Comparison of an Open-hardware Electroencephalography Amplifier
with Medical Grade Device in Brain-computer Interface Applications
Jérémy Frey
1,2
1
Univ. Bordeaux, Talence, France
2
Inria, Talence, France
Keywords:
BCI, EEG, Amplifiers Comparison, P300 Speller, Workload Classification.
Abstract:
Brain-computer interfaces (BCI) are promising communication devices between humans and machines. BCI
based on non-invasive neuroimaging techniques such as electroencephalography (EEG) have many applica-
tions, however the dissemination of the technology is limited, in part because of the price of the hardware. In
this paper we compare side by side two EEG amplifiers, the consumer grade OpenBCI and the medical grade
g.tec g.USBamp. For this purpose, we employed an original montage, based on the simultaneous recording of
the same set of electrodes. Two set of recordings were performed. During the first experiment a simple adapter
with a direct connection between the amplifiers and the electrodes was used. Then, in a second experiment,
we attempted to discard any possible interference that one amplifier could cause to the other by adding “ideal”
diodes to the adapter. Both spectral and temporal features were tested – the former with a workload monitoring
task, the latter with an visual P300 speller task. Overall, the results suggest that the OpenBCI board or a
similar solution based on the Texas Instrument ADS1299 chip – could be an effective alternative to traditional
EEG devices. Even though a medical grade equipment still outperforms the OpenBCI, the latter gives very
close EEG readings, resulting in practice in a classification accuracy that may be suitable for popularizing BCI
uses.
1 INTRODUCTION
Brain-computer interfaces (BCI) are communication
devices between humans and machines that rely only
on brain activity (i.e. no muscular input) to issue com-
mands or to monitor states (Wolpaw et al., 2002). BCI
is an emerging research area in Human-Computer
Interaction that offers new opportunities for inter-
action, beyond standard input devices (Tan and Ni-
jholt, 2010). In order to account for brain activity,
portable and non invasive neuroimaging techniques
are most commonly used, such as electroencephalog-
raphy – EEG, which measures electrical current onto
the scalp. The “interface” term covers many differ-
ent areas of applications, for people with or without
disabilities. However, while an increasing number of
systems are being developed, from BCI aimed at con-
trolling a cursor (Wolpaw et al., 2002) to adaptive sys-
tems (Zander and Kothe, 2011), more often than not
the use of the technology is limited.
The price of the hardware is one of the main rea-
sons that prevents the dissemination of non invasive
BCI. Recently, more affordable EEG amplifiers ap-
peared on the market, that could solve this issue.
Among them, the OpenBCI board
1
claims to bring
BCI to the many. Enthusiasts and laboratories have
started to use this board, but the quality of the record-
ings and the reliability of the resulting systems have
yet to be assessed. In this study, we compare side
by side the OpenBCI board with the g.tec g.USBamp
amplifier
2
, a device commonly used in BCI research.
The price tag of the g.tec solution is around 20 thou-
sands euros, 25 times more expensive than the 800
euros of 16 channels version of the OpenBCI board.
Both OpenBCI and g.USBamp amplifiers can record
up to 16 electrodes. This number of channels is suf-
ficient to setup various BCI. We compared OpenBCI
with g.USBamp for, on the one hand, a P300 speller
application and, on the other hand, a workload moni-
toring application. Doing so, we could study respec-
tively temporal and spectral features.
Note that here the question is not to assess which
amplifier is the best device per se. Instead, we inves-
tigate if in a context of popular interactions a nar-
row scope compared to the possibilities that offers the
1
http://www.openbci.com/
2
http://www.gtec.at/
Frey, J.
Comparison of an Open-hardware Electroencephalography Amplifier with Medical Grade Device in Brain-computer Interface Applications.
DOI: 10.5220/0005954501050114
In Proceedings of the 3rd International Conference on Physiological Computing Systems (PhyCS 2016), pages 105-114
ISBN: 978-989-758-197-7
Copyright
c
2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
105
g.USBamp it is conceivable for researchers from the
field or (well equipped) enthusiasts to make the leap.
To which extend should we employ devices coming
from the “DIY” (“Do It Yourself”) community for ac-
tual BCI applications?
To answers this question, we adopted an approach
somewhat different to what exists in the literature.
Many papers deal with the comparison of electrodes,
e.g. wet vs dry, with or without a conductive solution.
To do so, authors try to optimize the placement of
both sets of sensors in order to get measures that orig-
inate from the same spots. However, no matter their
efforts they could not merge sensors, and even clever
montages, with electrodes of one sort positioned be-
tween electrodes of the other sort (Tautan et al., 2013),
are not ideal. It will produce a slight offset, hence a
slight inaccuracy. Another alternative is to make sep-
arate measures by repeating the recordings with each
system (Nijboer et al., 2015), but once again the con-
ditions could not be exactly the same.
In the present study we do not attempt to assess
the quality of electrodes, but the behavior of ampli-
fiers that are attached to them. Not a whole system,
only the amplifiers. Therefore, we would not mind us-
ing the same electrodes during simultaneous record-
ings. This setup would ensure that the signal coming
in each amplifier’s inputs is exactly the same, avoid-
ing any bias regarding the source of the measures.
We made that possible by crafting a dedicated
adapter, one that basically splits in two the electrodes’
wires. Such parallel measurement works because the
amplifiers have high impedance circuits, that is to say
that they are designed to not draw any amount of cur-
rent from their source. As such, when one amplifier
is connected, the readings of the other stay the same.
Of course an infinite impedance cannot be achieved,
and no matter the precautions this setup may cause
a very slight difference compared to separate record-
ings. This is why in a second time we added to our
adapter a circuit that prevents any interference be-
tween the two amplifiers, using ideal diodes to block
current flows in one direction.
Using two different BCI applications, we inves-
tigated two types of EEG features. A task assess-
ing workload aimed at assessing spectral informa-
tion, and an oddball task sought temporal information.
For each amplifier we measured the performance of
a classifier based on those recordings, and addition-
ally we compared both by correlating the signals that
they recorded. No matter the financial aspects, the
qualities of the g.USBamp amplifier make it the per-
fect baseline to gauge new challengers. This is also
true for the electrodes developed by its manufacturer;
in this study we are using g.tec wet and active (pre-
Figure 1: Schematics of the direct adapter between elec-
trodes and amplifiers.
amplified) electrodes.
2 FIRST EXPERIMENT: DIRECT
CONNECTIONS
2.1 Experimental Setup
We acquired 16 EEG channels using the active
g.Ladybird electrodes from g.tec. In this system,
the electrodes are attached to a box that powers
their electrical components and retrieves the signal;
the g.GAMMAbox. After studying the wiring of
the g.GAMMAbox, we designed a printed circuit
board (PCB) to connect both amplifiers. Our adapter
plugs on one end to the D-sub 26 connector of the
g.GAMMAbox. Thanks to a pinout composed of
2.54mm connectors that gave access to all the chan-
nels (16 EEG + reference + ground), we attached the
OpenBCI board to the adapter ground set to “bias”
pin. On the other end of the adapter there was a D-
sub 26 female connector, onto which we could plug
the g.USBamp amplifier as if it were the regular end
of the g.GAMMAbox. The schematics of the 2 layers
PCB is presented in Figure 1.
The EEG channels were positioned according to
the 10-20 system at AFz, Fz, FCz, C3, C1, Cz, C2,
C4, CPz, P3, Pz, P4, POz, O1, Oz and O2 ground
at FPz, reference on the left earlobe. Since the mea-
sures between both amplifiers were identical, only
one recording session occurred, with one participant
there were no factors to counterbalance with re-
peated measures. The signals of both amplifier were
acquired using OpenViBE 1.0.1
3
; at 512Hz sampling
rate for the g.USBamp and 125Hz sampling rate for
the OpenBCI board.
3
http://openvibe.inria.fr/
PhyCS 2016 - 3rd International Conference on Physiological Computing Systems
106
The spectral features were investigated with a
workload monitoring application, i.e. a BCI that is
able to discriminate between several levels of mental
effort. The system was trained using the N-back task,
a well-known task to induce workload by playing on
memory load and time pressure (see, e.g. (Mühl et al.,
2014)). The protocol we used was similar to (Mühl
et al., 2014), there were 360 trials presented during 6
blocks of alternate difficulty levels (0-back vs 2-back
conditions). The recording session lasted approxi-
mately 12 minutes.
The temporal features were investigated using an
oddball task directly implemented within OpenViBE
with a visual P300 speller. P300 spellers are well-
established BCI applications during which letters that
randomly flash on the screen can be used to spell
words with the sole brain activity. Indeed, when the
letter that the user wants to spell flashes, a particular
event-related potential (ERP) arises within the EEG,
which possess a positive “peak” around t=300 ms af-
ter the stimulus onset. This is commonly referred to
as the “oddball paradigm” since the occurrence of rare
stimuli is used to elicit brain responses. During the
recordings a matrix of 6 by 6 letters and digits was
displayed in full screen on a 24-inch display. Only the
calibration session occurred, during which one need
to focus one’s attention on a predefined sequence of
letters. 32 letters composing a pangram were men-
tally “spelled” this way. The sentence was, without
spaces, “pack my box with five dozen liquor jugs”.
Letters were flashing for 0.2s. There were 24 flashes
per letter (12 times the row, 12 times the column),
hence due to the matrix disposition there were in total
4608 trials, among which 768 were targets “odd”
trials, i.e. the letters of the target sentence were flash-
ing. The recording session lasted approximately 30
minutes.
The acquisition of both amplifiers’ signals and
the P300 application occurred within the same Open-
ViBE scenario (script). The recordings of each am-
plifier were synchronized with the appropriate events
and exported in separate GDF files for later analyses.
There was also only one scenario involved in the syn-
chronization of all signals and events in the case of the
N-back task; stimulation from the python script sup-
porting this latter task were retrieved using the LSL
protocol, a network protocol dedicated to physiologi-
cal recordings which ensures accurate timings
4
.
2.2 Signal Processing
Two kinds of analyses were performed, using stan-
dard BCI signal processing pipelines. One aimed at
4
https://github.com/sccn/labstreaminglayer
assessing if and how the amplifiers differ in practice,
when used for classification. The second then looked
at the correlation between the acquired signals.
2.2.1 Classification
The signal processing of the data acquired during the
N-back task is analogous to (Mühl et al., 2014), i.e. 2s
time windows, 5 frequency bands delta (1-3 Hz),
theta (4-6 Hz), alpha (7-13 Hz), beta (14-25 Hz) and
gamma (26-40 Hz) and spatial filters. We used
common spatial patterns spatial filters to reduce the
16 channels to 6 “virtual” channels more discrimi-
nant between the workload conditions – see (Ramoser
et al., 2000). Additionally, we also tested a 3 fre-
quency bands version of our pipeline, that consider
only the lower frequencies, less prone to muscular ar-
tifacts – delta, theta and alpha.
Concerning the oddball task, we band-passed the
signal between 0.5Hz and 40Hz, downsampled it by a
factor 32 using the “decimate” Matlab function by
a factor 8 for OpenBCI because of the reduced sam-
pling rate –, and applied regularized Eigen Fisher spa-
tial filters a spatial filter specifically designed for
ERPs classification (Hoffmann et al., 2006) to re-
duce channels’ dimension from 16 to 5. We used 1s
time windows after stimuli onsets letters’ flashes
to epoch (“slice”) our signal. However, in order to
prevent data to overlap between consecutive stimuli
due to the rapid pace of the flashes, after a first pass
of epoching we discarded overlapping time windows
from further analyses. This ensured that no part of the
signal could be seen twice by the classifier between
the training phase and the testing phase and bias the
accuracy. The procedure was automatic, the first non-
overlapping epoch in order of appearance being kept.
As a result, in the end we obtained 48 target trials
and 240 distractor trials for classification, identical
between the g.USBamp and the OpenBCI recordings.
Both for the workload and the P300 speller tasks,
we used shrinkage LDA (linear discriminant analysis)
for classification (Ledoit and Wolf, 2004). To assess
the classifiers’ performance on the calibration data,
we used 4-fold cross-validation. I.e. we split the col-
lected data into 4 parts of equal size, used 3 parts to
calibrate the classifiers and tested the resulting clas-
sifiers on the unseen data from the remaining part.
This process occurred 3 more times so that in the end
each part was used once as test data. Finally, we aver-
aged the obtained classification accuracies. The accu-
racy was measured using the area under the receiver-
operating characteristic curve (AUROCC). The AU-
ROCC is a metric that is robust against unbalanced
classes, as it is the case with oddball tasks. A score
of “1” means a perfect classification, a score of “0.5”
Comparison of an Open-hardware Electroencephalography Amplifier with Medical Grade Device in Brain-computer Interface Applications
107
is chance. In order to make statistical comparisons
between both amplifiers for each type of features that
we studied, we ran 10 times the analyses the trials
were selected randomly for cross-validation.
2.2.2 Correlations
We compared, on the one hand, the frequency spectra
associated to the different workload conditions and,
on the other hand, the time course of the ERP that
were caused by the flashing target letters. To do so,
we used Pearson correlations, on par with the litera-
ture for similar analyses e.g. (Zander et al., 2011).
In order to ensure a 1-to-1 correspondence between
our sets of data, the recordings from the g.USBamp
were downsampled to 125Hz same sampling rate
as for the OpenBCI using the “resample” function
from Matlab R2014a signal processing toolbox.
Concerning the workload task, we first aggregated
the 2s time-windows corresponding to each condition
(0-back and 2-back). Then we used the “spectopo”
function of the EEGLAB toolbox (version 13.4.4b)
to compute the grand average power spectral between
1Hz and 40Hz, for each channel. The output of
the function was then passed on to R (version 3.0.2)
to compute correlations through the “rcorr” function
from the “Hmisc” package.
For the oddball task, we first band-passed the sig-
nals between 1Hz and 8Hz the approximate fre-
quency band used for classification. Then we ex-
tracted time epochs starting 0.5s prior to the flashing
of the target letters and ending 1s after stimuli on-
set. Contrary to what occurred for classification, we
did not prune overlapping epochs in the oddball task
when we compute the averaged ERP there was no
bias that could have been induced here. Finally, we
averaged the ERP per channel before exporting the
time points to the R environment.
2.3 Results
2.3.1 Classification
The results regarding classification accuracy are pre-
sented in Table 1, with the AUROCC scores for each
one of the 10 repetitions, for both amplifiers and both
tasks including the 3 and 5 frequency bands pipeline
for workload.
We tested for significance using Wilcoxon signed-
rank tests. There was a significant difference be-
tween amplifiers for the P300 tasks (p < 0.01). The
AUROCC mean score for the g.USBamp was 0.961
vs 0.918 for the OpenBCI. There were however no
significance but tendencies concerning the workload
task, with mean AUROCC scores between 0.85 and
0.86 for the 3 bands pipeline (p = 0.095) and between
0.89 and 0.90 for the 5 bands task (p = 0.079) see
Table 1 for details.
2.3.2 Correlations
When we first analyzed our data to seek correlations
regarding the oddball tasks, we realized that a shift
occurred during the recordings, as denoted in Figure 2
by the grand average of the ERP for target trials across
channels. This may have been caused by a software
issue (see Discussion). In order to correct the shift and
conduct proper comparisons between both amplifiers’
measures, we used a cross-correlation to estimate the
time shift, using the “ccf” function from the R “stats”
package. We found a delay of 88ms between the two
signals – 11 data points at 125Hz, see Figure 2.
In Figure 3a, the averaged ERP were shifted by as
much for each channel. Corresponding Pearson corre-
lation R scores, that were computed using the “rcorr”
function, are presented in Table 2. The mean R score
is 0.9965 and is statistically significant (p < 0.001).
There was also a significant correlation
(p < 0.001) for the spectral features, with a mean R
score of 0.9983 for the 0-back condition and 0.9979
the 2-back condition (see Table 2 for details). Among
the brain signals patterns that could be expected
during the completion of a difficult task, the decrease
nearby the alpha frequency band during the 2-task
condition can be observed within per-channel spectra
presented in Figures 3b and 3c. Note that we did not
correct time shifts prior to workload analyses due
to the nature of the features i.e. spectral and not
temporal.
2.4 Discussion
The correlation between both temporal and spectral
features tends so show that the signals acquired by
the g.USBamp and the OpenBCI are, if not identical,
very closely related. For every condition and channel
tested, the Pearson R score was greater than 0.99.
There were however more dissimilarities in the
classification accuracy obtained during the corre-
sponding tasks. While there were hardly a difference
between the AUROCC scores computed from both
amplifiers with the N-back tasks, the g.USBamp per-
formed significantly better than the OpenBCI during
the P300 speller task. The time shift observed after-
wards between the two amplifiers may partially ex-
plain this difference. Indeed, the detection of ERP is
particularly sensitive to signals’ latency, and a shift
between events’ timestamp and signal’s acquisition
could result in such degradation of performance when
temporal features are involved.
PhyCS 2016 - 3rd International Conference on Physiological Computing Systems
108
1
0
1
0.5 0.0 0.5 1.0
Time (s)
uV
Device gtec openbci
ERP for p300 target direct connection
30 20 10 0 10 20 30
0.0 0.2 0.4 0.6 0.8 1.0
Lag
CCF
ERP crosscorrelation direct connection
Figure 2: Left: Averaged ERP across channels of the target trials during the oddball task, before time shift correction. Right:
Cross-correlation between the amplifiers. The computed lag of 11 data points corresponds to 88ms. (Direct connection.)
AFz Fz FCz C3
C1 Cz C2 C4
CPz P3 Pz P4
POz O1 Oz O2
−2
0
2
4
−2
−1
0
1
2
−3
−2
−1
0
1
2
−2
−1
0
1
2
−3
−2
−1
0
1
2
−3
−2
−1
0
1
2
−2
−1
0
1
−2
−1
0
1
−2
−1
0
1
2
−3
−2
−1
0
1
−1
0
1
−2
0
2
−2
−1
0
1
2
−2
0
2
−2
−1
0
1
2
−2
−1
0
1
2
0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0
Time (s)
uV
Device gtec openbci
ERP for p300 − target − direct connection
(a)
AFz Fz FCz C3
C1 Cz C2 C4
CPz P3 Pz P4
POz O1 Oz O2
−10
0
10
20
30
−10
0
10
20
30
−10
0
10
20
30
−10
0
10
20
30
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Frequency (Hz)
Power 10*log
10
(uV
2
/Hz)
Device gtec openbci
Spectra for workload − 0−back − direct connection
(b)
AFz Fz FCz C3
C1 Cz C2 C4
CPz P3 Pz P4
POz O1 Oz O2
−10
0
10
20
30
−10
0
10
20
30
−10
0
10
20
30
−10
0
10
20
30
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Frequency (Hz)
Power 10*log
10
(uV
2
/Hz)
Device gtec openbci
Spectra for workload − 2−back − direct connection
(c)
Figure 3: Direct connection: averaged ERP for the target trials of the oddball task (a), averaged spectra for the 0-back (b) and
2-back (c) trials of the N-bak task.
Comparison of an Open-hardware Electroencephalography Amplifier with Medical Grade Device in Brain-computer Interface Applications
109
Table 1: Classification accuracy (AUROCC scores) for the P300 and workload tasks studied during the first experiment
direct connection between the electrodes and the amplifiers. The 4-fold cross validations were repeated 10 times. Two
pipelines are presented for the workload: 3 frequency bands (“WL 3b”, δ+ θ+ α) as well as 5 frequency bands pipeline (“WL
5b”, δ + θ + α+ β + γ). Significance was tested using Wilcoxon signed-rank tests.
Condition Amplifier 1 2 3 4 5 6 7 8 9 10 Mean SD
P300 g.USBamp 0.96 0.96 0.96 0.96 0.96 0.96 0.96 0.96 0.96 0.97 0.961 0.003
OpenBCI 0.92 0.92 0.91 0.91 0.92 0.92 0.93 0.92 0.92 0.91 0.918 0.006
WL 3b g.USBamp 0.85 0.85 0.85 0.86 0.85 0.87 0.87 0.87 0.87 0.86 0.860 0.009
OpenBCI 0.86 0.86 0.85 0.86 0.84 0.86 0.85 0.85 0.85 0.85 0.853 0.007
WL 5b g.USBamp 0.90 0.89 0.90 0.89 0.91 0.90 0.89 0.90 0.89 0.90 0.897 0.007
OpenBCI 0.91 0.89 0.87 0.88 0.89 0.88 0.89 0.89 0.89 0.90 0.889 0.011
Table 2: Pearson correlation R scores between g.USBamp and OpenBCI recordings at the 16 different electrode locations
with a direct connection. The “P300 target” condition corresponds to temporal features (ERP averaged across trials) and the
workload conditions to spectral features.
AFz Fz FCz C3 C1 Cz C2 C4 CPz
P300 target 0.998 0.997 0.997 0.997 0.997 0.994 0.996 0.992 0.994
Workload 0-back 0.999 0.999 0.998 0.998 0.998 0.998 0.998 0.999 0.998
Workload 2-back 0.999 0.999 0.998 0.998 0.998 0.998 0.998 0.999 0.997
P3 Pz P4 POz O1 Oz O2 Mean SD
P300 target 0.995 0.994 0.996 0.996 0.996 0.994 0.995 0.9965 0.0015
Workload 0-back 0.998 0.998 0.998 0.999 0.998 0.998 0.998 0.9983 0.0003
Workload 2-back 0.998 0.998 0.998 0.998 0.997 0.998 0.998 0.9979 0.0005
The radio transmission between the wireless
OpenBCI board and the dongle plugged to the com-
puter may be one of the cause of the situation. The
problem could also originate from the software. As
a matter of fact, the OpenViBE acquisition driver of
the OpenBCI board was released no so long before
our experiment, and was still labelled as “unstable”
as for version 1.0.1 of the software. One “oddity”
that may further highlight the youth of OpenBCI soft-
ware integration: we realized during our analysis that
the recorded signals were completely inverted on the
Y axis. The voltage reported by the board were the
opposite of what g.USBamp was claiming. Since on
numerous occasions we acknowledged the accuracy
of g.tec devices readings, it is the OpenBCI’s signals
that we inverted back to “normal” prior to correlation
analyses.
Beside time shifts issues, as mentioned during the
introduction we needed to strengthen those first in-
sights by discarding the eventuality that both EEG
signals may have influence each other due to the di-
rect wiring with the electrodes.
3 SECOND EXPERIMENT:
ISOLATED CONNECTIONS
The second set of recordings is very similar to the
what was described during the first study. The second
experiment only differs by the nature of the adapter
that was employed. As such we will only discuss the
changes that were made to the hardware and quickly
dive into the results.
3.1 Ideal Adapter
We modified the adapter that connects the amplifiers
to the g.GAMMAbox and by extent to the EEG
electrodes. Instead of a direct connection between
each amplifier’s inputs and the EEG channels, we in-
terposed “ideal” (or “super”) diodes on the branches
of the “Y” wiring.
Diodes are electrical components that let the cur-
rent flow in only one direction, the “forward” direc-
tion. Hence, this type of montage ensures that no
current could travel directly from one amplifier to the
other, contaminating the recordings. However, reg-
ular diodes cause a voltage drop. The voltage drop
varies depending on the materials used for their con-
struction, but it is at least 0.3V. Meaning that if the
current coming in the forward direction is lesser than
0.3V, no signal will pass through. 0.3V is an order of
magnitude superior to the range of EEG signal ap-
proximately a thousand time, therefore regular diode
could not be used.
To circumvent this problem, we utilized a particu-
lar montage that involved operational amplifiers (op-
PhyCS 2016 - 3rd International Conference on Physiological Computing Systems
110
Figure 4: Schematics of the adapter with the ideal diodes
montage. Note that there is a set of ideal diodes on the
ground channel, but they were shorted with a jumper during
our experiment.
amp). Op-amps are components widely used in elec-
trical circuits, acting as sorts of “building blocks”.
Notably, in combination with a regular diode, one
could use a precision rectifier configuration to ob-
tain an “ideal” diode. This particular montage is also
known as a “super” diode, since there will always be
a slight voltage drop, but in this case, thanks to the
gain of the op-amp, it becomes negligible.
We mounted 36 of such ideal diodes on the
adapter. One on each end of the “Y” section associ-
ated to the 16 EEG channels, plus 2 for the reference.
Due to the nature of the electrical recordings, only
the ground was left without such circuit. We utilized
Texas Instrument op-amps, model TLC2272ACPE4.
The TLC227xA series are more indicated for preci-
sion application, and with 2 op-amps per chip we
could limit the overall size of the adapter. The op-
erational amplifiers were powered by an external cir-
cuit with regulated -2.5 / +2.5 voltage. The schemat-
ics of the adapter also a 2 layers 2 layers PCB
is presented in Figure 4. The ideal diodes montage,
placed before amplifiers’ inputs, prevented any cur-
rent to flow in reverse direction from either amplifier
to the adapter; it ensured that one set of recordings
would not bias the other. One recording session oc-
curred for each application and each condition.
3.2 Results
The signal processing and the analyses were strictly
identical to the first experiment detailed above, refer
to the previous section for related information.
3.2.1 Classification
As with the first study, the results regarding classifica-
tion accuracy are presented in Table 3, with the AU-
ROCC scores for each one of the 10 repetitions, for
both amplifiers and both tasks including the 3 and
5 frequency bands pipeline for workload. We tested
for significance using Wilcoxon signed-rank tests. No
matter the task there was no significant difference, al-
though the 5% threshold was nearly reached for spec-
tral features p-value was 0.157 for the P300 task,
0.051 for the 3 bands version of the workload pipeline
and 0.286 for the 5 bands version.
3.2.2 Correlations
Concerning the P300 oddball task, there was a offset
of 88ms as well between the recordings of both ampli-
fier with the isolated connection – see Figure 5 for the
grand ERP average and the cross correlation. The per-
channel averaged ERP are plotted in figure 6a. Corre-
sponding Pearson correlation R scores are presented
in Table 4. The mean R score is 0.8847 and is statis-
tically significant (p < 0.001).
There was also a significant correlation
(p < 0.001) for the spectral features, with a
mean R score of 0.9976 for the 0-back condition
and 0.9987 the 2-back condition (see Table 4 for
details). The per-channel spectra are presented in
Figures 6b and 6c. As with the direct connection, the
band frequency changes between the 0-back and the
2-back conditions can be observed in the spectra.
3.3 Discussion
The results with the isolated connections are not that
different from what was obtained during the first ex-
periment. This would suggest that directly connecting
two high impedance amplifiers to the same EEG elec-
trodes could be a viable montage for a side-by-side
comparison.
Since with both types of connector there was only
one set of recordings, we could not draw any conclu-
sion about the lower classification accuracy obtained
with the isolated montage. The vigilance level of the
participant alone could explain these performances.
Thanks to signals’ correlations, however, we may
infer that noise was added to the system due to
the presence of additional electrical components in
the adapter. Indeed, while the spectra were once
again strongly correlated, the averaged ERP achieved
“only” a mean R score of 0.88. Here external factors
such as the metal state of the participant or the qual-
ity of electrodes contacts could not have influenced
one amplifier rather than the other. Since temporal
features are more sensitive than spectral features to
signal quality e.g. one “peak” in the signal vs os-
cillatory patterns over several seconds –, it is instead
Comparison of an Open-hardware Electroencephalography Amplifier with Medical Grade Device in Brain-computer Interface Applications
111
1
0
1
0.5 0.0 0.5 1.0
Time (s)
uV
Device gtec openbci
ERP for p300 target isolated connection
30 20 10 0 10 20 30
0.2 0.0 0.2 0.4 0.6 0.8
Lag
CCF
ERP crosscorrelation isolated connection
Figure 5: Left: Averaged ERP across channels of the target trials during the oddball task, before time shift correction. Right:
Cross-correlation between the amplifiers. The computed lag of 11 data points corresponds to 88ms. (Isolated connection.)
AFz Fz FCz C3
C1 Cz C2 C4
CPz P3 Pz P4
POz O1 Oz O2
−2.5
0.0
2.5
−2
0
2
−2
−1
0
1
2
−1
0
1
2
−2
−1
0
1
2
−2
−1
0
1
2
−1
0
1
2
−1
0
1
−1
0
1
2
−2
−1
0
1
−1
0
1
−3
−2
−1
0
1
2
−2
−1
0
1
2
−2
0
2
−5
0
5
−2
−1
0
1
2
0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0 0.0 0.5 1.0
Time (s)
uV
Device gtec openbci
ERP for p300 − target − isolated connection
(a)
AFz Fz FCz C3
C1 Cz C2 C4
CPz P3 Pz P4
POz O1 Oz O2
−10
0
10
20
−10
0
10
20
−10
0
10
20
−10
0
10
20
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Frequency (Hz)
Power 10*log
10
(uV
2
/Hz)
Device gtec openbci
Spectra for workload − 0−back − isolated connection
(b)
AFz Fz FCz C3
C1 Cz C2 C4
CPz P3 Pz P4
POz O1 Oz O2
−10
0
10
20
30
−10
0
10
20
30
−10
0
10
20
30
−10
0
10
20
30
0 10 20 30 40 0 10 20 30 40 0 10 20 30 40 0 10 20 30 40
Frequency (Hz)
Power 10*log
10
(uV
2
/Hz)
Device gtec openbci
Spectra for workload − 2−back − isolated connection
(c)
Figure 6: Isolated connection: averaged ERP for the target trials of the oddball task (a), averaged spectra for the 0-back (b)
and 2-back (c) trials of the N-bak task.
PhyCS 2016 - 3rd International Conference on Physiological Computing Systems
112
Table 3: Classification accuracy (AUROCC scores) for the P300 and workload tasks studied during the second experiment
isolated connection between the electrodes and the amplifiers. The 4-fold cross validations were repeated 10 times. Two
pipelines are presented for the workload: 3 frequency bands (“WL 3b”, δ+ θ + α) as well as 5 frequency bands pipeline (“WL
5b”, δ + θ + α + β + γ). Significance was tested using Wilcoxon signed-rank tests.
Condition Amplifier 1 2 3 4 5 6 7 8 9 10 Mean SD
P300 g.USBamp 0.84 0.82 0.83 0.83 0.81 0.85 0.83 0.84 0.83 0.84 0.832 0.011
OpenBCI 0.82 0.83 0.83 0.82 0.81 0.82 0.84 0.82 0.83 0.83 0.825 0.008
WL 3b g.USBamp 0.88 0.88 0.88 0.91 0.89 0.90 0.89 0.88 0.89 0.88 0.888 0.100
OpenBCI 0.88 0.88 0.89 0.89 0.88 0.89 0.87 0.88 0.88 0.87 0.881 0.007
WL 5b g.USBamp 0.92 0.91 0.92 0.90 0.91 0.89 0.90 0.91 0.91 0.90 0.907 0.009
OpenBCI 0.90 0.92 0.92 0.91 0.92 0.90 0.91 0.91 0.90 0.91 0.910 0.008
Table 4: Pearson correlation R scores between g.USBamp and OpenBCI recordings at the 16 different electrode locations
with an isolated connection. The “P300 target” condition corresponds to temporal features (ERP averaged across trials) and
the workload conditions to spectral features.
AFz Fz FCz C3 C1 Cz C2 C4 CPz
P300 target 0.976 0.934 0.892 0.846 0.872 0.881 0.838 0.811 0.912
Workload 0-back 0.999 0.998 0.998 0.997 0.998 0.998 0.998 0.996 0.998
Workload 2-back 0.999 0.999 0.999 0.998 0.998 0.999 0.999 0.999 0.999
P3 Pz P4 POz O1 Oz O2 Mean SD
P300 target 0.849 0.823 0.896 0.910 0.878 0.982 0.853 0.8847 0.0483
Workload 0-back 0.998 0.998 0.998 0.997 0.997 0.997 0.997 0.9976 0.0007
Workload 2-back 0.999 0.999 0.999 0.999 0.999 0.998 0.998 0.9987 0.0004
more plausible that the difference with the first exper-
iment comes from the adapter.
Nonetheless, even though the ideal diode montage
did not produce ideal signals, those results still ad-
vocate for a close proximity between the g.USBamp
and the OpenBCI. No device behaved “better” than
the other, because no matter the lower correlation be-
tween averaged ERP, the classification accuracy is in
practice comparable between both amplifiers. Each
one probably endured different fluctuations since each
had a dedicated set of ideal diodes.
4 CONCLUSION
During this preliminary study, we compared the
OpenBCI board to the g.tec g.USBamp amplifier. We
employed an original montage, based on the simulta-
neous recording of the same set of electrodes. While
as a first approach we used a simple adapter with a di-
rect connection between the amplifiers and the elec-
trodes, in a second experiment we attempted to dis-
card any possible interference that one amplifier could
cause to the other.
To do so, we built an adapter that embedded
“ideal” diodes, components that prevented electrical
currents to flow “backward”. This ensured that we
could test both devices in isolation. We did not try
to compare both adapters as the purpose was simply
to gather more insights about the possibility of simul-
taneous recordings this was a precaution to detect
a possible bias. For all applications and conditions
AUROCC scores were far beyond chancel level; the
OpenBCI amplifier came close to the g.USBamp in
terms EEG features and effective performance.
That is not to say that the OpenBCI could replace
an equipment such as the g.USBamp, though. For in-
stance, this open-hardware initiative does not aim at
medical applications, hence it should be employed in
sensitive contexts. It does not possess any certifica-
tion; one reason why so many cheap EEG device are
wireless is not for practicality, but to avoid any hazard
due to power supply. Connecting somehow a body to
the power grid requires extra precautions and a certi-
fied isolation, moreover when the impedance between
the electrodes and the brain is intentionally lowered.
There are also few issues with the current state of
the OpenBCI project. One concerns the sampling rate
of the board. While 125Hz may be enough for our
use-cases no frequencies beyond 40Hz were used
during the work presented here it may not suffice
others. The limitation of the sampling rate is caused
by the wireless protocol used for data transmission.
OpenBCI can deliver up to 250Hz signals to the com-
puter, but only on 8 channels instead of 16. Note that
this may be optimized in the future by updating the
Comparison of an Open-hardware Electroencephalography Amplifier with Medical Grade Device in Brain-computer Interface Applications
113
firmware or using alternate communications as far
as the board itself is concerned, the documentation of
the ADS1299 chip ensuring analog-to-digital conver-
sion claims a sampling rate up to 16,000Hz.
With those limitations in mind, overall the results
suggest that the OpenBCI board or a similar solution
based also on the Texas Instrument ADS1299 chip
could be an alternative to traditional EEG amplifiers.
Even though medical grade equipment possesses cer-
tification and still outperforms the OpenBCI board in
terms of classification, the latter gives very close EEG
readings. In practice, the obtained classification accu-
racy may be suitable for reliable BCI in popular set-
tings, widening the realm of applications and increas-
ing the number of potential users.
ACKNOWLEDGEMENTS
I thank Thibault Laine for his technical assistance dur-
ing this project.
REFERENCES
Hoffmann, U., Vesin, J., and Ebrahimi, T. (2006). Spatial
filters for the classification of event-related potentials.
In ESANN ’06.
Ledoit, O. and Wolf, M. (2004). A well-conditioned estima-
tor for large-dimensional covariance matrices. Journal
of Multivariate Analysis, 88(2):365–411.
Mühl, C., Jeunet, C., and Lotte, F. (2014). EEG-based
workload estimation across affective contexts. Fron-
tiers in Neuroscience, 8(8 JUN):1–15.
Nijboer, F., van de Laar, B., Gerritsen, S., Nijholt, A., and
Poel, M. (2015). Usability of Three Electroencephalo-
gram Headsets for Brain–Computer Interfaces: A
Within Subject Comparison. Interacting with Com-
puters, 27(5):500–511.
Ramoser, H., Müller-Gerking, J., and Pfurtscheller, G.
(2000). Optimal spatial filtering of single trial EEG
during imagined hand movement. IEEE Trans Reha-
bil Eng, 8(4):441–446.
Tan, D. S. and Nijholt, A. (2010). Brain-Computer Inter-
faces: Applying Our Minds to Human-Computer In-
teraction. Springer.
Tautan, A.-M., Serdijn, W., Mihajlovic, V., Grundlehner,
B., and Penders, J. (2013). Framework for evaluat-
ing EEG signal quality of dry electrode recordings.
2013 IEEE Biomedical Circuits and Systems Confer-
ence (BioCAS), pages 186–189.
Wolpaw, J. R., Birbaumer, N., McFarland, D. J.,
Pfurtscheller, G., and Vaughan, T. M. (2002). Brain-
computer interfaces for communication and control.
Clin. Neurophysiol., 113(6):767–91.
Zander, T. O. and Kothe, C. (2011). Towards passive brain-
computer interfaces: applying brain-computer inter-
face technology to human-machine systems in gen-
eral. J. Neural. Eng, 8(2):025005.
Zander, T. O., Lehne, M., Ihme, K., Jatzev, S., Cor-
reia, J., Kothe, C., Picht, B., and Nijboer, F.
(2011). A Dry EEG-System for Scientific Research
and Brain–Computer Interfaces. Frontiers in Neuro-
science, 5(May):1–10.
PhyCS 2016 - 3rd International Conference on Physiological Computing Systems
114