Spatio-temporal Pattern Analysis for EEG Classification in
Rapid Serial Visual Presentation Task
Bowen Li
1
, Zhiwen Liu
1
, Xiaorong Gao
2
, Yanfei Lin
1
*
1
School of Information and Electronics, Beijing Institute of Technology. Beijing, China
2
School of Medicine, Tsinghua University. Beijing, China
1
{3120160348, zwliu, linyf}@bit.edu.cn
2
gxr-dea@tsinghua.edu.cn
ABSTRACT
This study will explore an algorithm of spatio-temporal pattern
analysis for electroencephalographic (EEG) classification in the
rapid serial visual presentation (RSVP) task. In this algorithm, the
spatial low-rank and temporal-frequency sparse priors are
exploited to train the supervised spatial and temporal filters. The
discriminant features are extracted by the supervised spatio-
temporal filters and classified by support vector machine. The
EEG signals were recorded from a total of 12 subjects under
RSVP task and were used as training and testing data. The
average true positive rate of classification is 79%, and the average
false positive rate is only 3.4%. The classification results show
that the proposed algorithm has better performance in the target
detection than HDCA and SWFP.
CCS Concepts
• Applied computing ➝ Computer-aided design.
Keywords
EEG; Spatio-temporal pattern; Discriminant features; Low-rank;
Sparse prior
1. INTRODUCTION
Recently, the development in neuroscience has led to an intense
interest in Brain-computer Interface (BCI) studies. The BCI
systems can directly translate the human intention for computers
or machines by decoding the neural signals, hence the users can
control devices without the peripheral pathways [1]. In the BCIs,
electroencephalographic (EEG) is the most widely used neural
signal. The BCIs have been applied as the assistive technology for
disabled and healthy populations, such as the brain spellers [2, 3],
brain-controlled wheelchairs [4, 5] and intelligent manipulators
[6].
In addition, an important type of the BCIs focuses on enhancing
the visual perception abilities and allows human to quickly detect
the interest targets in the massive images. The rapid serial visual
presentation (RSVP) task is used in these BCIs, in which users are
able to detect the targets in the high-speed image stream [7]. The
targets and non-targets evoke distinct event-related potentials
(ERPs) in the EEG recordings [8]. Since the ERPs are phase-
* Corresponding author
locked relative to the stimuli, the targets can be accurately located
in the image dataset by classifying the ERPs [9].
In order to determine the category of ERPs in the EEG samples,
many spatial filtering methods have been proposed to analyze the
EEG components and extract the discriminant features. Common
spatial patterns (CSP) provides spatial filters to identify the
discriminant components by maximizing the variance ration of
two classes [10]. Common spatio-temporal patterns (CSTP)
algorithm was specifically constructed to classify the EEG for
RSVP tasks by adding the analysis of ERP temporal pattern based
on the CSP [11]. And xDawn is another method to maximize the
power ratio between target and non-target ERP components,
which can effectively improve classification performance [12].
Besides the traditional spatial filtering methods, the linear
discriminant analysis (LDA) is utilized to obtain the projection
weights which maximize the separation between two categories
and minimize the within-class scatters [13]. Compared to the CSP
methods, the LDA methods can provide better performance for
the ERP classification. Hierarchical discriminant components
analysis (HDCA) is built upon the LDA to extract the spatial and
temporal features [14-16]. Spatially Weighted FLD-PCA (SWFP)
also provides the spatio-temporal weights to extract features by
LDA [17].
However, the above algorithms do not make full use of the prior
knowledge and characteristics of ERPs. First, the number of
spatial filters cannot be automatically determined in these
algorithms, since the number of discriminant components is
unknown. The number of spatial filters determines the number of
extracted discriminant components, i.e., each spatial filter extracts
one component. The redundancy of invalid components and the
lack of discriminant components will both affect the classification
performance. Second, the frequency distributions and phases of
discriminant components are different from the background
components. Hence, the temporal weights should be learned for
the specific temporal-frequency components.
In this study, we propose a method of spatio-temporal pattern
analysis for EEG classification by exploiting two priors in a linear
discriminant framework. First, the number of ERPs is typically
less than that of electrodes, which indicates that the spatial
patterns of discriminant components are low-rank. The spatial
low-rank prior can be utilized to automatically determine the
number of spatial filters, which can extract the effective
discriminant components and lead to better classification. Second,
the temporal-frequency distribution of ERPs is sparse on the
wavelet dictionary, since the ERPs are low-frequency and phase-
locked. The sparse prior can be used to conduct a dictionary
learning process for temporal-frequency discriminant component
extraction, which can further increase the separation between
target and non-target. We compared the performance of the
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that
copies bear this notice and the full citation on the first page. To copy
otherwise, or republish, to post on servers or to redistribute to lists,
requires prior specific permission and/or a fee.
ICABB’1, December 19–21, 2019, Seoul, South Korea.