XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
Real-time decoding for fNIRS-based Brain Computer
Interface using adaptive Gaussian mixture model
classifier and Kalman estimator
Yao Zhang
College of Precision Instrument and
Optoelectronics Engineering
Tianjin University
Tianjin, China
viyoo_zhang@tju.edu.cn
Bingyuan Wang
College of Precision Instrument and
Optoelectronics Engineering
Tianjin University
Tianjin, China
wangbingyuan@tju.edu.cn
Feng Gao
College of Precision Instrument and
Optoelectronics Engineering
Tianjin Key Laboratory of Biomedical
Detecting Techniques and Instruments
Tianjin University
Tianjin, China
gaofeng@tju.edu.cn
Abstract— In this paper, we investigate a real-time analysis of
NIRS data by using an unsupervised Gaussian mixture model
adaptive classifier (GMMAC) for a framework consisting of the
general linear model (GLM) and the Kalman estimator to
improve decoding accuracy. Simulation experiment results
demonstrate that the GMMAC (91.9%) perform significantly
better than linear discriminant analysis (LDA, 51.9%) and
support vector machines (SVM, 51.1%) classifiers in binary
problems.
Keywords—Functional near-infrared spectroscopy, Brain-
computer interface, adaptive real-time decoding, General linear
model, Kalman estimator, Gaussian mixture model, variational
inference
I. I
NTRODUCTION
Functional near-infrared spectroscopy (fNIRS) is the
relatively new non-invasive neuroimaging modality for brain-
computer interface (BCI) to measure the changes of cerebral
blood oxygenation associated with brain activities [1, 2]. These
changes are caused by the concentration variation of oxy-
hemoglobin (HbO) and deoxy-hemoglobin (HbR), which are
primary absorbing chromophores in the capillaries of the brain
[1, 3]. Since fNIRS is less susceptible to environmental noises
and electrophysiological artifacts compared to EEG [3], it is
believed that fNIRS-BCI is ideally suited for the development
of portable and potentially clinical real-time system for future
applications.
Over the past few years, several research groups attempt to
develop real-time data analysis schemes to decoding different
experimental paradigms and assess brain activation at the same
time as data collection, and combination of more powerful and
faster computers [4]. At the same time, real-time fNIRS data
analysis can use biofeedback to develop experimental designs
for subjects. Therefore, an adaptive real-time high-precision
decoding method is proposed and applied to clinical
neurofeedback and neural rehabilitation training.
In our current paper, we have taken advantages of a real-
time NIRS data analysis framework combined with the GLM
and the Kalman estimator. The GLM has been established as a
standard method for fMRI data analysis, and has also been
applied to fNIRS studies using block and event-related designs
[5]. In group analysis, the GLM is used to produce a subject-
specific contrast relationship from X-weight in regression to the
hemodynamic response function (HRF). In our model, we used
the Kalman estimator to estimate the X coefficients recursively
[6]. Classification techniques are used to identify the different
brain signals that are generated by the subjects. In this article,
we present the GMMAC for adaptive classification of fNIRS
data. Then we use simulated experiments to test this method.
These simulations contained in fNIRS neural activation
patterns changes, including shift of activation region,
contraction and expansions of activation region, during a two-
class decoding task. Our results show that the GMMAC applied
to real-time decoding can well track the activation region
changes, and can get high-decoding accuracy.
II. M
ETHODS
A. General linear model
Time series of HbO and HbR signal data were analyzed
using a GLM [4], the model can be expressed as follows:
CX
ε
Δ= +H
, (1)
where the
ND
CR
×
Δ∈
represents the changes of hemodynamic
variable at the N points time series in D channels. There
NL
×
∈H
(where L is the number of explanatory variables) is
known as the design matrix including a set of explanatory
variables and models the observed NIRS time series. In our job,
the design matrix H contains five explanatory variables
representing canonical HRF (cHRF), heartbeat signal, Mayer
wave signal, breathing signal, and constant respectively. The
LD
R
×
∈
is the regression coefficients quantifying the
magnitude of the explanatory variables. The final term
ε
is the
residual error obeying independent zero-mean Gaussian
distribution. We used the Kalman estimator to solve the (1) for
real-time estimation of each point time.
B.
Kalman estimator
The Kalman filter method is an adaptive tracking scheme
that perform an optimal estimation of the state of a process
using a recursively regularized linear inversion routine [4]. In
this study, the Kalman filter was used as the GLM coefficients
2018 Asia Communications and Photonics Conference (ACP)
978-1-5386-6158-1/18/$31.00 ©2018 IEEE