Physiological Signal Processing for Emotional Feature Extraction

Peng Wu

, Dongmei Jiang

and Hichem Sahli

1,3

VUB-NPU Joint AVSP Research Lab

Department of Electronics and Informatics(ETRO), Vrije Universiteit Brussel (VUB), Pleinlaan 2, Brussels, Belgium

Shaanxi Provincial Key Lab on Speech and Image Information Processing,

Northwestern Polytechnical University (NPU), Xi’an, China

Interuniversity Microelectronics Centre (IMEC), Leuven, Belgium

Keywords:

Signal Denoising, QRS Detection, Electromyography Segmentation.

Abstract:

This paper introduces new approaches of physiological signal processing prior to feature extraction from

electrocardiogram (ECG) and electromyography (EMG). Firstly a new signal denoising approach based on

the Empirical mode decomposition (EMD) is presented. The EMD can decompose the noisy signal into

a number of Intrinsic Mode Functions (IMFs). The proposed algorithm estimates the noise level of each

IMF. Experiments show that the proposed EMD-based method provides better denoising results compared to

state-of-art. In addition, a real-time QRS detection approach is proposed to be directly applied on the noisy

ECG signals. Moreover, an adaptive thresholding approach is employed for the EMG segmentation. Both

approaches are validated using synthetic and real physiological data resulting in good performances.

1 INTRODUCTION

Affective Computing is a new area of computing

research described as ”computing which relates to,

arises from, or deliberately inﬂuences emotions” (Pi-

card, 2000). It emphasizes the importance of adding

new emotional features to the human-computer inter-

action. The use of physiological sensors as a means

of recognizing user’s affective state has a number of

advantages: (i) the size of such sensors is rapidly de-

creasing to the extent that it is nowadays incorporated

into body wireless network (Pantelopoulos and Bour-

bakis, 2008), (ii) they are less disturbing than being

”watched” by a camera as is the case with facial ex-

pression recognition, and (iii) they are less susceptible

to social masking (Kim, 2007).

However, emotional state recognition by means of

biosignals analysis is also problematic. This is due in

part to the fact that biosignals are usually corrupted by

white noise (

Ust

unda

g et al., 2012) and other various

types of noise, such as baseline wander, muscle con-

traction and electrode motion artifacts (Andrade et al.,

2006; Blanco-Velasco et al., 2008). Furthermore, de-

spite the evidence from psychophysiology suggesting

a strong correlation between human emotional states

and physiological responses, determining an appro-

priate mapping between them requires the estimation

of reliable features.

In this paper, we propose advanced signal process-

ing for the characterization of ECG and EMG phys-

iological signals. Firstly, a new signal denoising ap-

proach is introduced. The focus is on signals with

white Gaussian noise.

Wavelet-based denoising has been the dominant

technique in the area of non-linear and non-stationary

signals (e.g. physiological signals) denoising for

many years (Donoho, 1995). The goal is to estimate

the signal from the noisy observations such that the

Mean Square Error (MSE) is minimum. To achieve

this, the observed signal is transformed into wavelet

domain, which decomposes it into many subbands.

The small coefﬁcients in the subbands are dominated

by noise, while coefﬁcients with large absolute value

carry more signal information than noise. Replac-

ing noisy coefﬁcients (small coefﬁcients below cer-

tain value) by zero and an inverse wavelet transform

may lead to reconstruction signal that has lesser noise.

Often, hard thresholding and soft thresholding tech-

niques are used for such denoising process (Donoho,

1995).

The main drawback of the wavelet approach is

that the basis functions are ﬁxed, and do not nec-

essarily match signals with large variabilities, such

as the electromyography (EMG). Huang et al. pro-

Wu P., Jiang D. and Sahli H..

Physiological Signal Processing for Emotional Feature Extraction.

DOI: 10.5220/0004727500400047

In Proceedings of the International Conference on Physiological Computing Systems (PhyCS-2014), pages 40-47

ISBN: 978-989-758-006-2

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

posed the Empirical mode composition (EMD) as a

tool to adaptively decompose a signal into a num-

ber of components, called Intrinsic Mode Functions

(IMFs) (Huang et al., 1998). In contrast to the

wavelet approach, the EMD relies on a fully data-

driven mechanism that doesn’t require any predeﬁned

basis. A general scheme of signal denoising using

EMD is as follows: 1. Decompose the noisy signal

into a number of IMFs; 2. Estimate the noise level of

each IMF and threshold the estimated IMFs; 3. Re-

construct the denoised signal using the thresholded

IMFs.

EMD-based denoising has been applied to physio-

logical signals in many studies (Boudraa et al., 2005;

Andrade et al., 2006; Jing-tian et al., 2007; Blanco-

Velasco et al., 2008; Karagiannis and Constantinou,

2009; Agraﬁoti et al., 2012). However, the esti-

mation of the noise level, of each IMF, for a reli-

able threshold is still an open question. The strategy

of Donoho (Donoho and Johnstone, 1994) is widely

used (Boudraa et al., 2005; Jing-tian et al., 2007;

Karagiannis and Constantinou, 2009), nevertheless, it

is only suitable for noise dominant scales (i.e. IMFs),

thus applying this threshold to high scales causes sig-

nal distortion. This is due to the fact that each scale

of the EMD occupies lower frequencies than its pre-

ceding ones and most noise components lie in the ﬁrst

several scales. The high scales are signal dominated.

Using a single threshold may give misleading results.

State-of-art approaches either reconstruct denoised

signals using all thresholded IMFs while ignoring the

signal distortion (Boudraa et al., 2005), or by thresh-

olding the ﬁrst IMFs and keep the other IMFs un-

changed (Jing-tian et al., 2007; Blanco-Velasco et al.,

2008). For the latter, the number of IMFs to be thresh-

olded is decided empirically (Jing-tian et al., 2007) or

via statistical test (Blanco-Velasco et al., 2008). Alter-

natively, instead of estimating the noise level from the

whole IMF, Andrade (Andrade et al., 2006) manually

selected a window of noise from the original signal

and then the boundaries of this window were used for

the noise level estimation.

In this paper, we follow the idea of EMD-based

denoising, however different from state-of-art meth-

ods, rather than estimating the noise level of all IMFs

using Donoho’s strategy, we propose a novel ap-

proach for noise levels estimation of dominant IMFs.

In this study, we focus on the electrocardiogram

(ECG) and electromyography (EMG) signals which

have been conﬁrmed to be useful for emotion recog-

nition (Kreibig et al., 2007). For the ECG, heart rate

(HR) and heart rate variability (HRV) are the cardio-

vascular response features most often reported as in-

dicators of emotion (Kreibig, 2010). The ﬁrst step in

extracting them starts from the exact detection of R

peaks in the QRS complex (see Fig.1). Within the

last decade many approaches to R peaks detection

have been proposed (Kohler et al., 2002). However,

most of these approaches are off-line and targeting

the noiseless signal, which don’t meet the require-

ments of many real-time applications. To overcome

this problem, we make use of a Change Point Detec-

tion (CPD) algorithm proposed by Guralnik and Sri-

vastava (Guralnik and Srivastava, 1999) for event de-

tection in time series.

The EMG also plays an important role in expres-

sion recognition (Van Boxtel, 2010). A basic problem

in EMG applications is the determination of muscu-

lar active periods within a given EMG signal. Within

each muscle activity period the root mean square

(RMS) and absolute mean value (AMV), are used as

features (Farf

an et al., 2010). Most state-of-art ap-

proaches uses thresholding for muscle activity seg-

mentation. However, a ﬁxed threshold value may

cause misleading results (

Ozg

unen et al., 2010) as

muscular movement has different intensities. In this

study, inspired by the energy-based speech voice ac-

tivity detection (Van Gerven and Xie, 1997; Ghosh

et al., 2011), we propose an algorithm for the auto-

matic segmentation of EMG signals by dynamically

calculating instantaneous values for the estimation of

the segmentation threshold based on an adaptive scal-

ing parameter.

2 EMD-BASED SIGNAL

DENOISING

Signal denoising can be described as following: given

a noisy signal x(t):

x(t) = f (t) + n(t) (1)

where f (t) and n(t) are signal and noise components,

respectively. The objective is to estimate the noise

level σ(n) and then ﬁlter out the noise. Translating

this idea to the case of EMD-based denoising, ﬁrstly

the noisy signal is decomposed via EMD into a num-

ber of Intrinsic Mode Functions, c

, i.e.:

x(t) =

∑

i=1

(t) + r

(t) (2)

where N is the number of IMFs and r

is the residual.

The noise component, n(t), of the signal x(t) is now

decomposed and dispersed among the IMFs. Thus

we have to estimate the noise level σ

(n), where i =

1,...,N of each IMF.

In this study, we propose an approach for estimat-

ing σ

(n), which avoids removing useful information

PhysiologicalSignalProcessingforEmotionalFeatureExtraction

from the dominant IMFs, hence causing signal distor-

tion. The idea, is that only σ

(n) is computed using

Donoho’s approach (Donoho and Johnstone, 1994)

since the ﬁrst IMF basically consist of high frequency

noise component, i.e.:

(n) = MAD

/.6745 (3)

MAD

= Median(| c

(t) − Median(c

) |) (4)

Then the noise level of the ith IMF σ

(n) is computed

as:

(n) = δ

(n) (5)

The procedure for computing δ

is as following: at

ﬁrst the Hilbert transform (HT) is applied to each

IMF. The HT of the ith IMF c

(t) is deﬁned as (Huang

et al., 1999):

H[c

(t)] = c

∗

πt

= y

(6)

or using the convolution deﬁnition,

(t) =

∞

−∞

(γ)

t − γ

dγ (7)

Where P indicates the Cauchy principal value. From

(t) it is possible to deﬁne the analytical version of

the ith IMF:

(t) = c

(t) + iy(t) (8)

where i is the imaginary unit, or in polar form:

(t) = a

(t)e

iθ

(t)

(9)

in which:

(t) =

(t) + y

(t) (10)

(t) = arctan(

(t)

) (11)

The instantaneous frequency of the ith IMF ω

is then

obtained using the instantaneous variation of phase

(Huang et al., 1999).

(t) =

dθ(t)

(12)

the mean period

is determined using the mean value

of the instantaneous frequencies of the ith IMF:

mean(ω

(t))

(13)

Following, the noise energy density of the ith IMF, E

is needed. Since the ﬁrst IMF is always noise domi-

nated, E

can be easily obtained by:

∑

t=1

(t) (14)

The product of energy density of IMF and its corre-

sponding mean period must be a constant (Wu and

Huang, 2004), i.e.:

= C (15)

Based on the mean period of the ith IMF

and the

const C, the energy density of the ith IMF is obtained:

= C/

(16)

and then:

= E

(17)

Having the δ

, the noise level of the ith IMF σ

(n)

can be obtained using Eq.5. The soft-thresholding ap-

proach (Donoho, 1995) is the applied to the IMFs:

(t) =











(t) − T

, i f c

(t) ≥ T

0, i f | c

(t) |< T

(t) + T

, i f c

(t) ≤ −T

(18)

in which c

(t) is the thresholded version of the ith

IMF, where:

= σ

(n)

2log(L) (19)

Where L is the length of the signal x(t). Finally, the

denoised signal x

(t) is obtained via reconstruction

using the thresholded IMFs:

(t) =

∑

i=1

(t) + r

(t) (20)

3 Physiological Signal Analysis

3.1 Real-time QRS detection in ECG

The QRS complex is the most important segment

in ECG signal, which reﬂects the electrical activity

within the heart during the ventricular contraction.

The moment of its occurrence gives us much infor-

mation related to the current emotional state (Kreibig,

2010). Thus the detection of QRS complex, in par-

ticular R peak detection is the basis for ECG-based

emotion recognition.

The real-time QRS detection approach proposed

in this paper is based on the algorithm of change

point detection of (Guralnik and Srivastava, 1999).

Change point detection approaches apply data min-

ing techniques to identify the time points at which the

changes, i.e. events, occur. In (Guralnik and Srivas-

tava, 1999) a method has been proposed for the de-

tection of the appropriate set of number of points that

minimizes the error in ﬁtting a pre-deﬁned function

using maximum likelihood. There is no ﬁxed num-

ber of change-points to be detected. Moreover, no

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

constraints are imposed on the class of functions that

will be ﬁtted to the subsequences between successive

change-points.

Following the notation in (Guralnik and Srivas-

tava, 1999), let y(t), (t = 1,...,n) be the time series

to be segmented. It is assumed that the time series

can be modeled mathematically, where each model is

characterized by a set of parameters. The problem

of change-points detection, is formulated as ﬁnding a

piecewise segmented model, given by

Y = f

(t,v

) + e

(t),(1 < t ≤ θ

= f

(t,v

) + e

(t),(θ

< t ≤ θ

= .....................................

= f

(t,v

) + e

(t),(θ

l−1

< t ≤ n). (21)

Where f

(t,v

) is the function (with its vector of pa-

rameters v

) that is ﬁtted to the segment i. The θ

’s are

the change-points between successive segments, and

(t)’s are error terms. The change-point in the ﬁrst

stage is the j minimizing L (1, j) + L ( j + 1,n), say

∗

The approach of (Guralnik and Srivastava, 1999)

can be implemented in a batch or incremental. The

batch algorithm is useful only when data collection

precedes analysis. As in our case we are targeting

on-line processing, we use the incremental version

of the algorithm. The key idea is that if the next

data point collected by the sensor reﬂects a signiﬁ-

cant change in phenomenon, then the likelihood of

being a change-point is going to be smaller then the

likelihood that it is not. However, if the difference in

likelihoods is small, we cannot deﬁnitively conclude

that a change did occur, since it may be the artifact of

a large amount of noise in the data. Therefore a user-

deﬁned likelihood increase threshold is introduced.

no change

− L

change

)/L

no change

> δ, (22)

where δ is a user-deﬁned likelihood increase thresh-

old.

Suppose that the last change-point was detected at

time t

l−1

. At time t

the algorithm starts by collecting

enough data to ﬁt the regression model. Suppose at

time t

a new data point is collected. The candidate

change-point is found by determining t

, with likeli-

hood criterion L

min

(l, j), such that

min

(l, j) = min

l<i≤ j

L (l,i) + L (i + 1, j). (23)

If this minimum is signiﬁcantly smaller than

L (l, j), i.e. the likelihood criteria of no change-points

from t

to t

, then t

is a change-point. Otherwise, the

process should continue with the next point, i.e. t

j+1

In the incremental algorithm, execution time is a

signiﬁcant factor. If enough information is stored,

some of the calculations can be avoided. Thus, at time

j+1

to ﬁnd likelihood criteria

min

(l, j + 1) = min

l<i≤ j

L (l,i) + L (i + 1, j +1) (24)

it is only necessary to calculate L (i + 1, j + 1), since

(l,i) was calculated in the previous iteration.

Several types of basis functions, f

(t,v

), can

be considered, e.g. algebraic polynomials, wavelet,

Fourier, etc.. In our current implementation polyno-

mial ﬁtting functions of degree 1 have been selected

empirically. The user-deﬁned likelihood threshold δ,

has been set 0.97 in this study to avoid ﬁnding a lot

of irrelevant change points. The remaining irrelevant

change points (see Fig.1) can be ﬁltered out by a pre-

determined threshold. The threshold is selected as

15% of the maximum value of the detected change

points (Behbahani and Dabanloo, 2011).

Figure 1: Detected change points including R peaks (cross)

and irrelevant ones (circle)

3.2 EMG Segmentation

EMG signals are characterized by a discontinuous

signal since information is carried only when mus-

cles are active, such segments are the regions where

activity information exists and are referred to as ’ac-

tive segments’. The pauses between them are called

’inactive segments’. The features extracted from ac-

tive segments are considered more suitable for ex-

pression recognition (Hamedi et al., 2011). For ex-

ample, the active features of facial EMG is helpful

for not only recognizing facial expressions, but also

learning the actual mechanism of how facial expres-

sions are formed (Aoi et al., 2011), which is less

susceptible to social masking and may better reveal

the emotional state behind the facial expression. The

decision of determining to what class an EMG seg-

ment belongs reassembles the voice activity detec-

tion (VAD) in speech. Inspired by the energy-based

speech VAD (Van Gerven and Xie, 1997; Ghosh et al.,

PhysiologicalSignalProcessingforEmotionalFeatureExtraction

2011) approaches, in this paper we propose an RMS-

based EMG segmentation. RMSs are extracted from

each L = 100 ms of the EMG signal as an indicator of

the muscle activation:

RMS(i) =

∑

t=1

x(t)

(25)

where i is the index of the windows within EMG sig-

nal x(t). The duration of the optimal analysis window

depends on the purpose of the study. The window

size of 100 ms is decided in this study to track the

fast dynamic changes in facial expression (Van Box-

tel, 2010).

The adaptive thresholding method for EMG seg-

mentation is based on RMS levels, i.e., RMS

min

and

RMS

max

, representing the minimum and maximum

RMS value of the incoming windows respectively.

The threshold T is calculated as (Ghosh et al., 2011):

T = αRMS

min

+ (1 − α)RMS

max

(26)

where α is an adaptive factor and computed

as (Sakhnov et al., 2009):

α =

RMS

max

− RMS

min

RMS

max

(27)

RMS

min

and RMS

max

are initialized by the ﬁrst second

of the EMG signal, i.e., the ﬁrst 10 windows. The

RMS

min

, RMS

max

and corresponding T are then up-

dated for every incoming window before threshold-

ing. A control factor, called inactive counter (IC), is

deﬁned as the number of consecutive windows with

RMS value which is smaller than T . If the RMS value

of the incoming window i, i.e. RMS(i) is larger than

T , it is classiﬁed as active; If RMS(i) is smaller than

T , and IC > 3, the window i is classiﬁed as inactive,

otherwise it’s still labeled as active.

4 RESULTS AND DISCUSSION

4.1 Signal Denoising

In order to test the validity of the proposed denoising

approach, three ECG (101, 103, 118) and two EMG

(brux1, brux2) noiseless signals are chosen arbitrarily

from the MIT-BIH arrhythmia database (Moody and

Mark, 2001) and the CAP sleep database (Terzano

et al., 2001) respectively through PhysioNet (Gold-

berger et al., 2000). Then simulations for several dif-

ferent cases are carried out. A white Gaussian noise

n(t) is added to the noiseless signal f (t), obtaining

the synthetic noisy signal x

∗

(t) (such as 101

∗

, 103

∗

118

∗

, brux1

∗

, brux2

∗

). The quantitative evaluation is

assessed by the signal-to-error ratio (SER) (Blanco-

Velasco et al., 2008)

SER =

∑

L−1

t=0

(t)

∑

L−1

t=0

[ f (t) − x

(t)]

(28)

where x

(t) is the denoised version of x

∗

(t). The

signal-to-noise ratio (SNR) is given by:

SNR =

∑

L−1

t=0

(t)

∑

L−1

t=0

(t)

(29)

Fig.2 and Fig.3 illustrate the waveforms of the

original (clean) signal, noisy signal, and denoised sig-

nal using the proposed EMD-based approach, the tra-

ditional EMD-based approach and the wavelet-based

approach, for the ECG and the EMG respectively.

Here, the traditional EMD-based approach (Jing-tian

et al., 2007) and the wavelet-based approach (Tikka-

nen, 1999) are implemented for comparison. With

respect to the latter, the 4-level discrete wavelet

transform with the Daubechies (db4) and Symlets

(sym2) basis are used for the ECG and the EMG sig-

nals respectively. The soft-thresholding proposed by

Donoho (Donoho, 1995) is then applied to the dis-

crete wavelet transform coefﬁcients with ﬁxed thresh-

olds for each level. As can be seen from Fig.2 and

Fig.3, the proposed EMD-based approach yields good

results in terms of visual quality and outperforms

the traditional EMD-based approach and the wavelet-

based approach for both the ECG and the EMG cases.

Figure 2: Example of ECG denoising for white Gaussian

noise

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

Table 1: Signal denoising experiments carried out over several records.

SNR = 2 db SNR = 6 db SNR = 10 db

Record EMD tEMD WT EMD tEMD WT EMD tEMD WT

101

∗

7.25 5.48 5.94 16.01 10.84 10.42 21.73 15.59 13.34

103

∗

6.80 4.68 5.83 15.92 10.86 11.30 19.80 18.05 15.40

118

∗

8.71 7.41 6.27 13.6 11.32 12.25 23.05 16.30 14.09

brux1

∗

8.25 2.41 3.26 18.58 2.65 4.08 28.38 3.19 4.25

brux2

∗

6.05 2.05 1.53 11.00 2.95 1.72 15.75 3.83 1.80

Figure 3: Example of EMG (the submentalis muscle) de-

noising for white Gaussian noise

Moreover, we study the behavior of the proposed

method quantitatively, using multiple realizations of

white Gaussian noise at different SNRs (2 db, 6 db

and 10 db). The results are presented in Table.1 in

terms of SER for the corresponding methods (EMD:

proposed EMD-based, tEMD: traditional EMD-based

and WT: wavelet-based). Again, as can be seen from

Table.1, the proposed EMD-based method shows bet-

ter ability to deal with the white Gaussian noise than

the other two methods.

4.2 Real-time QRS Detection

We use both the synthetic and real ECG data to test

the validity of proposed real-time QRS detection ap-

proach. The real ECG data were acquired for the

analysis of emotional facial expressivity of Parkin-

son patients (Verbraeck, 2012). In the experiments

for data acquisition, six emotions (happiness, sadness,

anger, disgust, surprise, and fear) and a neutral state

Table 2: The results of the QRS detection.

Record No. of Beats FP FN Err(%)

101

∗

1865 1 13 0.75

103

∗

2084 1 15 0.77

118

∗

2278 6 14 0.88

4431 5 39 0.99

4781 3 20 0.48

5288 2 9 0.21

5169 6 21 0.52

3924 12 93 2.68

4220 10 79 2.11

were induced via watching movies clips in a group

of eight healthy subjects and seven persons suffering

from Parkinson’s disease. The EMG from two facial

muscles (the levator labii superioris, and the orbicu-

laris oculi), the ECG, as well as the face of each sub-

ject were recorded (Verbraeck, 2012). In total, the

data used for validation in this study include 3 syn-

thetic signals (101

∗

, 103

∗

, 118

∗

), 4 real signals from

subjects with normal heartbeats (N1, N2, N3, N4) and

2 real signals from subjects with arrhythmia (A

, A

All the data are noisy signals.

In Table.2, false positive (FP), indicates the

method detects a beat when no beat is present; while

FN, false negative, indicates that the method failed to

detect a beat. The Err is the total detection failure

rate, deﬁned as:

Err =

∑

FP +

∑

× 100% (30)

where S is the total number of beats. Consequently,

based on the obtained results, the proposed QRS de-

tection approach is effective on noisy ECG signals.

4.3 EMG Segmentation

The proposed EMG segmentation approach is ap-

plied to both the synthetic data (brux1

∗

, brux2

∗

) and

the real EMG signals acquired in (Verbraeck, 2012).

Fig.4 shows the example of segmentation for an EMG

signal from the levator labii superioris using the pro-

posed approach. As can be seen, the validity of the

PhysiologicalSignalProcessingforEmotionalFeatureExtraction

Figure 4: Example of EMG (the levator labii superioris muscle) segmentation on real data.

proposed approach has been conﬁrmed: the EMG sig-

nals are segmented into several active segments be-

tween the green and the following red lines. Even

short and weak active segments have been detected.

5 CONCLUSIONS

This paper introduces novel approaches of physiolog-

ical signal processing for emotional feature extrac-

tion. Both the ECG and the EMG signals are ad-

dressed. Firstly an EMD-based denoising approach is

proposed where the noise levels of all IMFs are effec-

tively estimated and processed to successfully achieve

the denoising. The validity of the proposed method is

conﬁrmed through several experiments. Results indi-

cate that it is an effective tool for physiological sig-

nals denoising, especially for the case of white Gaus-

sian noise. In addition, a real-time QRS detection ap-

proach is proposed which can be applied directly on

the noisy ECG signals. Finally, an adaptive thresh-

olding approach is employed for EMG segmentation.

Both approaches for QRS detection and EMG seg-

mentation are validated using synthetic and real phys-

iological data resulting in good performances.

ACKNOWLEDGEMENTS

The research work reported in this paper is supported

by a CSC-VUB scholarship grant, and the VUB-

interdisiplinary research project ”Objectifying assess-

ment of human emotional processing and expression

for clinical and psychotherapy applications (EMO-

App)”.

REFERENCES

Agraﬁoti, F., Hatzinakos, D., and Anderson, A. K. (2012).

Ecg pattern analysis for emotion detection. IEEE

Transactions on Affective Computing, 3:102–115.

Andrade, A. O., Nasuto, S., Kyberd, P., Sweeney-Reed,

C. M., and Kanijn, F. R. V. (2006). Emg signal ﬁlter-

ing based on empirical mode decomposition. Biomed-

ical Signal Processing and Control, 1:44–55.

Aoi, M., Kamijo, M., and Yoshida, H. (2011). Relationship

between Facial Expression and Facial Electromyo-

gram (f-EMG) Analysis in the Expression of Drowsi-

ness. In International Conference on Biometrics and

Kansei Engineering, pages 65–70.

Behbahani, S. and Dabanloo, N. (2011). Detection of

qrs complexes in the ecg signal using multiresolution

wavelet and thresholding method. In Computing in

Cardiology, 2011, pages 805–808.

Blanco-Velasco, M., Weng, B., and Barner, K. E. (2008).

Ecg signal denoising and baseline wander correction

based on the empirical mode decomposition. Comput-

ers in Biology and Medicine, 38:1–13.

Boudraa, A. O., Cexus, J. C., and Saidi, Z. (2005). Emd-

based signal noise reduction. Signal Processing, 1:33–

37.

Donoho, D. L. (1995). Denoising by soft-thresholding.

IEEE Transactions on Information Theory, 41:613–

627.

Donoho, D. L. and Johnstone, I. M. (1994). Ideal spatial

adaption via wavelet shrinkage. Biometrika, 81:425–

455.

Farf

an, F., Politti, J., and Felice, C. (2010). Evaluation of

PhyCS2014-InternationalConferenceonPhysiologicalComputingSystems

emg processing techniques using information theory.

Biomedical engineering online, 9(1):72.

Ghosh, P. K., Tsiartas, A., and Narayanan, S. (2011). Ro-

bust voice activity detection using long-term signal

variability. Audio, Speech, and Language Processing,

IEEE Transactions on.

Goldberger, A. L., Amaral, L. A., Glass, L., Hausdorff,

J. M., Ivanov, P. C., Mark, R. G., Mietus, J. E., Moody,

G. B., Peng, C.-K., and Stanley, H. E. (2000). Phys-

iobank, physiotoolkit, and physionet components of

a new research resource for complex physiologic sig-

nals. Circulation, 101(23):e215–e220.

Guralnik, V. and Srivastava, J. (1999). Event detection from

time series data. In Knowledge Discovery and Data

Mining, pages 33–42.

Hamedi, M., Salleh, S.-H., and Swee, T. T. (2011). Sur-

face electromyography-based facial expression recog-

nition in bi-polar conﬁguration. Journal of Computer

Science, 7(9):1407–1415.

Huang, N. E., Shen, Z., and Long, S. R. (1999). A new

view of nonlinear water waves: The hilbert spectrum

1. Annual review of ﬂuid mechanics, 31(1):417–457.

Huang, N. E., Shen, Z., Long, S. R., Wu, M. C., Shih,

H. H., Zheng, Q., Yen, N.-C., Tung, C. C., and Liu,

H. H. (1998). The empirical mode decomposition and

the hilbert spectrum for nonlinear and non-stationary

time series analysis. Proceedings of The Royal Society

A: Mathematical, Physical and Engineering Sciences,

454:903–995.

Jing-tian, T., Qing, Z., Yan, T., Bin, L., and Xiao-kai, Z.

(2007). Hilbert-huang transform for ecg de-noising.

In Bioinformatics and Biomedical Engineering, 2007.

ICBBE 2007. The 1st International Conference on,

pages 664–667.

Karagiannis, A. and Constantinou, P. (2009). Noise com-

ponents identiﬁcation in biomedical signals based on

empirical mode decomposition. In IEEE EMBS Inter-

national Conference on Information Technology Ap-

plications in Biomedicine, pages 1–4.

Kim, J. (2007). Bimodal emotion recognition using speech

and physiological changes. Robust speech recognition

and understanding, pages 265–280.

Kohler, B.-U., Hennig, C., and Orglmeister, R. (2002). The

principles of software qrs detection. IEEE Engineer-

ing in Medicine and Biology Magazine, 21:42–57.

Kreibig, S. D. (2010). Autonomic nervous system activ-

ity in emotion: A review. Biological psychology,

84(3):394–421.

Kreibig, S. D., Wilhelm, F. H., Roth, W. T., and Gross, J. J.

(2007). Cardiovascular, electrodermal, and respiratory

response patterns to fear-and sadness-inducing ﬁlms.

Psychophysiology, 44(5):787–806.

Moody, G. B. and Mark, R. G. (2001). The impact of the

mit-bih arrhythmia database. Engineering in Medicine

and Biology Magazine, IEEE, 20(3):45–50.

Ozg

unen, K. T., C¸ elik, U., and Kurdak, S. S. (2010). De-

termination of an optimal threshold value for muscle

activity detection in emg analysis. Journal of Sports

Science and Medicine, 9:620–628.

Pantelopoulos, A. and Bourbakis, N. (2008). A survey on

wearable biosensor systems for health monitoring. In

Engineering in Medicine and Biology Society, 2008.

EMBS 2008. 30th Annual International Conference of

the IEEE, pages 4887–4890.

Picard, R. W. (2000). Affective computing. MIT press.

Sakhnov, K., Verteletskaya, E., and Simak, B. (2009). Ap-

proach for energy-based voice detector with adaptive

scaling factor. IAENG Internat. J. Comput. Sci, 36(4).

Terzano, M. G., Parrino, L., Sherieri, A., Chervin, R.,

Chokroverty, S., Guilleminault, C., Hirshkowitz, M.,

Mahowald, M., Moldofsky, H., Rosa, A., et al. (2001).

Atlas, rules, and recording techniques for the scor-

ing of cyclic alternating pattern (cap) in human sleep.

Sleep medicine, 2(6):537–553.

Tikkanen, P. E. (1999). Nonlinear wavelet and wavelet

packet denoising of electrocardiogram signal. Biolog-

ical Cybernetics, 80(4):259–267.

Ust

unda

g, M., G

okbulut, M., S¸eng

ur, A., and Ata, F.

(2012). Denoising of weak ecg signals by using

wavelet analysis and fuzzy thresholding. Network

Modeling Analysis in Health Informatics and Bioin-

formatics, 1(4):135–140.

Van Boxtel, A. (2010). Facial emg as a tool for inferring af-

fective states. In Proceedings of Measuring Behavior

2010, pages 104–108.

Van Gerven, S. and Xie, F. (1997). A comparative study of

speech detection methods. In Eurospeech, volume 97.

Verbraeck, F. (2012). Objectifying human facial expres-

sions for clinical applications. Master’s thesis, Vrije

Universiteit Brussel, Belgium.

Wu, Z. and Huang, N. E. (2004). A study of the character-

istics of white noise using the empirical mode decom-

position method. Proceedings of The Royal Society

A: Mathematical, Physical and Engineering Sciences,

460:1597–1611.

PhysiologicalSignalProcessingforEmotionalFeatureExtraction