COMPUTER AIDED DIAGNOSIS FOR MENTAL HEALTH CARE
On the Clinical Validation of Sensitive Machines
Frans van der Sluis
1,2
, Ton Dijkstra
3
and Egon L. van den Broek
1,2,4
1
Human-Media Interaction (HMI), University of Twente, P.O. Box 217, 7500 AE Enschede, The Netherlands
2
Karakter University Center, Radboud University Medical Center Nijmegen,
P.O. Box 9101, 6500 HB Nijmegen, The Netherlands
3
Donders Institute for Brain, Cognition, and Behaviour, Radboud University,
P.O. Box 9104, 6500 HE Nijmegen, The Netherlands
4
TNO Technical Sciences, P.O. Box 5050, 2600 GB Delft, The Netherlands
Keywords:
Stress, Mental health care, Speech, Computer aided diagnostics (CAD), Artificial neural network, Validation.
Abstract:
This study explores the feasibility of sensitive machines; that is, machines with empathic abilities, at least
to some extent. A signal processing and machine learning pipeline is presented that is used to analyze data
from two studies in which 25 Post-Traumatic Stress Disorder (PTSD) patients participated. The feasibility
of speech as a stress detector was validated in a clinical setting, using the Subjective Unit of Distress (SUD).
13 statistical parameters were derived from five speech features, namely: amplitude, zero crossings, power,
high-frequency power, and pitch. To achieve a low dimensional representation, a subset of 28 parameters was
selected and, subsequently, compressed into 11 principal components (PC). Using a Multi-Layer Perceptron
neural network (MLP), the set of 11 PC were mapped upon 9 distinct quantizations of the SUD. The MLP
was able to discriminate between 2 stress levels with 82.4% accuracy and up to 10 stress levels with 36.3%
accuracy. With stress baptized as being the black death of the 21st century, this work can be conceived as an
important step towards computer aided mental health care.
1 INTRODUCTION
In contrast to animals, humans have the ability to
make cognitive representations of events, both from
the past and for the future. Although such repre-
sentations aid our daily work and living, they have
their down side. In the case of stressful life events,
cognitive representations can catalyze worrying and,
hence, facilitate chronic stress, unknown to animal
species (Brosschot, 2010). This effect is strength-
ened by the influence of stress on our memory sys-
tems (Schwabe et al., 2010). Chronic stress often pro-
duces similar physiological responses to those occur-
ring during the stressful events from which it orig-
inates. In turn, this can cause pervasive and struc-
tural chemical imbalances in our physiological sys-
tems, including the autonomic nervous, central ner-
vous, neuroendocrine, immune system, and even in
the brain (Brosschot, 2010). Due to the complexity
of our physiological systems, their continuous inter-
action, and their inherent dynamic nature, a thorough
understanding of ‘chronic stress’ is still missing.
Current day practice in treatment of chronic stress
focusses on the treatment of either cognitiverepresen-
tations, our habit memory system, or both (Schwabe
et al., 2010). In general, under stressful events, the
habit memory system tends to dominate over the cog-
nitive memory (or representations) system; however,
their precise relation remains unknown (Schwabe
et al., 2010). This lack of understanding makes treat-
ment inherently complex and requires a very high
level of expertise from the clinician. Moreover, most
indicators of patients’ progress rely on behavior mea-
sures and the clinician’s expertise.
This article presents the development of the back-
end of a computer aided diagnosis (CAD) for mental
health care, in particular for the treatment of chronic
stress related disorders. This backend will be val-
idated with two previously gathered sets of clinical
data (van den Broek et al., 2009; van den Broek et al.,
2011; van der Sluis et al., 2010; van der Sluis et al.,
2011). Its foundation lay in speech signal analysis
and the processing of the signal’s features by a Multi-
Layer Perceptron neural network (MLP). The envi-
493
van der Sluis F., Dijkstra T. and L. van den Broek E..
COMPUTER AIDED DIAGNOSIS FOR MENTAL HEALTH CARE - On the Clinical Validation of Sensitive Machines.
DOI: 10.5220/0003891404930498
In Proceedings of the International Conference on Health Informatics (BSSS-2012), pages 493-498
ISBN: 978-989-8425-88-1
Copyright
c
 2012 SCITEPRESS (Science and Technology Publications, Lda.)

sioned CAD system can be used as a decision support
system in everyday life (e.g., at work) and in mental
health care settings.
The remaining article is composed as follows.
Next, we will present a brief section that will describe
the data set with which the CAD framework will be
clinically validated. Section 3 will describe relevant
speech signal features and their parameters. Section 4
will present the results of the actual stress classifica-
tion and validation employed in this study. Last, Sec-
tion 5 will present a general discussion, with which
this article end.
2 DATA SET
Various stimuli have been applied in the endeavor to
trigger stress in a controlled manner, including, im-
ages, sounds, (fragments of) films (van den Broek and
Westerink, 2009), and real-world experiences (Healey
and Picard, 2005). However, how do we know which
methods actually triggered participants’ true stress?
This is a typical concern of validity, which is a crucial
issue for stress assessment.
Although understandable from a measurement-
feasibility perspective, stress measurements are of-
ten done in controlled lab settings (cf. the Trier So-
cial Stress Test (Kirschbaum et al., 1993)). This
makes results poorly generalizable to real-world ap-
plications (van den Broek, 2010). Moreover, under
normal circumstances, in our every day lives, bursts
of significant stress are sparse, which makes it even
more difficult to obtain such data (in a limited time
frame) (Picard, 2010). However, luckily, previously
we already obtained two data sets of clinically vali-
dated data that comprises bursts of authentic chronic
stress (van den Broek et al., 2009; van den Broek
et al., 2011; van der Sluis et al., 2010; van der Sluis
et al., 2011), which we will briefly introduced here.
In total, 25 female Dutch Post-Traumatic Stress
Disorder (PTSD) patients (mean age: 36; SD: 11.32)
participated of their own free will. All patients suf-
fered from panic attacks, agoraphobia, and panic dis-
order with agoraphobia (S´anchez-Meca et al., 2010).
Before the start of the studies, all patients signed an
informed consent form and all were informed of the
tasks they could expect. The data from one patient
with problems in both studies were omitted from fur-
ther analysis. Hence, the data of 24 patients were used
for further analysis.
All patients took part in two studies: a storytelling
(ST) study and a reliving (RL) study. Possible fac-
tors of influence (e.g., location, apparatus, therapist,
and experiment leader) were kept constant. In the
ST study, the participants read aloud both a stress-
provoking and a positive story; see also (van den
Broek et al., 2011; van der Sluis et al., 2010; van der
Sluis et al., 2011). This procedure allows consider-
able methodological control over the invoked stress,
in the sense that every patient reads exactly the same
stories. The fictive stories were constructed in such a
way that they would induce certain relevant emotional
associations. The complexity and syntactic structure
of the two stories were controlled to exclude the ef-
fects of confounding factors. In the RL study, the
participants re-experienced their last panic attack and
their last joyful occasion; see also (van den Broek
et al., 2009; van den Broek et al., 2011). For RL
, a panic attack approximates the trauma in its full
strength, as with the during admission of a patient.
The condition of telling about the last experienced
happy event resembles that of a patient who is relaxed
or (at least) in a ‘normal’ stress condition.
To evaluate the quality of our speech analysis, we
compared the results of our speech analysis to those
obtained by means of a standard questionnaire: the
Subjective Unit of Distress (SUD) (Wolpe, 1958).
The SUD has repeatedly proven itself as a reliable
measure of a person’s experienced stress. The SUD
will serve as the ground truth in our quest to develop a
CAD system for mental health care. The CAD should
be robust, enable to process a variety of data. For
this purpose, we have chosen to treat both the speech
and the SUD data of both studies as one set. Con-
sequently, the assessment of their relation was put to
the test by such a diverse data set and enabled the il-
lustration of the robustness of this relation. For more
detailed analyses, we refer to (van den Broek et al.,
2012).
The SUD is measured by a Likert scale that reg-
isters the degree of distress a person experiences at a
particular moment in time. In our case, we used a lin-
ear scale with a range between 0 and 10 on which the
experienced degree of distress was indicated by a dot
or cross. The participants in our study were asked to
fill in the SUD test once every minute; consequently,
it became routine during the experimental sessions.
3 SPEECH SIGNAL PROCESSING
Speech was recorded using a personal computer, an
amplifier, and a microphone. The sample rate of the
recordings was 44.1 kHz, mono channel, with a res-
olution of 16 bits. All recordings were divided into
epochs of approximately one minute of speech. For
each of the two conditions of both experiments, 3
epochs of one minute of speech were taken. Because
HEALTHINF 2012 - International Conference on Health Informatics
494

the therapy sessions were held under controlled con-
ditions in a room shielded from noise, high quality
speech signals were collected.
This section will describe the features that were
extracted from the speech signal, using the Praat soft-
ware package
1
. Subsequently, the basic preprocess-
ing conducted on the speech signal will be described,
namely: outlier removal, data normalization, and pa-
rameter derivation from the complete set of features.
3.1 Speech Signal Features
From the speech signals, various features have been
shown to be sensitive to experienced stress; for a re-
cent survey, see (El Ayadi et al., 2011). In this re-
search, we extracted five features from the speech sig-
nal, namely: i) its wave amplitude (Scherer, 2003); ii)
power, often used interchangeablywith energyand in-
tensity, which is also described as the Sound Pressure
Level (SPL) , relativeto the auditory threshold P
0
(i.e.,
in decibel (dB) (SPL )) (Cowie et al., 2001); iii)the
zero-crossings rate of the speech signal (Yang and
Lugger, 2010); iv) the high-frequency power (Banse
and Scherer, 1996; Cowie et al., 2001; Yang and Lug-
ger, 2010): the power for the domain [1000, 22000],
denoted in Hz; and v) the fundamental frequency (F0
or perceived pitch) (Cowie et al., 2001; Scherer, 2003;
Yang and Lugger, 2010), extracted using an autocor-
relation function (i.e., the cross-correlation of the sig-
nal with itself).
13 statistical parameters were derived from the 5
speech signal features: mean, median, standard de-
viation, variance, minimum value, maximum value,
range, the quantiles at 10%, 90%, 25%, and 75%,
the inter-quantile-range 10% − 90%, and the inter-
quantile-range 25% − 75%. The features and statis-
tical parameters were computed for each minute of
speech sample over a time window of 40 ms, us-
ing a step length of 10 ms (i.e., computing each fea-
ture every 10 ms over the next 40 ms of the sig-
nal). This short term processing approach takes care
of time varying spectral information and is in line
with the generally accepted standards (Rossing et al.,
2007). Two variations of amplitude are reported, one
in which the parameters are calculated from the mean
amplitude per window of 40 ms, and one where the
features are calculated over the full signal (reported
as amplitude(full)). In total, 6 × 13) = 78 parameters
were determined on the basis of the speech signal fea-
tures.
The same procedure for outlier removal was exe-
cuted for all speech features. It was founded on the
1
http://www.praat.org by P. P. G. Boersma and D. J. M.
Weenink [Last accessed on December 09, 2011]
inter-quartile range iqR, which we define as: q
3
− q
1
,
with q
1
being the 25th percentile and q
3
being the
75th percentile. Next, data points x were removed
from the data set if q
1
− 3iqR ≥ x ≥ q
3
+ 3iqR. All
other data points (i.e., that satisfied this requirement)
were kept in the data set.
To achieve good classification results with pat-
tern recognition and machine learning methods, the
set of selected input features is crucial. The same
holds for classifying stress. To limit this enormous
feature space, a Linear Regression Model (LRM) -
based heuristic search was applied, using α ≤ 0.1,
which can be considered as a soft threshold. An LRM
was generated using all available data, starting with
the full set of parameters, and then reducing it in 32
iterations by means of backward removal, to a set
of 28 parameters. The LRM model explained 59%
(F(28,351) = 18.22, p < .001) of the variance.
4 CLASSIFICATION AND
VALIDATION
A principal component analysis (PCA) was used
to (further) reduce the dimensionality of the set of
speech signal parameters, while preserving its vari-
ation as much as possible. 28 parameters were fed
into the PCA transformation. Subsequently, the first
11 principal components from the PCA transforma-
tion were selected, covering 95% of the variance as
was explained by the fill set of 28 parameters. These
principal componentsprovided a condensed represen-
tation of the LRM and, as such, served as input for the
MLP that will be introduced next.
The MLP has been used as state-of-the-art ma-
chine learning technique. It are universal approxima-
tors, capable of modeling complex functions. More-
over, MLPs can adequately handle irrelevant inputs
and noise and can adapt their weights and/or topology
in response to environmental changes. They are used
for classification, providing discrete outputs, but also
for regression with numeric outputs and reinforce-
ment learning when output is not perfectly known.
For a proper introduction to this classifier, we re-
fer to the many handbooks and survey articles that
have been published. Here, we will only specify
the MLP, for purpose of replication. We computed
WEKA’s (Hall et al., 2009) MLP trained by a back-
propagation algorithm (in its binary mode). It used
gradient descent with moment and adaptive training
parameters. In our case, a MLP with 3 layers with 7
nodes in the hidden layer was shown to have optimal
topology. This topology was trained with 500 cycles.
The nodes in this network were all sigmoid. For all
COMPUTER AIDED DIAGNOSIS FOR MENTAL HEALTH CARE - On the Clinical Validation of Sensitive Machines
495

other settings, the defaults of WEKA were used (Hall
et al., 2009).
We conducted a cross-validation of the (precision
of the) SUD with the parameters of the speech sig-
nal features that are classified by the MLP. On the
one hand, this verifies the validity of the SUD ; on
the other hand, this determines the performance of the
classifier in objective stress detection. The MLP was
tested using 10-fold cross-validation. Its average per-
formance is reported in Table 1.
4.1 Validation
The SUD ranged from 0 to 10, giving 11 classes to
classify. However, none of the patients used SUD
score 10; hence, this class was omitted from the clas-
sification process. We classified the SUD scores over
both studies, including both conditions and their base-
lines; see also Section 2.
The SUD is an established instrument in psychol-
ogy; nevertheless, to the authors’ knowledge the in-
strument’s precision has not been assessed. People’s
interoception is said to be unreliable (Craig, 2002),
which calls for an assessment of the reliability of
a SUD with a relatively high precision (i.e., range:
0 − 10). However, although interoception is possi-
bly indeed prone to errors, it has been reported that
patients with anxiety disorders are (over)sensitive to
interoception (Domschke et al., 2010). Please note
that this can possibly be explained by the influence
chronic stress has on human memory (Schwabe et al.,
2010)
We used the SUD as a ground truth. To assess
the precision of the SUD , the scale was quantized
into all possible numbers of levels (i.e., from 2 to 10);
see Table 1. This quantization is performed by reas-
Table 1: The classification results (in %) of the Multi-
Layer Perceptron neural network (MLP). Correct classifi-
cation (C
N
), chance level for classification (µ
N
), delta clas-
sification rate (dC
N
; see Eq. 1) and relative classification
rate (rC
N
; see Eq. 2) are reported. The Subjective Unit of
Distress (SUD) was taken as ground truth. N indicates the
number of SUD levels employed.
N C
N
µ
N
dC
N
rC
N
2 82.4 % 50.0 % 32.4 % 64.7 %
3 72.4 % 33.3 % 39.0 % 117.1 %
4 57.4 % 25.0 % 32.4 % 129.5 %
5 49.0 % 20.0 % 29.0 % 144.7 %
6 47.6 % 16.7 % 31.0 % 185.8 %
7 42.4 % 14.3 % 28.1 % 196.6 %
8 41.6 % 12.5 % 29.1 % 232.6 %
9 34.7 % 11.1 % 23.6 % 212.6 %
10 36.3 % 10.0 % 26.3 % 263.2 %
signing the SUD responses into N steps, with a step
size of r/N, where r is the range of the SUD values
(i.e., 0− 9). This quantization allowed the assessment
of the reliability of the SUD in relation to the LRM.
Moreover, to provide a fair report on the MLP’s clas-
sification, we provide both the correct classification
rate (C
N
), delta classification rate (dC
N
), and the rel-
ative classification rate (rC
N
) for each of the N bins.
dC
N
is a standard correction, also known as delta or
reaction score/classification, which is defined as
C
N
− µ
N
, (1)
with µ
N
being the chance level for N classes. rC
N
expresses the improvement of the classification com-
pared to chance level, which is defined as
C
N
− µ
N
µ
N
× 100, (2)
rC
N
is also known as a range correction and used
more often in health research (Fillingim et al., 1992).
The relative classification rate (see Eq. 2) enables
the assessment of the true classification performance
on each level of quantization of the SUD . Figure 1
shows that the MLP has an almost monotone linear
increase in relative classification rate. Its linear fit fol-
lows the data in Table 1 nicely (explained variance:
R
2
= .95). Moreover, Figure 1 shows that the classi-
fication rate of the MLP is almost constant, indepen-
dent of the level of quantization of the SUD . These
fits underline the validity of the SUD as an instrument
to assess people’s stress levels. The fits also illus-
trate the SUD ’s high concurrent validity, with its abil-
ity to discriminate between up to 10 levels of stress.
Moreover, the fits indicate that the use of the SUD as
ground truth for stress assessment is adequate.
5 DISCUSSION
We have explored the feasibility of CAD for men-
tal health care, which can help both in daily life
and in therapy. To assure a clinically valid assess-
ment of stress, previously collected data of 25 PTSD
patients was used, see also Section 2. The stress
level of the patients was assessed by the SUD and
their speech characteristics were mapped upon the
SUD ; hence, a behavioral and an indirect physiolog-
ical measure. The MLP neural network was used to
classify the speech sample, with the SUD scores as
ground truth. Correct classification rates of 82.4%,
72.4%, and 36.3% were achieved on, respectively, 2,
3, and 10 SUD levels. Given the fact that the complete
research is conducted on patients in their regular clin-
ical setting, this underlined the feasibility of a CAD
for mental health care.
HEALTHINF 2012 - International Conference on Health Informatics
496

0
20
40
60
80
100
120
140
160
180
200
220
240
260
280
300
2 3 4 5 6 7 8 9 10
Delta and relative classification rates (in %)
Levels of SUD (N)
dC
N
rC
N
linear fit dC
N
linear fit rC
N
Figure 1: The relative classification results (in %) of 11
principal components based on 28 parameters of speech
features using a Multi-Layer Perceptron neural network
(MLP). The Subjective Unit of Distress (SUD) was taken as
ground truth, using quantizations between 2 and 10. Both
the delta classification rate (dC
N
; see Eq. 1) and the relative
classification rate (rC
N
; see Eq. 2) are reported as well as
the linear fit of their performance in relation to the level of
quantization of the SUD.
A detailed report of the two studies conducted
can be found in (van den Broek et al., 2011). Ad-
ditional analyses in line with the analyses presented
in this article are reported in (van den Broek et al.,
2012). These analyses distinguish between the two
studies conducted. Moreover, in addition to the MLP,
k-nearest neighbors (k-NN) and a support vector ma-
chine were employed. The interested reader is kindly
directed to (van den Broek et al., 2012). Moreover,
it would be of interest to validate the current signal
processing and pattern recognition pipeline on new
(unseen) data sets. Such data sets could comprise,
for example, different patient groups and/or different
methods of emotion elicitation.
A limitation of the current research can be found
in the unbalanced data set (He and Garcia, 2009). Not
all SUD scores were chosen equally by the patients
nor is their distribution Gaussian. Possibly, this issue
is even more prominently present in the quantization
of the SUD scores in a smaller number of bins. Hence,
the chance level for classification (µ
N
) as reported in
Table 1 is not the actual chance level. The linear fits
presented in Figure 1 are possibly not as strong as pre-
sented in this figure (He and Garcia, 2009). However,
the unbalanced data set could also have declined the
chance level and, hence, the to be expected classifica-
tion level. Therefore, we feel it is justified to give this
straightforward intuitive representation of the classi-
fication rates for the different quantizations.
The success of machines in sensing emotions by
way of the speech signal ranges from 25% correct
classification on 14 emotions (Banse and Scherer,
1996) to 73.5% correct classification on 6 emo-
tions (Yang and Lugger, 2010). However, these re-
sults are not stable and have been shown to be hard
to replicate. This is well illustrated by the structured
benchmark conducted by (Schuller et al., 2011), who
report up to 71% (2 classes) and 44% (5 classes)
correct classification. Apart from the differences in
classification rate and the number of classes to be
distinguished, these studies can both be questioned
with respect to their ecological validity of the ex-
perienced emotions. Often they employ methods to
elicit stress that are not validated on their effective-
ness. Exceptions to this, such as the Trier Social
Stress Test (Kirschbaum et al., 1993), are seldom used
in engineering and applied sciences. Therefore, we
feel the need to stress that a careful interpretation of
laboratory results is needed. A one-on-one mapping
between lab and real-world results cannot be taken for
granted (Picard, 2010; van den Broek, 2010). The
current research deviates from common practice of
speech-based stress recognition in its use of a clini-
cally valid data set.
In sum, a leap was made towards modeling stress
through an acoustic model, which can be applied in
our daily lives and in mental health care settings. By
the specific research design, it was ensured that bursts
of authentic chronic stress were measured. The pre-
cision of the SUD as an instrument to assess expe-
rienced stress, as was claimed by clinical practition-
ers, was confirmed. Moreover, a set of features de-
rived from the speech signal was shown to enable
the detection of stress using an MLP neural network.
This shows that unobtrusive and ubiquitous automatic
assessment experienced stress is both possible and
promising and can already be applied as a reliable in-
strument in clinical settings.
ACKNOWLEDGEMENTS
We thank all Post-Traumatic Stress Disorder (PTSD)
patients who participated voluntarily in this research.
We are grateful to four reviewers for their helpful
suggestions on an earlier draft of this article. More-
over, we thank Lynn Packwood (HMI, University of
Twente, NL) for proofreading this article.
REFERENCES
Banse, R. and Scherer, K. R. (1996). Acoustic profiles in
vocal emotion expression.
Journal of Personality and
Social Psychology
, 70(3):614–636.
COMPUTER AIDED DIAGNOSIS FOR MENTAL HEALTH CARE - On the Clinical Validation of Sensitive Machines
497

Brosschot, J. F. (2010). Markers of chronic stress: Pro-
longed physiological activation and (un)conscious
perseverative cognition.
Neuroscience & Biobehav-
ioral Reviews
, 35(1):46–50.
Cowie, R., Douglas-Cowie, E., Tsapatsoulis, N., Votsis,
G., Kollias, S., Fellenz, W., and Taylor, J. G. (2001).
Emotion recognition in human–computer interaction.
IEEE Signal Processing Magazine , 18(1):32–80.
Craig, D. A. (2002). How do you feel? Interoception: The
sense of the physiological condition of the body.
Na-
ture Reviews Neuroscience
, 3(8):655–666.
Domschke, K., Stevens, S., Pfleiderer, B., and Gerlach,
A. L. (2010). Interoceptive sensitivity in anxiety and
anxiety disorders: An overview and integration of
neurobiological findings.
Clinical Psychology Re-
view
, 30(1):1–11.
El Ayadi, M., Kamel, M. S., and Karray, F. (2011). Sur-
vey on speech emotion recognition: Features, classi-
fication schemes, and databases.
Pattern Recognition ,
44(3):572–587.
Fillingim, R. B., Roth, D. L., and Cook III, E. W. (1992).
The effects of aerobic exercise on cardiovascular, fa-
cial EMG, and self-report responses to emotional im-
agery.
Psychosomatic Medicine , 54(1):109–120.
Hall, M., Frank, E., Holmes, G., Pfahringer, B., Reutemann,
P., and Witten, I. H. (2009). The WEKA data mining
software: An update.
ACM SIGKDD Explorations
Newsletter
, 11(1):10–18.
He, H. and Garcia, E. A. (2009). Learning from imbalanced
data.
IEEE Transactions on Knowledge and Data En-
gineering
, 21(9):1263–1284.
Healey, J. A. and Picard, R. W. (2005). Detecting stress dur-
ing real-world driving tasks using physiological sen-
sors.
IEEE Transactions on Intelligent Transportation
Systems
, 6(2):156–166.
Kirschbaum, C., Pirke, K. M., and Hellhammer, D. H.
(1993). The “Trier Social Stress Test”– A tool for in-
vestigating psychobiological stress responses in a lab-
oratory setting.
Neuropsychobiology , 28(1–2):76–81.
Picard, R. W. (2010). Emotion research by the people, for
the people.
Emotion Review , 2(3):250–254.
Rossing, T. D., Dunn, F., Hartmann, W. M., Campbell,
D. M., and Fletcher, N. H. (2007).
Springer handbook
of acoustics
. Berlin/Heidelberg, Germany: Springer-
Verlag.
S´anchez-Meca, J., Rosa-Alc´azar, A. I., Mar´ın-Mart´ınez, F.,
and G´omez-Conesa, A. (2010). Psychological treat-
ment of panic disorder with or without agorapho-
bia: A meta-analysis.
Clinical Psychology Review ,
30(1):37–50.
Scherer, K. R. (2003). Vocal communication of emotion:
A review of research paradigms.
Speech Communica-
tion
, 40(1–2):227–256.
Schuller, B., Batliner, A., Steidl, S., and Seppi, D. (2011).
Recognising realistic emotions and affect in speech:
State of the art and lessons learnt from the first chal-
lenge.
Speech Communication , 53(9–10):1062–1087.
Schwabe, L., Wolf, O. T., and Oitzl, M. S. (2010). Memory
formation under stress: Quantity and quality.
Neuro-
science & Biobehavioral Reviews
, 34(4):584–591.
Van den Broek, E. L. (2010). Robot nannies: Future or
fiction?
Interaction Studies , 11(2):274–282.
Van den Broek, E. L., van der Sluis, F., and Dijkstra, T.
(2009). Therapy Progress Indicator (TPI): Combin-
ing speech parameters and the subjective unit of dis-
tress. In
Proceedings of the IEEE 3rd International
Conference on Affective Computing and Intelligent
Interaction, ACII
, volume 1, pages 381–386, Amster-
dam, The Netherlands. IEEE Press.
Van den Broek, E. L., van der Sluis, F., and Dijkstra,
T. (2011).
Telling the story and re-living the past:
How speech analysis can reveal emotions in post-
traumatic stress disorder (PTSD) patients
, volume 12
of
Philips Research Book Series , chapter 10, pages
153–180. Dordrecht, The Netherlands: Springer Sci-
ence+Business Media B.V.
Van den Broek, E. L., van der Sluis, F., and Dijkstra, T.
(2012). Cross-validation of bimodal health-related
stress assessment.
Personal and Ubiquitous Comput-
ing
, 16:[in press].
Van den Broek, E. L. and Westerink, J. H. D. M. (2009).
Considerations for emotion-aware consumer products.
Applied Ergonomics , 40(6):1055–1064.
Van der Sluis, F., van den Broek, E. L., and Dijkstra, T.
(2010). Towards semi-automated assistance for the
treatment of stress disorders. In
HealthInf 2010: Pro-
ceedings of the Third International Conference on
Health Informatics
, pages 446–449, Valencia, Spain.
INSTICC – Institute for Systems and Technologies of
Information, Control and Communication.
Van der Sluis, F., van den Broek, E. L., and Dijkstra, T.
(2011). Towards an artificial therapy assistant: Mea-
suring excessive stress from speech. In
Proceedings
of the International Conference on Health Informatics
,
pages 357–363, Rome, Italy. Portugal: SciTePress.
Wolpe, J. (1958).
Psychotherapy by reciprocal inhibition .
Stanford, CA, USA: Stanford University Press.
Yang, B. and Lugger, M. (2010). Emotion recognition from
speech signals using new harmony features.
Signal
Processing
, 90(5):1415–1423.
HEALTHINF 2012 - International Conference on Health Informatics
498