N400 Extraction from Fewer-trial EEG Data Using a
Supervised Signal-to-noise Ratio Maximizer Method
Bowen Li, Zhiwen Liu, and Yanfei Lin*
School of Information and Electronics
Beijing Institute of Technology
Beijing, 100081, P.R.China
Abstract—N400 is a kind of event-related potential (ERP), which
is related to language processing of brain and can be used for the
evaluation of clinical psychological diseases. There still remain
some problems in the accurate N400 waveform extraction from
fewer-trial EEG data under the low signal-to-noise ratio (SNR)
level. In this study, a supervised signal-to-noise ratio maximizer
(SSM) method to obtain N400 waveform from multi-channel
EEG data is proposed. The SSM algorithm designs a spatial filter
for low-rank ERP component and extracts the N400 by 40-trial
EEG datasets of each subject. The algorithm has more excellent
performance in estimating the accurate N400 waveform from
simulation data and real EEG data, compared to SIM and the
regularized SOBI algorithms. The results show that the proposed
method can effectively achieve the N400 extraction from fewer-
trial EEG data.
Keywords-EEG; N400 extraction; fewer-trial; supervised
signal-to-noise ratio maximizer; low-rank component
I. INTRODUCTION
N400 is an event-related potential, which is extensively
used to study neurocognitive mechanism. It is a negative wave
occurring at 400 ms post-stimulus, whose amplitude is
positively correlated with the level of semantic deviation [1]. It
can be used to evaluate the language function of brains and be
helpful to diagnose the clinical diseases [2]. Therefore, it is
very meaningful to extract N400 waveform accurately.
However, there are some difficulties in extracting N400
from fewer-trial EEG data. First, in the previous studies, since
the signal-to-noise ratio of N400 component in EEG data is
very low, grand-average of more than 40-trial EEG data for all
subjects was used to estimate N400 amplitude. However, it is
not practical to clinical application for single subject. Second,
N400 is obtained from the amplitude difference between EEG
signals under consistent and inconsistent conditions. In the
statistical sense, it can be observed by the grand-average of
multiple trials, but not by single trial. Therefore, it is
reasonable to extract N400 from a few trials EEG data. It also
can be done by making full use of the spatial information of
multichannel EEG signals. The spatial information can be
extracted by training spatial filters based on multiple-trial EEG
data. In the previous studies, a variety of spatial filtering
algorithms for ERPs have been proposed, such as principal
component analysis (PCA) [3], independent component
analysis (ICA) combined with wavelet filtering [4], and the
regularized second-order blind identification (SOBI), whose
spatial filter was optimized by focusing on the extraction of
phase-locked components [5]. The above algorithms usually
obtained principal components in the the maximum variance
direction, based on PCA [3]. If there were other phase-locked
components with much larger variance, the objective ERPs
might not be extracted. In pursuing optimal spatial filters to
estimate ERPs, a signal-to-noise ratio maximizer algorithm
(SIM) for event-related potentials was proposed in [6], which
specifically maximized the SNR using an efficient spatial filter
design. xDAWN algorithm was proposed in [7], which
estimated the principal components of ERPs subspace by
providing the best SNR. To solve the small sample problem, a
method for spatially smoothing xDAWN spatial filters was
recently proposed in [8], which gave a subspace constraint to
the parameter space of the spatial filters. SIM and xDAWN
provide effective ERP waveform estimation to classify the
EEG data in the BCI system. However, both of them give
priority to the components with larger variance, which are
similar to those algorithms based on PCA. Therefore, the
interference components with greater variance might exist in
the waveforms estimated using SIM and xDAWN, which
would lead to the inaccuracy of the ERP amplitude.
In order to accurately estimate the N400 waveform from
fewer-trial EEG data, a supervised signal-to-noise ratio
maximizer (SSM) algorithm based on training a denoising
projection and low-rank component analysis is proposed in this
study. Compared to previous methods, the proposed method
has less amount of computation and better performance on the
estimation of N400 amplitude. The efficiency of the algorithm
is analyzed using both simulated and real EEG data.
II. M
ETHOD
The ERP waveform is phase-locked, and the inter-trial
variability in ERP amplitude and latency is typically small
compared to that of spontaneous EEG activities [6]. Since each
channel of ERP waveform is a linear mixture of the same ERP
sources, ERP component is represented by a low rank matrix.
As is mentioned above, an EEG signal can be modeled. Let
k
,
c
,
n
( [1, ]kK ,
[1, ]cC
,
[1, ]nN
) denote the indexes for
the trials, channels, and sampled time points of EEG data,
respectively. The
thk
trial observed EEG signal in time
domain is modeled as follows:
kk
XSN
(1)
where
CN
k
X denotes the observed multi-channel EEG
This work was supported by National Natural Science Foundation o
f
China (No. 61601028).
*The corresponding author: linyf@bit.edu.cn.
2018 11th International Congress on Image and Signal Processing, BioMedical Engineering and Informatics (CISP-BMEI 2018)
978-1-5386-7604-2/18/$31.00 ©2018 IEEE
资源评论
