Multi-Person Brain Activity Recognition via Comprehensive EEG Signal Analysis ACM Reference format

Abstract

An electroencephalography (EEG) based brain activity recognition is a fundamental eld of study for a number of signiicant applications such as intention prediction, appliance control, and neurological disease diagnosis in smart home and smart healthcare domains. Existing techniques mostly focus on binary brain activity recognition for a single person, which limits their deployment in wider and complex practical scenarios. erefore, multi-person and multi-class brain activity recognition has obtained popularity recently. Another challenge faced by brain activity recognition is the low recognition accuracy due to the massive noises and the low signal-to-noise ratio in EEG signals. Moreover, the feature engineering in EEG processing is time-consuming and highly relies on the expert experience. In this paper, we aaempt to solve the above challenges by proposing an approach which has beeer EEG interpretation ability via raw Electroencephalography (EEG) signal analysis for multi-person and multi-class brain activity recognition. Speciically, we analyze inter-class and inter-person EEG signal characteristics, based on which to capture the discrepancy of inter-class EEG data. en, we adopt an Autoencoder layer to automatically reene the raw EEG signals by eliminating various ar-tifacts. We evaluate our approach on both a public and a local EEG datasets and conduct extensive experiments to explore the eeect of several factors (such as normalization methods, training data size, and Autoencoder hidden neuron size) on the recognition results. e experimental results show that our approach achieves a high accuracy comparing to competitive state-of-the-art methods, indicating its potential in promoting future research on multi-person EEG recognition.

Full text

Multi-Person Brain Activity Recognition via Comprehensive
EEG Signal Analysis
Xiang Zhang
University of New South Wales
Sydney, Australia
xiang.zhang3@student.unsw.edu.au
Lina Yao
University of New South Wales
Sydney, Australia
lina.yao@unsw.edu.au
Dalin Zhang
University of New South Wales
Sydney, Australia
dalin.zhang@student.unsw.edu.au
Xianzhi Wang
Singapore Management University
Singapore
xzwang@smu.edu.sg
an Z. Sheng
Macquarie University
Sydney, Australia
michael.sheng@mq.edu.au
Tao Gu
RMIT University
Melbourne, Australia
tao.gu@rmit.edu.au
ABSTRACT
An electroencephalography (EEG) based brain activity recogni-
tion is a fundamental eld of study for a number of signicant
applications such as intention prediction, appliance control, and
neurological disease diagnosis in smart home and smart healthcare
domains. Existing techniques mostly focus on binary brain activity
recognition for a single person, which limits their deployment in
wider and complex practical scenarios. erefore, multi-person
and multi-class brain activity recognition has obtained popularity
recently. Another challenge faced by brain activity recognition is
the low recognition accuracy due to the massive noises and the
low signal-to-noise ratio in EEG signals. Moreover, the feature
engineering in EEG processing is time-consuming and highly re-
lies on the expert experience. In this paper, we aempt to solve
the above challenges by proposing an approach which has beer
EEG interpretation ability via raw Electroencephalography (EEG)
signal analysis for multi-person and multi-class brain activity recog-
nition. Specically, we analyze inter-class and inter-person EEG
signal characteristics, based on which to capture the discrepancy
of inter-class EEG data. en, we adopt an Autoencoder layer to
automatically rene the raw EEG signals by eliminating various ar-
tifacts. We evaluate our approach on both a public and a local EEG
datasets and conduct extensive experiments to explore the eect of
several factors (such as normalization methods, training data size,
and Autoencoder hidden neuron size) on the recognition results.
e experimental results show that our approach achieves a high
accuracy comparing to competitive state-of-the-art methods, indi-
cating its potential in promoting future research on multi-person
EEG recognition.
KEYWORDS
Brain computer interface, EEG classication, Activity recognition,
Auto-encoder
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for prot or commercial advantage and that copies bear this notice and the full citation
on the rst page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
Mobiquitous 2017, MELBOURNE, AU
©2017 Copyright held by the owner/author(s). 123-4567-2567/08/06. . .$15.00
DOI: 10.475/123 4
ACM Reference format:
Xiang Zhang, Lina Yao, Dalin Zhang, Xianzhi Wang, an Z. Sheng, and Tao
Gu. 2017. Multi-Person Brain Activity Recognition via Comprehensive EEG
Signal Analysis. In Proceedings of Mobiquitous 2017, MELBOURNE, AU, Nov.
2017, 10 pages.
DOI: 10.475/123 4
1 INTRODUCTION
Brain activity recognition is one of the most promising research
areas over the last few years. It has the potential to revolutionize a
wide range of applications such as ICU (Intensive Care Unit) mon-
itoring [
18
], appliance control [
13
,
28
], assisted living of disabled
people and elderly people [
5
], and diagnosis of neurological dis-
eases [
9
]. For instance, people with Amyotrophic Lateral Sclerosis
(ALS) generally have only limited physical capacities and they are
unable to communicate with the outer world, such as performing
most daily activities, e.g., turning on/o a light. In such occasions,
brain activity recognition can help interpret their demands and
assist them to live more independently with dignity through the
mind-control intelligence. Brain activities are mostly represented
by Electroencephalography (EEG) signals, which record the voltage
uctuations of brain neurons with the electrodes placed on the
scalp in a non-invasive way.
Although brain activity recognition has been widely investigated
over the past decade, it still faces several challenges such as multi-
person and multi-class classication. First, despite several studies
on multi-person EEG classication, e.g., [
6
] employed a LDA (linear
discriminant analysis) classier to classify two datasets with nine
and three subjects, there still has space for improvement over the
existing methods in terms of the classication accuracy (86.06%
and 93% over the two datasets in [
6
]). Second, to the best of our
knowledge, most existing applications that adopt EEG classication
are for diseases diagnosis (such as epilepsy and Alzheimer’s dis-
eases), which requires only binary classication (normal or abnor-
mal). However, there exist various other deployment occasions (e.g.,
smart home and assisted living) that demand multi-class EEG classi-
cation. For instance, EEG-based assisting robots require more than
two commands (such as walking straight, turning le/right, and
raising/lowering hands) to complete assisted living tasks. Regard-
ing this, only some preliminary research exists, such as [
26
], which
adopted SVM to classify a four-class EEG dataset and achieved the
accuracy of 70%.

Mobiquitous 2017, Nov. 2017, MELBOURNE, AU Xiang Zhang, Lina Yao, Dalin Zhang, Xianzhi Wang, an Z. Sheng, and Tao Gu
In this paper, we propose a novel brain activity recognition ap-
proach to classifying the multi-person and multi-class EEG data.
We analyze the similarity of EEG signals and calculates the cor-
relation coecients matrix in both inter-class and inter-person
conditions. en, on top of data similarity analysis, we extract
EEG signal features by the Autoencoder algorithm, and nally feed
the features into the XGBoost classier to recognize categories of
EEG data, with each category corresponding to one specic brain
activity. e main contributions of this paper are summarized as
follows:
•
We present a novel brain activity recognition approach
based on comprehensive EEG analysis. e proposed ap-
proach directly works on the raw EEG data, which en-
hances the ductility, relieves from EEG signal pre/post-
processing, and decreases the need of human expertise.
•
We calculate the correlation coecients matrix and mea-
sure the self-similarity and cross-similarity under both
inter-class and inter-person conditions. Based on the simi-
larity investigation, we propose three favorable conditions
of multi-person and multi-class EEG classication.
•
We adopt the Autoencoder, an unsupervised neuron net-
work algorithm, to rene EEG features. Moreover, we
investigate the size of hidden-layer neurons to optimize
the neurons size to optimize the classication accuracy.
•
We conduct an experiment to evaluate the proposed ap-
proach on a public EEG dataset (containing 560,000 samples
from 20 subjects and 5 classes) and obtain the accuracy
of 79.4%. Our approach achieves around 10% accuracy im-
provement compared with other popular EEG classication
methods.
•
We design a case study to evaluate the proposed approach
on a local dataset which consists of 172,800 samples col-
lected from 5 subjects and 6 classes. Our approach obtains
the accuracy of 74.85% and outperforms the result of the
state-of-the-art methods.
e rest of this paper is organized as follows. Some existing
studies related to this paper are introduced in Section 2. Section 3
investigates the EEG data characteristic and provides the EEG sam-
ple similarity intra-class and inter-class. Section 4 describes the
methodology details of the approach adopted in this paper. e
experimental results and evaluation are presented in Section 5. A
local experiment case study is introduced in Section 6. Finally, we
summarized the paper and highlight the future work in Section 7.
2 RELATED WORK
Over the last decade, much aention has been drawn to brain data
modeling, a crucial pathway to translating human brain activity
into computer commands to realize Brain-Computer Interaction
(BCI). BCI systems are an alternative way to allow paralyzed or
severely muscular disordered patients to recover communication
and control abilities, as well as to save scarce medical care resources.
Recent research has also found its application in virtual reality [
25
]
and space applications [
20
]. As EEG signals are the most commonly
used brain data for BCI system [
2
,
30
,
31
], signicant eorts have
been devoted to build accurate and eective models for EEG-based
brain activity analysis [3, 15, 21, 27].
EEG Feature Representation Method.
Feature representa-
tion of EEG raw data has great impact on classication accuracy due
to the complexity and high dimensionality of EEG signals. Vzard
et al. [
24
] employed common spatial paern (CSP) along with LDA
to pre-process EEG data and obtained an accuracy of 71.59% on bi-
nary alertness states. Meisheri et al. [
14
] exploited multi-class CSP
(mCSP) combined with Self-Regulated Interval Type-2 Neuro-Fuzzy
Inference System (SRIT2NFIS) classier for four EEG-based motor
imagery classes (movement imagination of le hand, right hand,
both feet, and tongue) classication and achieved the accuracy of
54.63%, which is signicantly lower than the accuracy of binary
classication. Shiratori et al. [
23
] achieved a similar accuracy of
56.7% using mCSP coupled to the random forests for a three-class
EEG-based motor imagery task. e autoregressive (AR) modeling
approach, a widely used algorithm for EEG feature extraction, is
also broadly combined with other feature extraction techniques to
gain a beer performance [
19
]. For example, [
29
] investigated two
methods EEG with AR and feature extraction combination: 1) AR
model and approximate entropy, 2) AR model and wavelet packet
decomposition. ey employed SVM as the classier and showed
that AR can eectively improve classication performance. Duan
et al. [
7
] introduced the Autoencoder method for feature extraction
and obtained an accuracy of 86.69%.
EEG Multi-person Classication.
Multi-person EEG clas-
sication investigates mental signals from multiple participants,
each of whom undergoing the same brain activities. It is the re-
quirement of future ubiquitous application of EEG instruments to
capture the underlying consistency and inter-subject variations
among EEG paerns of dierent subjects. Kang et al. [
12
] pre-
sented a Bayesian CSP model with Indian Buet process (IBP) to
investigate the shared latent subspace across subjects for EEG clas-
sication. eir experiments on two EEG datasets containing ve
and nine subjects showed the superior performance of approxi-
mate 70% accuracy. Djemal et al. [
6
] utilized two multi-person
multi-class EEG datasets to validate sequential forward oating
selection (SFFS) and a multi-class LDA algorithm. Eugster et al.
[
8
] involved forty participants in their experiments to perform rel-
evance judgment tasks. ey also recorded the EEG signals for
further classication research. Ji et al. [
11
] investigated a dataset
containing nine subjects for analyzing and evaluating a hybrid
brain-computer interface.
EEG Multi-class Classication.
Multi-class classication
is a major challenge in EEG signal analysis, given that current EEG
classication research is mostly focused on binary classication.
Usually, an algorithm achieves only inferior performance when
handling multi-classication than in handling binary classication.
Anh et al. [
1
] used Articial Neural Network trained with output
weight optimization back-propagation (OWO-BP) training scheme
for dual- and triple-mental state classication problems. ey got
a classication accuracy of 95.36% on dual mental state for triple
classication problems, the algorithm performance fell o to 76.84%.
With the four-class problem, Olivier et al. [
17
] got an accuracy of
around 50% when using a voting ensemble neural network classier.
Aiming at four-class EEG classication, Wang et al. [
26
] employed
four preprocessing steps and a simple SVM classier and got an
average classication accuracy of 70%.

Multi-Person Brain Activity Recognition Mobiquitous 2017, Nov. 2017, MELBOURNE, AU
In summary, diering from previous work, this paper proposes
an Autoencoder+XGBoost algorithm to address the multi-class multi-
person EEG signal classication problem, which is a core challenge
in applying brain activity recognition technologies to many impor-
tant domains. e proposed algorithm engages the Autoencoder for
EEG feature representation to explore the relevant EEG features.
Also, it emphasizes on the generalization over participants by solv-
ing an EEG classication problem with as much as ve classes
and taking twenty subjects. e present approach is supposed to
improve the accuracy and practical feasibility of EEG classication.
3 EEG CHARACTERISTIC ANALYSIS
To gain knowledge about EEG data characteristics and prepare for
the further EEG classication, we quantify the similarity between
EEG samples by calculating their Pearson correlation coecients,
using the following equation:
ρ(A,B)=1
¯
N−1
¯
N
Õ
¯
i=1
(A¯
i−¯
µA
¯
σA
)( B¯
i−¯
µB
¯
σB
),¯
i=1,2, . . . , ¯
N
where
A
and
B
denote two EEG vector samples, each containing
¯
N
elements.
µA
and
σA
denote the mean and standard deviation of
A
.
µB
and
σB
denote the mean and standard deviation of
B
. e Person
correlation coecient is positively correlated with the similarity,
and both are in the range of [0,1].
We introduce two similarity concepts used in our measurement:
self-similarity and cross-similarity. e self-similarity is dened by
the similarity of EEG signals within the same EEG category while
the cross-similarity is dened by the similarity of EEG signals of
two dierent EEG categories. Both the self-similarity and cross-
similarity are measured under two conditions: inter-class and inter-
person, respectively.
Inter-class measurement.
Under the inter-class situation,
we measure the correlation coecient matrix for every specic
subject and calculate the average matrix by calculating the mean
value of all the matrix. For example, there are 5 classes for the
specic subject, we calculate a 5
∗
5 correlation coecient matrix.
In this matrix,
ρ˘
i,˘
j
denotes the correlation coecient between the
samples of the class
˘
i
and the samples of the class
˘
j
. e self-
similarity indicates the similarity between two dierent samples
from the same class. e cross-similarity indicates the average of
similarity of each possible class pair of samples belonging to the
specic subject.
Inter-person measurement.
Under the inter-person situa-
tion, we measure the correlation coecients matrix for every specic
class and then calculate the average matrix. e self-similarity in-
dicates the similarity between two dierent samples from the same
class of the same subject. e cross-similarity denotes the average
of similarity of each possible subject pair of samples belonging to
the specic class.
Table 1 shows the inter-class correlation coecient matrix and
the corresponding statistical self- and cross-similarity. e last
column (PD) denotes the Percentage Dierence between the self-
similarity and cross-similarity. We can observe from the results that
the self-similarity is always higher than the cross-similarity for all
classes, meaning that the samples’ intra-class cohesion is stronger
than the inter-class cohesion. e percentage dierence has a
noticeable uctuation, indicating the varying intra-class cohesion
over dierent class pairs. Class 1 is easier to be distinguished due
to its highest percentage dierence, while in contrast, class 0 and
class 4 are dicult to be accurately classied.
Similarly, Table 2 shows the inter-person correlation coecient
matrix and gives an alternative visualization of the results. Again,
we nd that, for each class, the self-similarity is higher than cross-
similarity with varying percentage dierence. e standard devia-
tions of cross-similarity in the ve classes are similar. is indicates
the steady and even distribution of the dataset between dierent
subjects and dierent classes.
e above analysis results basically satisfy our following hypoth-
esis for multi-person multi-class classication: 1) the self-similarity
is consistently higher than cross-similarity both under inter-class
and inter-person conditions; 2) the higher inter-class percentage
dierence, the beer classication results; 3) lower average percent-
age dierences and standard deviations of the subjects result in the
beer classication performance under the inter-person condition.
4 METHODOLOGY
In this section, we review the algorithm by rst normalizing the
input EEG data and then automatically explore the feature repre-
sentation of the normalized data. At last, we adopt the XGBoost
classier to classify the trained features. e methodology ow-
chart is shown in Figure 1.
4.1 Normalization
Normalization plays a crucial role in a knowledge discovery process
for handling dierent units and scales of features. For instance,
given one input feature ranges from 0 to 1 while another ranges
from 0 to 100, the analysis results will be dominated by the laer
feature. Generally, there are three widely used normalization meth-
ods: Min-Max Normalization, Unity Normalization, and Z-score
Scaling (also called standardization).
Min-Max Normalization.
Min-Max Normalization projects
all the elements in an vector to the range of
[
0
,
1
]
. is method
maps features to the same range despite of their original means
and standard deviations. e formula of Min-Max normalization is
given below:
xnew =x−xmi n
xmax −xm in
where
xmi n
and
xmax
separately denotes the minimum and maxi-
mum in the feature x.
Unity Normalization.
Unity Normalization re-scales the fea-
tures by the percentage or the weight of each single element. It
calculates the sum of all the elements and then divides each element
by the sum. e equation is:
xnew =x
Íx
where
Íx
denotes the sum of feature
x
. Similar to Min-Max Nor-
malization, the results of this method also belong to the range of
[0,1].
Z-score Scaling.
Z-score Scaling forces features under nor-
mal Gaussian distribution (zero mean and unit variance), using the

Mobiquitous 2017, Nov. 2017, MELBOURNE, AU Xiang Zhang, Lina Yao, Dalin Zhang, Xianzhi Wang, an Z. Sheng, and Tao Gu
Table 1: Inter-class correlation coecients matrix. e correlation coecients matrix (upper le section) is the average of 20
correlation coecients matrix separately from 20 subjects.
Class 0 1 2 3 4 Self-similarity Cross-similarity Percentage dierence
0 0.4010 0.2855 0.4146 0.4787 0.3700 0.401 0.3872 3.44%
1 0.2855 0.5100 0.0689 0.0162 0.0546 0.51 0.1063 79.16%
2 0.4146 0.0689 0.4126 0.2632 0.3950 0.4126 0.2854 30.83%
3 0.4787 0.0162 0.2632 0.3062 0.2247 0.3062 0.2457 19.76%
4 0.3700 0.0546 0.3950 0.2247 0.3395 0.3395 0.3156 7.04%
Range 0.1932 0.4938 0.3458 0.4625 0.3404 0.2038 0.2809 75.72%
Average 0.3900 0.1870 0.3109 0.2578 0.2768 0.3939 0.2680 28.05%
STD 0.0631 0.1869 0.1334 0.1487 0.1255 0.0700 0.0932 27.33%
Table 2: Inter-person correlation coecients matrix. STD denotes Standard Deviation, SS denotes Self-similarity, CS denotes
Cross-similarity, and PD denotes Percentage Dierence.
Class 0 Class 1 Class 2 Class 3 Class 4
subjects SS CS PD SS CS PD SS CS PD SS CS PD SS CS PD
subject1 0.451 0.3934 12.77% 0.2936 0.1998 31.95% 0.3962 0.3449 12.95% 0.4023 0.1911 52.50% 0.5986 0.4375 26.91%
subject2 0.3596 0.2064 42.60% 0.3591 0.1876 47.76% 0.5936 0.3927 33.84% 0.2354 0.2324 1.27% 0.3265 0.2225 31.85%
subject3 0.51 0.3464 32.08% 0.3695 0.2949 20.19% 0.3979 0.3418 14.10% 0.4226 0.3702 12.40% 0.4931 0.4635 6.00%
subject4 0.3196 0.1781 44.27% 0.4022 0.1604 60.12% 0.3362 0.2682 20.23% 0.4639 0.3905 15.82% 0.3695 0.2401 35.02%
subject5 0.4127 0.2588 37.29% 0.3961 0.2904 26.69% 0.3128 0.2393 23.50% 0.4256 0.1889 55.62% 0.3958 0.3797 4.07%
subject6 0.33 0.2924 11.39% 0.3869 0.3196 17.39% 0.3369 0.3281 2.61% 0.4523 0.1905 57.88% 0.4526 0.3321 26.62%
subject7 0.4142 0.3613 12.77% 0.3559 0.342 3.91% 0.3959 0.3867 2.32% 0.4032 0.3874 3.92% 0.4862 0.2723 43.99%
subject8 0.362 0.1784 50.72% 0.4281 0.2121 50.46% 0.4126 0.2368 42.61% 0.3523 0.1658 52.94% 0.4953 0.2438 50.78%
subject9 0.324 0.2568 20.74% 0.3462 0.2987 13.72% 0.3399 0.3079 9.41% 0.3516 0.1984 43.57% 0.3986 0.177 55.59%
subject10 0.335 0.1889 43.61% 0.3654 0.2089 42.83% 0.2654 0.2158 18.69% 0.3326 0.2102 36.80% 0.3395 0.2921 13.96%
subject11 0.403 0.1969 51.14% 0.3326 0.2066 37.88% 0.3561 0.3173 10.90% 0.4133 0.1697 58.94% 0.5054 0.44 12.94%
subject12 0.4596 0.2893 37.05% 0.4966 0.3702 25.45% 0.3326 0.2506 24.65% 0.4836 0.3545 26.70% 0.3968 0.3142 20.82%
subject13 0.3956 0.2581 34.76% 0.4061 0.3795 6.55% 0.3965 0.3588 9.51% 0.3326 0.1776 46.60% 0.3598 0.3035 15.65%
subject14 0.3001 0.299 0.37% 0.3164 0.2374 24.97% 0.4269 0.3763 11.85% 0.3856 0.1731 55.11% 0.4629 0.3281 29.12%
subject15 0.3629 0.3423 5.68% 0.3901 0.2278 41.60% 0.7203 0.2428 66.29% 0.3623 0.3274 9.63% 0.3862 0.3303 14.47%
subject16 0.3042 0.1403 53.88% 0.3901 0.3595 7.84% 0.4236 0.331 21.86% 0.4203 0.1634 61.12% 0.4206 0.3137 25.42%
subject17 0.396 0.1761 55.53% 0.3001 0.2232 25.62% 0.6235 0.3579 42.60% 0.5109 0.198 61.24% 0.3339 0.2608 21.89%
subject18 0.4253 0.3194 24.90% 0.3645 0.2286 37.28% 0.6825 0.222 67.47% 0.4236 0.3886 8.26% 0.4936 0.3017 38.88%
subject19 0.5431 0.3059 43.68% 0.3526 0.2547 27.77% 0.4326 0.3394 21.54% 0.5632 0.3729 33.79% 0.4625 0.219 52.65%
subject20 0.3964 0.3459 12.74% 0.3265 0.2849 12.74% 0.4025 0.3938 2.16% 0.3265 0.1873 42.63% 0.3976 0.2338 41.20%
Min 0.3001 0.1403 0.37% 0.2936 0.1604 3.91% 0.2654 0.2158 2.16% 0.2354 0.1634 1.27% 0.3265 0.177 4.07%
Max 0.5431 0.3934 55.53% 0.4966 0.3795 60.12% 0.7203 0.3938 67.47% 0.5632 0.3905 61.24% 0.5986 0.4635 55.59%
Range 0.2430 0.2531 55.16% 0.2030 0.2191 56.21% 0.4549 0.1780 65.31% 0.3278 0.2271 59.97% 0.2721 0.2865 51.53%
Average 0.3902 0.2667 31.40% 0.3689 0.2643 28.14% 0.4292 0.3126 22.96% 0.4032 0.2519 36.84% 0.4288 0.3053 28.39%
STD 0.0644 0.0723 0.1695 0.0456 0.0636 0.1518 0.1223 0.0589 0.1853 0.0717 0.0890 0.2066 0.0690 0.0759 0.1485
equation below:
xnew =x−µ
σ
where
µ
denotes the expectation of feature
x
and
σ
denotes the
standard deviation.
Depending on the feature characteristics of datasets, these 3
categories of normalization methods may lead to diered analysis
results.
4.2 Feature Representation
To exploit the deeper correlationship between EEG signals, we
adopt Autoencoder to have a beer representation of EEG. e
Autoencoder [
16
] is an unsupervised machine learning algorithm
that aims to explore a lower-dimensional representation of high-
dimensional input data for dimensionality reduction. In structure,
Autoencoder is a multi-layer back propagation neural network
that contains three types of layers: the input layer, the hidden
layer, and the output layer. e procedure from the input layer
to the hidden layer is called encoder while the procedure from
the hidden layer to the output layer is called decoder. Both the
encoder and the decoder yield a set of weights
W
and biases
b
.
Autoencoder is called either Basic Autoencoder when there is only
one hidden layer or Stacked Autoencoder when there are more than
one hidden layers. Based on our prior experiment experience, basic
Autoencoder works beer than stacked Autoencoder when dealing
with EEG signals. erefore, in this paper, we adopt the basic
Autoencoder structure.

Multi-Person Brain Activity Recognition Mobiquitous 2017, Nov. 2017, MELBOURNE, AU
Figure 1: e methodology owchart. e collected EEG data ow into the Feature Representation component to seek for the
appropriately representation and interpretation. Iiand I0
iseparately indicate the input and output EEG data. xi,hi, and x0
i
indicate the neurons in the input layer, the hidden layer and the output layer, respectively. e learned feature representation
hwill be sent to an XGBoost classier with Ktrees. e classier’s predict result is corresponding to the user’s brain activity,
which indicates the user’s intention such as closing eye, moving le hand or moving right hand.
Let
X={Xi|i=
1
,
2
· · · ,N},X∈RN,Xi∈Rd
be the entire
training data (unlabeled), where
Xi
denotes the
i
-th sample,
N
de-
notes the number of training samples, and
d
denotes the number of
elements in each sample.
hi={hij |j=
1
,
2
,· · · ,M},hi∈RM
rep-
resents the learned feature in the hidden layer for the
i
-th sample,
where
M
denotes the number of neural units in current layer (the
number of elements in
hi
). For simplicity, we use
x
and
h
to repre-
sent the input data and the data in the hidden layer, respectively.
First, the encoder transforms the input data
x
to the correspond-
ing representation hby the encoder weights Wen and the encoder
biases ben :
h=Wen x+be n
en, the decoder transforms the hidden layer data
h
to the output
layer data
x0
by the decoder weights
Wde
and the decoder biases
bde :
x0=Wde h+bd e
e function of the decoder is to reconstruct the encoded feature
h
and make the reconstructed data
x0
as similar to the input data
x
as possible. e discrepancy between
x
and
x0
is calculated by the
MSE (mean squared error) cost function which is optimized by the
RMSPropOptimizer.
In summary, training Autoencoder is the task of optimizing
the parameters to achieve the minimum cost between the input
x
and the reconstructed data
x0
. At last, the hidden layer data
h
would contain the rened information. Such information can be
regarded as representation of the input data, which is also the nal
outcome of Autoencoder. In above formulation, the dimension
of the input data
x
and the rened feature (the hidden layer data
h
) are
d
and
M
, respectively. e function of the Autoencoder is
either dimensionality reduction if
d>M
or dimensionality ascent
if d<M.
4.3 Brain Activity Recognition
To recognize the brain activity based on the represented feature, in
this section, we employ the XGBoost [
4
] classier. XGBoost, also
known as Extreme Gradient Boosting, is a supervised scalable tree
boosting algorithm derived from the concept of Gradient Boosting
Machine [
10
]. Compared with gradient boosting algorithm, XG-
Boost proposes a more regularized model formalization to prevent
over-ing, with the engineering goal of pushing the limit of com-
putation resources for boosted tree algorithms to achieve beer
performance.
Consider
n
sample pairs
D={(xi0,yi0)}
,
(|D|=n,xi0∈Rm,yi0∈
R)
where
xi0
denotes a
m
-dimensional sample and
yi0
denotes the
corresponding label. XGBoost aims to predict the label
˜
yi0
of every
given sample xi0.
e XGBoost model is the ensemble of a set of classication and
regression trees (CART), each having its leaves and corresponding
scores. e nial results of tree ensemble is the sum of all the
individual trees. For a tree ensemble model of
K0
trees, the predict
output is:
˜
yi0=
K
Õ
k0=1
fk0(xi0),fk0∈F
where Fis the space of all trees and fk0denotes a single tree.
e objective function of XGBoost includes loss function and
regularization. e loss function evaluates the dierence between
each ground truth label
yi0
and the predict result
˜
yi0
. It can be cho-
sen based on various conditions such as cross-entropy, logistic, and
mean square error. e regularization part is the most outstanding
contribution of XGBoost. It calculates the complexity of the model
and a more complex structure brings larger penalty.
e objective function is dened as:
Ψ=
n
Õ
i
l(˜
yi0,yi0)+
K
Õ
k0
Ω(fk0)(1)
where
l(˜
yi0,yi0)
is the loss function and
Ík0Ω(fk0)
is the regular-
ization term. e complexity of a single tree is calculated as
Ω(fk0)=γT+1
2λkωk2(2)
where
T
is the number of leaves in the tree,
kωk2
denotes the
square of the L2-norm of the weights in the tree,
γ
and
λ
are the

Mobiquitous 2017, Nov. 2017, MELBOURNE, AU Xiang Zhang, Lina Yao, Dalin Zhang, Xianzhi Wang, an Z. Sheng, and Tao Gu
coecients. e regularized objective helps deliver a model of
simple structure and predictive functions. More specically, the rst
term,
Ω
, penalizes complex structures of the tree (fewer leaves lead
to a smaller
Ω
), while the second term penalizes the overweights
of individual trees in case of the overbalanced trees dominating
the model. Moreover, the second term helps smooth the learned
weights to avoid overing.
5 EXPERIMENT
In this section, we evaluate the proposed approach on a public EEG
dataset and report the results of our experimental studies. Firstly,
we introduce the experimental seing and evaluation criterion.
en, we provide the classication results, followed by the analysis
of inuencing factors (e.g., normalization method, training data
size, and neuron size in the Autoencoder hidden layer). Additional
experiments are conducted to study the eciency and robustness
by comparing our approach with the state-of-the-art methods.
5.1 Experimental Setting
We use the EEG data from PhysioNet eegmmidb (EEG motor move-
ment/imagery database) database
1
, a widely used EEG database
collected by the BCI2000 (Brain Computer Interface) instrumenta-
tion system
2
[
22
], to evaluate the proposed method. In particular,
the data is collected by the BCI 2000 system, which owns 64 chan-
nels and an EEG data sampling rate of 160 Hz. During the collection
of this database, the subject sits in front of one screen and performs
the corresponding action as one target appears in dierent edges of
the screen. According to the tasks, dierent annotations are labeled
and can be downloaded from PhysioBank ATM
3
. e actions in
dierent tasks are as follows:
Task 1: e subject closes his or her eyes and keeps relax.
Task 2: A target appears at the le side of the screen and then
the subject focuses on the le hand and imagines he/she is opening
and closing the le hand until the target disappears.
Task 3: A target appears at the right side of the screen and then
the subject focuses on the right hand and imagines he/she is opening
and closing the right hand until the target disappears.
Task 4: A target appears on the top of the screen, and the subject
focuses on both hands and imagines he/she is opening and closing
both hands until the target disappears.
Task 5: A target appears on the boom of the screen, and the
subject focuses on both feet and imagines he/she is opening and
closing both feet until the target disappears.
Specically, we select 560,000 EEG samples from 20 subjects
(28,000 samples each subject) for our experiments. Every sample is
one vector which includes 64 elements corresponding to 64 chan-
nels. Each sample corresponding to one task (from task 1 to task 5
separately is eye closed,focus on le hand,focus on right hand,focus
on both hands and focus on both feet). Every task is labeled as one
class, and there are totally 5 classes labels (from 0 to 4).
5.2 Evaluation
Basic denitions related to classication problems include:
1hps://www.physionet.org/pn4/eegmmidb/
2hp://www.schalklab.org/research/bci2000
3hps://www.physionet.org/cgi-bin/atm/ATM
•
True Positive (TP): the ground truth is positive and the
prediction is positive;
•
False Negative (FN): the ground truth is positive but the
prediction is negative;
•
True Negative (TN): the ground truth is negative and the
prediction is negative;
•
False Positive (FP): the ground truth is negative but the
prediction is positive;
Based on these concepts, we dene criteria to evaluate the perfor-
mance of the classication results as follows:
Accuracy.
e proportion of all correctly predicted samples.
Accuracy is a measure of how good a model is.
accur acy=T P +F N
FP +F N +T P +T N
e test error used in this paper refers to the incorrectly predicted
samples’ proportion, which equals to 1 minus accuracy.
Precision.
e proportion of all positive predictions that are
correctly predicted.
Pr ecision =T P
T P +F P
Recall.
e proportion of all real positive observations that
are correctly predicted.
Recall =T P
T P +F N
F1 Score.
A ‘weighted average’ of precision and recall. e
higher F1 score, the beer the classication performance.
F1Scor e =2pr ecision ∗recall
precision +recall
ROC.
e ROC (Receiver Operating Characteristic) curve
describes the relationship between TPR (True Positive Rate) and
FPR (False Positive Rate) at various threshold seings.
AUC.
e AUC (Area Under the Curve) represents the area
under the ROC curve. e value of AUC drops in the range
[
0
.
5
,
1
]
.
e higher the AUC, the beer the classier.
5.3 Experiments and Results
In our experiments, the Autoencoder model is trained by the train-
ing dataset then the testing dataset is fed into the trained Autoen-
coder model for feature extraction. e extracted features of the
training dataset are used by the XGBoost classier, which will be
evaluated by the features of the testing dataset. e number of
neurons in the input and output layers in the Autoencoder model
xed at 64 (the input EEG data contains 64 dimensions), and the
learning rate is set as 0
.
01. Parameter tuning experience shows
that Autoencoder performs beer with more hidden layer neu-
rons. For XGBoost, we set the objective function as somax for
multi-class classication through pre-experiment experience. e
parameters of XGBoost are selected based on the parameters tuning
document
4
. More specically, we set the learning rate
η=
0
.
7, the
parameter related to the minimum loss reduction and the number
of leaves
дamma =
0, the maximum depth of a tree
maxd ep t h =
6
(too large
maxd ep t h
may lead to overing), the subsampling ra-
tio of training instance
subsample =
0
.
9 (to prevent overing),
4hps://github.com/dmlc/xgboost/blob/master/doc/parameter.md

Multi-Person Brain Activity Recognition Mobiquitous 2017, Nov. 2017, MELBOURNE, AU
Table 3: e confusion matrix of 5-class EEG classication
and the performance evaluation (Precision, Recall, F1 score,
and AUC)
Ground Truth Evaluation
Predict
Label
0 1 2 3 4 Precision Recall F1 Score AUC
0 3745 0 300 235 417 0.7973 0.7703 0.7836 0.9454
1 385 7857 515 445 488 0.8108 0.9219 0.8628 0.9572
2 245 174 3929 341 212 0.8017 0.7556 0.7780 0.9492
3 129 125 209 3304 153 0.8429 0.7294 0.7820 0.9506
4 358 367 247 205 3397 0.7427 0.7279 0.7352 0.9258
total amount 4862 8523 5200 4530 4667 3.9954 3.9051 3.9416 4.7282
average 0.7991 0.7810 0.7883 0.9456
and
numcl as s =
5, since we have samples of 5 categories. All the
other parameters are set as default value. Without specic explana-
tion, all the Autoencoder and XGBoost classiers are taking above
parameters seing.
e hardware used in experiments is a GPU-accelerated ma-
chine with Nvidia Titan X pascal GPU, 768G memory, and 1.45TB
PCIe based SSD. e training time is listed in related experiments,
respectively.
Multi-person Multi-class EEG Classication.
To evaluate
the proposed approach, 560,000 EEG sample patches are utilized
in this experiment. Each sample patch contains a feature vector
of 64 dimensions and a ground truth label. e raw EEG data is
normalized by the z-score scaling method and then randomly split
into training dataset (532,000 samples) and testing dataset (28,000
samples). e representative features are extracted by Autoencoder
with 121 hidden neurons and are input to the XGBoot classier. e
confusion matrix of the results is listed in Table 3. e classication
accuracy of 28,000 testing samples (from 20 subjects and belong
to 5 classes) is 0
.
794. e average precision, recall, F1 score, and
AUC are 0
.
7991, 0
.
781, 0
.
7883, and 0
.
9456, respectively. Among
the evaluation standards, the class 1 obtains the highest precision,
recall, F1 score, and AUC. is means that the class 1 samples have
the most obvious divergence and are most distinguishable. On
the contrary, class 4 is most confusing. is conclusion is highly
consistent with our similarity analysis results in Section 3. From
the ROC curve, shown as Figure 2, we can deduce to the same
conclusion. All the classes achieved the AUC higher than 0
.
92,
indicating that the classier is steady and of high quality, according
to the characteristics of AUC mentioned in Section 5.2.
Eect of Normalization Method.
e Autoencoder compo-
nent regards the input data as the training target and calculates the
discrepancy between them for error back propagation. is charac-
ter of Autoencoder determines that the feature extraction quality
and training cost are aected by the amplitude of the input data.
In data pre-processing stage, the data values are directly related to
the normalization method.
To explore the impact of the normalization method, 560,000 EEG
samples from 20 subjects are randomly split into a training dataset
of 532,000 samples (95% proportion) and a testing dataset of 28,000
samples (5% proportion). By seing 121 neurons for the hidden
layer of Autoencoder, the XGBoost test error under three kinds
of normalization methods is shown in Figure 3. e gure shows
that the z-score scaling normalization earns the lowest test error
while the unity normalization obtains the highest test error. All the
curves trend to convergence aer 1,600 iterations. Without specic
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
False Positive Rate
0
0.2
0.4
0.6
0.8
1
True Positive Rate
class 0
class 1
class 2
class 3
class 4
Figure 2: ROC curve for 5-class classication by XGBoost.
Five curves separately indicate the ROC curve of ve classes.
e dotted diagonal line denotes the random classier
where TPR=FPR. e closer the ROC curve to the upper le
corner, the better performance the classier has. It is clear
to notice that the class 1 has the best classication perfor-
mance.
0 200 400 600 800 1000 1200 1400 1600 1800
The number of iterations
0.2
0.3
0.4
0.5
0.6
0.7
Test error
Min-Max
Z-score
Unity
Figure 3: e eect of normalization method. e three test
error curves denote Min-Max, Z-score, and Unity normaliza-
tion method, respectively.
explanation, all the remaining experiments in this paper use the
z-score scaling method.
Eect of Training Data Size.
We explore in this section the
relationship between the classication performance and the train-
ing data size. We design ve experiments with the training data
proportion of 60%, 70%, 80%, 90%, and 95%, respectively. Each exper-
iment is repeated 5 times and the test error’s error bar is shown in
Figure 5. e training time is positively correlated with the training
data proportion. e test error arrives at the lowest point 0
.
206,
with an acceptable training time, while the proportion is 95%. All
the following experiments in this paper will take 95% proportion.
e relationships between test error and the iterations under
various training data proportions are shown in Figure 4. All the
curves trend to convergence aer 1,600 iterations and the higher
proportion leads to lower test error.
Eect of Neuron Size in Autoencoder Hidden Layer.
e
neuron size in the hidden layer of Autoencoder indicates the number
of dimensions of the extracted features. us it has great impact on
the quality of feature extraction as well as the classication results.

Mobiquitous 2017, Nov. 2017, MELBOURNE, AU Xiang Zhang, Lina Yao, Dalin Zhang, Xianzhi Wang, an Z. Sheng, and Tao Gu
Table 4: Comparison of various classication methods. e rst nine groups investigate the proper EEG data classier and
the last 7 groups illustrate the most ecient feature representation method.
No. 1 2 3 4 5 6 7 8
Classier SVM RNN LDA RNN+SVM CNN DT AdaBoost RF
Acc 0.3333 0.6104 0.3384 0.6134 0.5729 0.3345 0.3533 0.6805
No. 9 10 11 12 13 14 15 16
Classier XGBoost PCA+XGBoost PCA+AE+XGBoost EIG+AE+XGBoost EIG+PCA+XGBoost DWT+XGBoost Stacked AE+XGBoost AE+XGBoost
Acc 0.7453 0.7902 0.6717 0.5125 0.6937 0.7221 0.7048 0.794
0 200 400 600 800 1000 1200 1400 1600 1800
The number of iterations
0.2
0.3
0.4
0.5
0.6
0.7
Test error
95%
90%
80%
70%
60%
Figure 4: e relationships between test error and the itera-
tions under various training data proportions
55 60 65 70 75 80 85 90 95 100
Training data proportion (%)
0.2
0.22
0.24
0.26
0.28
0.3
0.32
Test error
4.5
5
5.5
6
6.5
7
7.5
8
8.5
Training time (s)
10 4
Figure 5: e relationship between the test error with error
bars, the training time and the training data proportion
We design the experiment with the neuron size ranges from 30 to
200 and the experimental results (the test error and the training
time) are shown in Figure 6.
In the rst stage (0-120), the test error keeps decreasing with the
increase of the neuron size; in the second stage (larger than 120),
the test error stands at around 0
.
21 with slight uctuation. e
training time curve has a linear relationship with the neuron size
on the whole. Although the gap between the test error curve and
the training time curve arrives at the minimum around 100 neurons,
the test error is still high. e test error reaches the boom while
the neuron size is 121, and the training time is acceptable at this
point. Moreover, the test error curve keeps steady aer 121. We set
the hidden layer neuron size for all other experiments as 121.
5.4 Comparison
In our approach, we employ XGBoost as the classier to classify the
rened EEG features yielded by Autoencoder. To demonstrate the
eciency of this method, in this section, we compare the proposed
approach with several widely used classication methods. All the
20 40 60 80 100 120 140 160 180 200 220
The neuron size in Autoencoder hidden layer
0.2
0.25
0.3
0.35
Test error
2
4
6
8
10
12
14
Training time (s)
10 4
64
Dimensionality
Ascent
Dimensionality
Reduction
Figure 6: e eect of neuron size in Autoencoder hidden
layer. Since the input data is 64-dimension (marked as red
line), the le part (smaller than 64) is dimensionality reduc-
tion area while the right part (larger than 64) is dimension-
ality ascent area.
classiers work on the same EEG dataset and their corresponding
performance is listed in Table 4.
In Table 4, LDA denotes Linear Discriminant Analysis; SVM
denotes Support Vector Machine; RNN denotes (Recurrent Neuron
Network) alongside LSTM denotes Long-Short Term Memory (kind
of RNN architecture); AdaBoost denotes Adaptive Boosting; RF
denotes Random Forest; DT denotes Decision Tree; EIG denotes
the eigenvector-based dimensionality reduction method used in
Eigenface recognition
5
; PCA denotes Principal Components Analy-
sis which is a commonly used dimensionality reduction method;
DWT denotes Discrete Wavelet Transform, which is the wavelet
transformation with the wavelets discretely sampled. e stacked
Autoencoder contains 3 hidden layers with 100, 121, 100 neurons,
respectively.
e comparison is divided into two aspects: the classier and
the feature representation method. At rst, we classify our dataset
separately by 9 commonly used sensing data classier (e.g., SVM, RF,
RNN, and CNN) to investigate the most suitable classier for raw
EEG data. en 7 categories feature extraction method (e.g., PCA,
AE, and DWT) are conducted to investigate the most appropriately
EEG feature representation approach. e comparison results show
that the XGBoost classier outperforms its counterpart (without
pre-processing and feature extraction) and obtains the accuracy
of 0
.
74, which means that XGBoost is more suitable to solve this
problem. On the other hand, some feature extraction is positive to
the classication whilst some are negative. rough the comparison,
5hp://www.vision.jhu.edu/teaching/vision08/Handouts/case study pca1.pdf

Multi-Person Brain Activity Recognition Mobiquitous 2017, Nov. 2017, MELBOURNE, AU
(a) EEG collection (b) EEG raw data
Figure 7: EEG collection and the raw data . e pure EEG data is selected for recognition and the data, which is contaminated
by eye blink and other noise, is not included in the local dataset (dropped).
Table 5: Comparison of various classication methods over the case study dataset
No. 1 2 3 4 5 6 7 8
Classier SVM RNN LDA RNN+SVM CNN DT AdaBoost RF
Acc 0.356 0.675 0.343 0.6312 0.5291 0.305 0.336 0.6609
No. 9 10 11 12 13 14 15 16
Classier XGBoost PCA+XGBoost PCA+AE+XGBoost EIG+AE+XGBoost EIG+PCA+XGBoost DWT+XGBoost StackedAE+XGBoost AE+XGBoost
Acc 0.6913 0.7225 0.6045 0.4951 0.6249 0.6703 0.6593 0.7485
we nd that Autoencoder (121 hidden neurons) achieves the highest
multi-person classication accuracy as 0.794.
6 CASE STUDY
In this section, to further demonstrate the feasibility of the proposed
approach, we conduct a local experiment and present the classi-
cation result. At rst, we design an EEG collection experiment.
Secondly, we report the recognition classication results under the
optimal hyper-parameters. Aerwards, the tuning procedure of the
optimal hyper-parameters is provided. Subsequently, we show the
comparison between our approach and the state-of-the-art meth-
ods.
6.1 Experimental Setting
is experiment is carried on by 5 subjects (3 males and 2 females)
aged from 24 to 30. During the experiment, the subject wearing the
Emotiv Epoc+
6
EEG collection headset, facing the computer screen
and focus on the corresponding mark which appears on the screen
(shown in Figure 7). e Emotiv Epoc+ contains 14 channels and
the sampling rate is set as 128 Hz. e marks are shown on the
screen and the corresponding brain activities and labels used in this
paper are listed in Table 6. Summarily, this experiment contains
172,800 samples with 34,560 samples for each subject.
6hps://www.emotiv.com/product/emotiv-epoc-14- channel-mobile- eeg/
Table 6: Mark in experiment and corresponding brain activ-
ity and label in case study
Mark up arrow down arrow le arrow right arrow central cycle close eye
Brain Activity upward downward leward rightward center relax
Label 0 1 2 3 4 5
Table 7: e confusion Matrix and the evaluation
Ground truth Evaluation
Predicted label
0 1 2 3 4 5 Precision Recall F1 AUC
0 2314 197 126 188 258 57 0.7369 0.7555 0.7461 0.8552
1 165 2271 153 171 214 75 0.7448 0.7407 0.7428 0.8395
2 131 176 2363 180 164 74 0.7652 0.7930 0.7788 0.8931
3 177 156 142 2179 216 85 0.7374 0.7230 0.7301 0.8695
4 191 190 118 219 2269 79 0.7401 0.6986 0.7187 0.8759
5 85 76 78 77 127 1539 0.7765 0.8062 0.7911 0.9125
6.2 Recognition Results and Comparison
e dataset is divided into a training set (155,520 samples) and a
testing set (17,280 samples). ere are 9 mini-batches and the batch
size is 17,280. All the other parameters are the same as listed in
Section 5. e proposed approach achieves the 6-class classica-
tion accuracy of
0.7485
. e confusion matrix and evaluation is
reported in Table 7. Clearly, the 5th class brain activity (eye closed
and keep relax) has the highest precision and is the easiest activity
to be recognized.
Subsequently, to demonstrate the eciency of the proposed ap-
proach, we compare our method with the state-of-the-art methods
and the results are shown in Table 5.

Mobiquitous 2017, Nov. 2017, MELBOURNE, AU Xiang Zhang, Lina Yao, Dalin Zhang, Xianzhi Wang, an Z. Sheng, and Tao Gu
7 CONCLUSION
In this paper, we have focused on multi-class EEG signal classica-
tion based on EEG data that come from dierent subjects (multi-
person). To achieve this goal, we aim at discovering the paerns
in the discrepancy between dierent EEG classes with robustness
over the dierence between various subjects. Firstly, we analyze
three widely used normalization methods in pre-processing stage.
en, we feed the normalized EEG data into the Autoencoder and
train the Autoencoder model. Autoencoder transforms the original
64-dimension features to 121-dimension features and essentially
maps the data to a new feature space when meaningful features
play a dominating role. Finally, we evaluate our approach over an
EEG dataset of 560,000 samples (belongs to 5 categories) and achieve
the accuracy of
0.794
. Compared with the accuracy of around 0
.
34
achieved by traditional methods (e.g., SVM, AdaBoost, Decision
Tree, and RNN), our results 0
.
794 show signicant improvement.
Furthermore, we explore the eect of two factors (the training data
size and the neuron size in Autoencoder hidden layer) on the train-
ing results. At last, we conduct a case study to gather 6 categories
of brain activities and obtain the classication accuracy of 0.7485.
As part of our future work, we will build multi-view model of
multi-class EEG signals to improve the classication performance.
In particular, we plan to establish multiple models with each single
model dealing with a single class. Following this philosophy, the
correlation between test sample and each model can be calculated
in the test stage and the sample can be classied to the class with
minimum correlation coecient. Besides, the work introduced in
this paper only represents our preliminary study on exploring the
common paerns of brain activities. Establishing a universal and
ecient EEG classication model will be a major goal of our future
research.
REFERENCES
[1]
Nguyen e Hoang Anh, Tran Huy Hoang, Vu Tat ang, TT yen Bui, and
others. 2016. An Articial Neural Network approach for electroencephalographic
signal classication towards brain-computer interface implementation. In Com-
puting & Communication Technologies, Research, Innovation, and Vision for the
Future (RIVF), 2016 IEEE RIVF International Conference on. IEEE, 205–210.
[2]
Mahnaz Arvaneh, Cuntai Guan, Kai Keng Ang, and Chai ek. 2013. Optimizing
spatial lters by minimizing within-class dissimilarities in electroencephalogram-
based brain–computer interface. IEEE Transactions on Neural Networks and
Learning Systems 24, 4 (2013), 610–619.
[3]
Saugat Bhaacharyya, Abhronil Sengupta, Tathagatha Chakraborti, Amit Konar,
and DN Tibarewala. 2014. Automatic feature selection of motor imagery EEG
signals using dierential evolution and learning automata. Medical & biological
engineering & computing 52, 2 (2014), 131–139.
[4]
Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system.
In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining. ACM, 785–794.
[5]
D De Venuto, VF Annese, M de Tommaso, E Vecchio, and AL Sangiovanni Vincen-
telli. 2015. Combining EEG and EMG signals in a wireless system for preventing
fall in neurodegenerative diseases. In Ambient Assisted Living. Springer, 317–327.
[6]
Ridha Djemal, Ayad G Bazyed, Kais Belwa, Soen Gannouni, and Walid
Kaaniche. 2016. ree-Class EEG-Based Motor Imagery Classication Using
Phase-Space Reconstruction Technique. Brain Sciences 6, 3 (2016), 36.
[7]
Lijuan Duan, Yanhui Xu, Song Cui, Juncheng Chen, and Menghu Bao. 2016.
Feature extraction of motor imagery EEG based on extreme learning machine
auto-encoder. In Proceedings of ELM-2015 Volume 1. Springer, 361–370.
[8]
Manuel JA Eugster, Tuukka Ruotsalo, Michiel M Spap
´
e, Ilkka Kosunen, Oswald
Barral, Niklas Ravaja, Giulio Jacucci, and Samuel Kaski. 2014. Predicting term-
relevance from brain signals. In Proceedings of the 37th International ACM SIGIR
Conference on Research & Development in Information Retrieval. 425–434.
[9]
Oliver Faust, U Rajendra Acharya, Hojjat Adeli, and Amir Adeli. 2015. Wavelet-
based EEG processing for computer-aided seizure detection and epilepsy diagno-
sis. Seizure 26 (2015), 56–64.
[10]
Jerome H Friedman. 2001. Greedy function approximation: a gradient boosting
machine. Annals of Statistics (2001), 1189–1232.
[11]
Hongfei Ji, Jie Li, Rongrong Lu, Rong Gu, Lei Cao, and Xiaoliang Gong. 2016. EEG
classication for hybrid brain-computer interface using a tensor based multiclass
multimodal analysis scheme. Computational Intelligence and Neuroscience 2016
(2016), 51.
[12]
Hyohyeong Kang and Seungjin Choi. 2014. Bayesian common spatial paerns
for multi-subject EEG classication. Neural Networks 57 (2014), 39–50.
[13]
Wei Tuck Lee, Humaira Nisar, Aamir S Malik, and Kim Ho Yeap. 2013. A brain
computer interface for smart home control. In Consumer Electronics (ISCE), 2013
IEEE 17th International Symposium on. IEEE, 35–36.
[14]
Hardik Meisheri, Nagaraj Ramrao, and Suman K Mitra. 2016. Multiclass common
spatial paern with artifacts removal methodology for EEG signals. In Compu-
tational and Business Intelligence (ISCBI), 2016 4th International Symposium on.
IEEE, 90–93.
[15] Gernot R M ¨
uller-Putz, Patrick Ofner, Andreas Schwarz, Joana Pereira, Andreas
Pinegger, Catarina Lopes Dias, Lea Hehenberger, Reinmar Kobler, and An-
dreea Ioana Sburlea. 2017. Towards non-invasive EEG-based arm/hand-control
in users with spinal cord injury. In Brain-Computer Interface (BCI), 2017 5th
International Winter Conference on. IEEE, 63–65.
[16]
anh Nguyen, Saeid Nahavandi, Abbas Khosravi, Douglas Creighton, and Imali
Heiarachchi. 2015. EEG signal analysis for BCI application using fuzzy system.
In Neural Networks (IJCNN), 2015 International Joint Conference on. IEEE, 1–8.
[17]
Tobie Erhard Olivier, Shengzhi Du, Barend Jacobus van Wyk, and Yskandar
Hamam. 2016. Independent components for EEG signal classication. In Pro-
ceedings of the 2016 International Conference on Intelligent Information Processing.
ACM, 33.
[18]
Gamaleldin Osman, Daniel Friedman, and Lawrence J Hirsch. 2017. Diagnosing
and Monitoring Seizures in the ICU: e Role of Continuous EEG for Detection
and Management of Seizures in Critically Ill Patients, Including the Ictal-Interictal
Continuum. In Seizures in Critical Care. Springer, 31–49.
[19]
Mohammad Rahman, Wanli Ma, Dat Tran, and John Campbell. 2012. A compre-
hensive survey of the feature extraction methods in the EEG research. Algorithms
and Architectures for Parallel Processing (2012), 274–283.
[20]
Luca Rossini, Dario Izzo, and Leopold Summerer. 2009. Brain-machine interfaces
for space applications. In Engineering in Medicine and Biology Society, 2009. EMBC
2009. Annual International Conference of the IEEE. IEEE, 520–523.
[21]
Roberto Santana, Laurent Bonnet, Jozef Leg
´
eny, and Anatole L
´
ecuyer. 2012.
Introducing the use of model-based evolutionary algorithms for EEG-based
motor imagery classication. In Proceedings of the 14th Annual Conference on
Genetic and Evolutionary Computation. ACM, 1159–1166.
[22]
Gerwin Schalk, Dennis J McFarland, ilo Hinterberger, Niels Birbaumer, and
Jonathan R Wolpaw. 2004. BCI2000: a general-purpose brain-computer interface
(BCI) system. IEEE Transactions on Biomedical Engineering 51, 6 (2004), 1034–
1043.
[23]
T Shiratori, H Tsubakida, A Ishiyama, and Y Ono. 2015. ree-class classication
of motor imagery EEG data including rest state using lter-bank multi-class
Common Spatial paern. In Brain-Computer Interface (BCI), 2015 3rd International
Winter Conference on. IEEE, 1–4.
[24]
Laurent V
´
ezard, Pierrick Legrand, Marie Chavent, Fr
´
ed
´
erique Fa
¨
ıta-A
¨
ınseba, and
Leonardo Trujillo. 2015. EEG classication for the detection of mental states.
Applied So Computing 32 (2015), 113–131.
[25]
Maitreyee Wairagkar, Ioannis Zoulias, Victoria Oguntosin, Yoshikatsu Hayashi,
and Slawomir Nasuto. 2016. Movement intention based Brain Computer Interface
for Virtual Reality and So Robotics rehabilitation using novel autocorrelation
analysis of EEG. In Biomedical Robotics and Biomechatronics (BioRob), 2016 6th
IEEE International Conference on. IEEE, 685–685.
[26]
Deng Wang, Duoqian Miao, and Gunnar Blohm. 2012. Multi-class motor imagery
EEG decoding for brain-computer interfaces. Frontiers in Neuroscience 6 (2012),
151.
[27]
Bahia Yahya-Zoubir, Maouia Bentlemsan, ET-TahirZemouri, and Karim Ferroudji.
2015. Adaptive Time Window for EEG-based Motor Imagery Classication. In
Proceedings of the International Conference on Intelligent Information Processing,
Security and Advanced Communication. ACM, 83.
[28]
Xiang Zhang, Lina Yao, Chaoran Huang, an Z Sheng, and Xianzhi Wang. 2017.
Intent Recognition in Smart Living rough Deep Recurrent Neural Networks.
e 24th International Conference on Neural Information Processing (2017).
[29]
Yong Zhang, Bo Liu, Xiaomin Ji, and Dan Huang. 2016. Classication of EEG
Signals Based on Autoregressive Model and Wavelet Packet Decomposition.
Neural Processing Leers (2016), 1–14.
[30]
Yu Zhang, Guoxu Zhou, Jing Jin, Xingyu Wang, and Andrzej Cichocki. 2014.
Frequency recognition in SSVEP-based BCI using multiset canonical correlation
analysis. International Journal of Neural Systems 24, 04 (2014), 1450013.
[31]
Yu Zhang, Guoxu Zhou, Jing Jin, Qibin Zhao, Xingyu Wang, and Andrzej Ci-
chocki. 2016. Sparse bayesian classication of EEG for brain–computer interface.
IEEE Transactions on Neural Networks and Learning Systems 27, 11 (2016), 2256–
2267.