0009). The antenna can cover a full hemi-
sphere with a maximum signal variation of 13 dB as shown in
Fig. 6, which is well within the signal margin for a 20-m link
distance (see Table II). This wide angular antenna coverage is
Fig. 6. Measured relative radiation plot of the HermesD transmitter antenna
assembled on the enclosure at 3.9 GHz. The 0
angle corresponds to the bore-
sight of the antenna.
TABLE II
H
ERMESDCOMMUNICATIONS LINK
BUDGET
a very desirable feature for reliable communications when the
animal subjects are freely moving. The radiation pattern was
measured in a standard room with the HP8720B vector network
analyzer, using a bandpass time-domain technique [23] to miti-
gate the effect of signal reflections on the pattern.
Both receiver and transmitter antennas employ right-hand
circular polarization (RHCP). This is particularly useful to
attenuate single-bounce reflections that occur in an indoor
environment, which are the main cause of signal fluctuation due
to destructive interference. The first and all odd-ordered bounce
reflections that arrive at the receiver have their polarization
rotation reversed since incidence angles are generally below
the pseudo-Brewster angle of typical wall material (60
to 70 )
[24]. These echoes are attenuated by the receiving antenna
which is configured for the original polarization rotation. In
order to quantitatively assess the advantage gain of the cir-
cular polarization, we performed channel measurements in a
8.8 m
5.8 m 2.8 m room at 4 GHz, using three different
transmitter antennas: an RHCP patch, a linear-polarized (LP)
patch, and a quarter-wavelength monopole for the same RHCP
Authorized licensed use limited to: Stanford University. Downloaded on May 29,2010 at 15:18:49 UTC from IEEE Xplore. Restrictions apply.
