Optimization of Sitting Posture Classification based on
Anthropometric Data
Leonardo Martins
1,2
, Bruno Ribeiro
1
, Rui Almeida
1
, Hugo Pereira
1
, Adelaide Jesus
1,3
,
Cláudia Quaresma
1,3
and Pedro Vieira
1,3
1
Department of Physics, Faculdade de Ciências e Tecnologias, Universidade Nova de Lisboa,
Quinta da Torre, 2829-516, Caparica, Portugal
2
UNINOVA, Institute for the Development of New Technologies, Quinta da Torre, 2829-516, Caparica, Portugal
3
LIBPhys-UNL, Department of Physics, Faculdade de Ciências e Tecnologias, Universidade Nova de Lisboa,
2829-516 Caparica, Portugal
Keywords: Intelligent Chair, Pressure Sensors, Sitting Posture, Classification, Algorithmic Optimization.
Abstract: An intelligent chair prototype was developed in order to detect and correct the adoption of bad sitting
postures during long periods of time. A pneumatic system was enclosed in the chair (4 air bladders inside
the seat pad and 4 in the backrest) to classify 12 standardized sitting postures, with a classification score of
80.9%. Recently we used algorithmic optimization applied to the existing classification algorithm (based on
Neural Networks) to split users (using Classification Trees) by their sex and used two different previously
trained Neural Networks (Male and Female) to get an improved classification of 89.0% when the user was
identified and 87.1% for unidentified users. In this work we aim to investigate the usage of the
anthropometric information (height and weight) to further optimize our classification process. Here we use
four Machine Learning Techniques (Neural Networks, Support Vector Machines, Classification Trees and
Naive Bayes) to automatically split the users in 2 classes (above and below the specific anthropometric
median value). Results showed that Classification Trees worked best on automatically separating the body
characteristics (i.e. Height) with a global optimization of 88.3%. During the classification process, if the
user is identified, we skip the splitting step, and this optimization increases to 90.2%.
1 INTRODUCTION
There has been a growing interest in developing
intelligent chairs capable of detecting a person’s
sitting posture and alerting that person to improve
his or her sitting posture. Numerous researchers
applied sheets of surface-mounted pressure sensors
placed as if in a 2D array or used statistical
techniques to find the best way to place singular
force-sensitive resistors (Zhu et al., 2003; Tan et al.,
2001; Zheng and Morrell, 2010; Mutlu et al., 2007;
Daian et al., 2007; Goossens et al., 2012). Other
groups implemented sensing textiles in the chair
(Forlizzi et al., 2005). Most of these chairs alert the
users by using vibrotactile motors or by computer
pop-ups (Haller et al., n.d.; Schrempf et al., 2011;
Zheng and Morrell, 2010). Another group used 36
intelligent pneumatic actuators over sensing plates to
detect and guide the sitting posture (Faudzi et al.,
2010).
These intelligent chairs, which have shown the
capability of monitoring physiological parameters
(e.g. heart rate) (Griffiths and Saponas, 2014) or
monitor everyday activities, are starting to be
implemented in real homes for year-long tests
(Palumbo et al., 2014) and they are needed because
our society spends long periods of time in the
workplace and even at home in the sitting position.
This sedentary lifestyle has been associated with an
increased risk of cardiovascular and musculoskeletal
diseases, although some studies have not been able
to prove direct and causal correlation between sitting
time and those disorders (Chau et al., 2010;
Hartvigsen et al., 2000; Owen et al., 2010; Owen et
al., 2014; Roffey et al., 2010). Musculoskeletal
disorders were recognized as one of main causes of
work-related disability and loss of productivity in
industrialized countries (Ramdan et al., 2014;
Punnett and Wegman, 2004), so there is a necessity
for monitoring and prevention of those health
406
Martins, L., Ribeiro, B., Almeida, R., Pereira, H., Jesus, A., Quaresma, C. and Vieira, P.
Optimization of Sitting Posture Classification based on Anthropometric Data.
DOI: 10.5220/0005790104060413
In Proceedings of the 9th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2016) - Volume 5: HEALTHINF, pages 406-413
ISBN: 978-989-758-170-0
Copyright
c
2016 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
dysfunctions.
When an individual is sitting, most of the
bodyweight is supported by the ischial tuberosities,
the thigh and gluteal muscles, while the rest is
transferred by the feet and armrests when they are
present (Pynt et al., 2001). During extended periods
of sitting, there is a decrease of the lumbar lordosis,
which has been implicated in increasing the physical
risk factors related to back, neck and shoulder pain
(Ariëns et al., 2001; Juul-Kristensen et al., 2004),
due to anatomical changes and degeneration of
intervertebral disks and joints, especially the lumbar
disks (Adams and Hutton, 1986; Kingma et al.,
2000; Billy et al., 2014). If a person is sitting in a so
called ‘bad posture’ (for example sitting in a leaned
back position without lumbar support, see Figure 3
for other examples), the risk of musculoskeletal
disorders increases (Lis et al., 2007)
The increase in these health disorders supports
the necessity for their monitoring and prevention,
leading to the development of chair prototypes that
identify several sitting positions and then alert or
correct the adoption of bad postures over extended
periods. Our first prototype had 4 air bladders placed
in the seat pad and 4 in the backrest, with pressure
sensors that measured the internal pressure of the
bladder. We used Artificial Neural Networks (ANN)
to classify 11 standard sitting postures, with 70%
accuracy, and we were able to do a real-time
classification of 8 postures, with 90% accuracy. This
prototype had had a rudimentary correction
algorithm based on Boolean logic (Martins et al.,
2014; Martins et al., 2013).
The second prototype was built in order to
overcome the gaps identified in the first prototype,
mainly the introduction of a vacuum pump to control
efficiently the air inside the bladders, the design of
industrially constructed air bladders and the
reorganization of the communication protocols
(Pereira et al., 2015). We then revised our
classification and correction algorithms and
introduced Fuzzy Logic to the existing ANN
algorithms, which was able to integrate time spent in
each posture (recognized by the ANN) and was able
to identify intermediate postures, other than the 12
standard ones and correct them based on fuzzy logic
actuators (Ribeiro et al., 2015). This work precedes
our previous implementation of algorithmic
optimization, applied to the second prototype in
order to improve posture classification performance,
based on the sex of the users (Ribeiro et al., 2015). It
continues the trend of classification optimization by
using the anthropometric information of the users
(height and weight) to surpass the previous
classification Accuracy, by testing various
classification methods to split the users. This study
was also driven by the discovery that our previous
classification algorithms (with leave-one-out
strategy to train with 49 users and test with the last
one) had some difficulties in the classification of
users with weights between 60 and 73 Kg and
heights between 173 and 190 cm (highlighted in the
red square in Figure 1).
Figure 1: (A) Classification Performance for each
participant regarding their Anthropometric Data (Height
and Weight), based on a Neural Net-work with 3 Layers
and 15 Neurons.
2 EXPERIMENTS AND
METHODS
2.1 Equipment – Sensors and
Pneumatic Actuators
For this work we use the second prototype that was
previously built, with 8 industrially made
polyurethane bladders, and with new features that
were improved from the first prototype, as
previously mentioned (Pereira et al., 2015). The
main objective was that the bladders covered and
distinguished the anatomical areas involved in the
weight transfer during the seated posture (Pynt et al.,
2001), such as the scapula, the ischial tuberosities,
the posterior thigh region, the lumbar spine (Martins
et al., 2014). The air bladders (see Figure 2-A for
configuration) were placed inside the original
padding foam (as can be observed in Figure 2-B, the
chair maintains the original integrity) (Pereira et al.,
2015). All the sensors and the pneumatic circuits
were integrated in eight small boxes that were
inserted in the backrest and connected to the lower
part of the seat pad, which makes all the electronics
and pneumatic circuit invisible to the user (Pereira et
Optimization of Sitting Posture Classification based on Anthropometric Data
407
al., 2015).
Figure 2: (A) Air bladder schematic (B) External aspect of
the chair prototype.
2.2 Experimental Design - Participants
and Procedure
The same dataset is used as in the previous
optimization work (Ribeiro et al., 2015) (see Table 1
for the participants information). We split the users
based on their weight and height (see dashed lines in
Figure 1). Just as in the previous work protocol, we
use a value of 5 sec for bladder inflation and also
asked the subjects to empty their pockets and adjust
the stool to their popliteal height and to keep their
hands on their thighs (Ribeiro et al., 2015).
Table 1: Data of the participants in the experiment,
namely, Sex, Age, Weight and Height. Note:
a
Values for
Average±Standard Deviation and (M/F) corresponds to
(Male/Female).
Participants Information
Value
Median
Number of Subjects
(M/F)
50 (25/25) -
Age (years)
a
26,4±9,5 -
Weight (Kg)
a
66,8±12,8 67
Height (cm)
a
170,5±9,8 171
Our experiment consists of two tests, the first
involved showing a presentation of the postures P1
to P12 (see Figure 2), each for a duration of 20
seconds, asking the subject to mimic those postures
without leaving the chair. In the second we used the
same presentation, repeating every posture two
times, but after every 20 seconds we ask the subject
to leave the chair, take a few steps and sit back. The
twelve postures (P1 to P12 – see Figure 2) used in
this experiment represent the most common sitting
postures found in office settings (Vergara and Page,
2000; Mutlu et al., 2007; Zheng and Morrell, 2010;
Martins et al., 2014).
Figure 3: Seated postures classes: (P1) seated upright,
leaning (P2) forward (P3) back (P4) back with no lumbar
support (P5) left (P6) right (P7) right leg crossed (P8) right
leg crossed, left lean (P9) left leg crossed (P10) left leg
crossed, right lean (P11) left leg over right (P12) right leg
over left.
Not all of the 20 sec of acquisition were used (as
in previous experiments), due to the existence of a
Transition zone, where the sensor values are not
stable (Martins et al., 2014; Ribeiro et al., 2015;
Pereira et al., 2015). We extracted 100 data-points,
corresponding to 12.5 sec with a sampling of 8 Hz.
Pressure maps were done, by averaging 20
acquisitions, obtaining 5 maps so out of the 100
data-points, for a total of 9000 (50 subjects * 3
repetitions * 5 pressure maps * 12 postures)..
This (P1) pressure is used as a baseline by
subtracting its average from the entire data-points
(9000 maps). After the calibration, the maps are
normalized to an interval of [-1, 1] to use as inputs
for the Posture Classification Algorithm based on
Artificial Neural Networks (ANNs), based on the
average pressure values of the P1 posture of each
subject (Pereira et al., 2015; Martins et al., 2014). To
create ANNs we use the MATLAB® Neural
Network Toolbox™. The optimization of the
Posture Classification is based on using the baseline
pressure as an input to a Pre-Process Classification
Algorithm that is going to classify the participants
according to their anthropometric information.
HEALTHINF 2016 - 9th International Conference on Health Informatics
408
2.3 Classification Algorithms
Here we use four supervised machine learning (ML)
techniques: Artificial Neural Networks (ANNs),
Support Vector Machines (SVM), Classification
Trees (CT) and Naive Bayes (NB) to create a Pre-
Process Classification Algorithm that splits the
participants based on their anthropometric
information. These techniques are widely used in
biomedical applications (Kotsiantis, 2007; Singh et
al., 2014), are the most reliable in supervised
learning and can be easily implemented with specific
libraries (Abeel, 2009) in simple computational
architectures, such as a single-board computers (e.g.
Raspberry Pi) or Mobile Devices (smartphones or
tablets). To train and test each method we use the
MATLAB® Neural Network Toolbox™ (MNNT)
and the MATLAB® Statistics Toolbox™ (MST). To
estimate the performance of each ML technique we
used the 10-fold cross validation, using the
cvpartition’ function. The results are obtained by
calculating the Accuracy of the 2 class separation
problem (below and above the specific
anthropometric information), as the above the
Median can be considered the True Positive and the
Below the Median the True Negative of the test.
ANN-based algorithms have been shown to be
useful in many engineering and biomedical
applications (Paliwal and Kumar, 2009). We already
use ANNs for the Posture Classification, as they
showed the ability to handle very well that
multiclass problem. They also have an advantage of
being easily exported to mobile applications (using
the weights and bias matrices).
The Classification and Regression Trees (CART)
methods that are still being widely used in
biomedical applications (Podgorelec et al., 2002),
were first presented by Breiman and colleagues in
1984 (Breiman et al., 1984). In this work we use the
fitctree from the MST.
SVM techniques were first presented to separate
a binary class problem (Boser et al., 1992) and have
been applied to Biomedical and Biotechnology
applications, such as face recognition (Cyran et al.,
2013) or using gene expression to classify different
cancers (Noble, 2006) and classifying objects such
mass spectra (Noble, 2003; Cyran et al., 2013),
proteins (Noble, 2003), DNA sequences (Noble,
2003). Here we have also binary classification
problem, so we used the fitcsvm function, present in
the MST.
Naive Bayes is a simple and scalable technique
that has been introduced in the 1950’s and has also
been used in biomedical applications (Singh et al.,
2014). Here we use the ‘fitcnb’ function, from MST
and then changed the kernel distributions.
3 RESULTS AND DISCUSSION
3.1 Classification Optimization based
on Anthropometric Information
3.1.1 Neural Network Optimization
To search for the optimized parameters of the
Posture Classification based on Neural Networks,
we tested various combinations of layers, neurons
(as can be seen in Table 2), using the ‘tansig’
transfer function and the ‘scaled conjugate gradient
backpropagation’ (SCG) training function (using the
default parameters), which proved to be the most
accurate parameters our previous work (Martins et
al., 2014; Pereira et al., 2015). As can be seen in
Table 2, the best overall result was with 15 Neurons
and 1 Layer with an overall classification of 95.8%
(overall separation of 95.6% for Height and 96.0%
for Weight). Training with 3 Layers is not shown as
the results were lower or around 90%. It is noted that
the 1 layer-15 neurons also had the best results for
the posture classification algorithm in the first
prototype.
Table 2: Results from the Neural Network Optimization.
Number
of
Neurons
Class
Above
the
Median
Below
the
Median
Overall
Class
Separation
15
Height 97.9 93.3 95.6
Weight 97.6 94.4 96.0
20
Height 93.1 91.7 92.4
Weight 94.9 94.7 94.8
25
Height 93.9 96.7 95.3
Weight 92.8 93.9 93.4
30
Height 96.8 94.7 95.7
Weight 95.7 92.3 94.0
15/15
Height 94.4 92.2 93.3
Weight 96.8 93.3 95.0
20/20
Height 93.7 93.0 93.4
Weight 95.1 95.5 94.8
25/25
Height 97.0 92.4 94.7
Weight 93.9 93.5 93.7
30/30
Height 96.1 94.0 95.0
Weight 95.5 94.7 95.1
3.1.2 Classification Trees Optimization
Using the default values from the fitctree function,
we changed the splitting criterion from the Gini's
Diversity Index to the Twoing rule (Breiman et al.,
Optimization of Sitting Posture Classification based on Anthropometric Data
409
1984) and then to the calculation of the node
deviance (Ritschard, 2006). The best score (97.8%)
were obtained with the Gini Index with an overall
separation of 97.8% for Height and 97.9% for
Weight, as seen in Table 3.
Table 3: Classification Trees Optimization results.
Splitting
Criterion
Class
Above
the
Median
Below
the
Median
Overall
Class
Separation
Gini
Height 97.6 98.1 97.8
Weight 98.9 96.8 97.9
Twoing
Height 97.1 97.6 97.3
Weight 98.4 97.9 98.1
Deviance
Height 97.3 98.9 98.1
Weight 97.1 97.3 97.2
3.1.3 Support Vector Machine Optimization
We started the SVM optimization with the default
parameters. In 'Change 1', we standardized the
predictors (using the 'Standardize' flag). In 'Change
2', we changed the 'KernelScale' to automatic, which
uses heuristic procedure to select the kernel scale
value.
In 'Change 1+2' we combined both flags, which
gave the best overall classification of 78.2% (overall
separation of 73.1% for Height and 83.2% for
Weight). In 'Change 3', we changed the 'Box
Constraint' flag to 10 and 0.1 (default is 1), along
with the flags from 'Change 1+2' (see Table 4 for all
Classification Accuracies).
Table 4: Support Vector Machine Optimization results.
Parameter
change
Class
Above
the
Median
Below
the
Median
Overall
Class
Separation
Default
Height 66.7 50.2 58.5
Weight 60.5 69.1 64,8
Change 1
Height 76.8 65.9 71.3
Weight 76.3 88.3 82.3
Change 2
Height 79.2 63.5 71.3
Weight 80.3 85.6 80.3
Change
1+2
Height 80.2 66.1 73.1
Weight 83.1 88.3 83.2
Change 3
(10)
Height 78.4 65.1 71.7
Weight 75.7 88.5 82.1
Change 3
(0.1)
Height 80.0 61.9 70.9
Weight 73.6 81.3 77.5
3.1.4 Naïve Bayes Optimization
Employing the the ‘fitcnb’ function, we started with
the default parameters, and adapted the data
distribution from 'normal' to 'kernel' with 4 possible
kernels: 'normal', 'box', 'epanechnikov' and 'triangle'.
The best results (see Table 5) was obtained with
a 'normal' kernel, with a global score of 79.8%
(78.9% for Height and 80.7% for Weight).
Table 5: Naïve Bayes Optimization results.
Parameter
change
Class
Above
the
Median
Below
the
Median
Overall
Class
Separation
Default
Height 52.3 71.7 62.0
Weight 54.1 78.4 66.3
Kernel
normal
Height 72.5 85.3 78.9
Weight 72.8 88.5 80.7
Kernel
box
Height 72.8 85.9 79.3
Weight 67.2 86.9 77.1
Kernel
epanech-
nikov
Height 73.1 84.3 78.7
Weight 66.7 86.4 76.5
Kernel
triangle
Height 72.3 81.9 77.1
Weight 65.6 84.3 74.9
3.2 Sitting Posture Classification based
on Neural Networks
After doing the class separation (above and below
the median height and weight), we now rely on
using Neural Networks to classify the 12 standard
Sitting Postures.
The chosen parameters were based on the best
results obtained in the previous experiments (Pereira
et al., 2015; Martins et al., 2014), so we also fixed
the SCG algorithm training function and ‘tansig’ for
the transfer functions and tested the Number of
Neurons (15 and 30) and the amount of Layers (1, 2
or 3).
Table 6 shows the obtained results, the best
result was with 15 Neurons and 1 Layer with an
overall classification of 90.0% (overall separation of
90.2% for height and 89.8% for weight).
This simpler configuration is also advantageous
to use, especially in real-time classification, to avoid
the overfitting problem (Martins et al., 2014).
Overtraining of the Algorithms was avoided by
using the ‘cvpartition’ (with 10-fold option), which
then test’s with 10% of the data and trains with 90%,
and repeats this process 10 times and averages the
results. Although there are a lot of parameters that
could have been used for each of the previous
machine learning algorithm, as we expressed in the
previous sections, we wanted to use a simple
approach to the classification process, because we
want to export the Algorithms to a small single-
board computer (e.g. Raspberry Pi) or to a mobile
application.
HEALTHINF 2016 - 9th International Conference on Health Informatics
410
Table 6: Results for Posture Classification based on
Neural Networks.
Number
of
Neurons
Class
Above
the
Median
Below
the
Median
Overall
Separation
15
Height 90.5% 89.9% 90.2%
Weight 89.8% 89.7% 89.8%
30
Height 87.9% 88.3% 88.1%
Weight 86.9% 87.3% 87.1%
15/15
Height 89.5% 87.2% 88.4%
Weight 87.9% 90.0% 89.0%
30/30
Height 89.8% 86.4% 88.1%
Weight 87.7% 90.6% 89.2%
15/15/15
Height 89.0% 90.4% 89.7%
Weight 88.2% 89.2% 88.7%
30/30/30
Height 89.2% 88.6% 88.9%
Weight 87.8% 89.3% 88.5%
4 CONCLUSIONS AND FUTURE
WORK
In prior works, we developed two intelligent sensing
chair prototypes. The first one was developed to
classify 11 standardized sitting postures using 8
pneumatic bladders connected to pressure sensors
(Martins et al., 2014). The second solved the
identified limitations of the first one (using a
vacuum pump to control the deflation of the
bladders, the design of industrially built bladders
and the use of simple computational architectures)
and had a classification score of 80.9% of 12
standard sitting postures . This work aimed to
demonstrate how we could optimize this
classification based on the identification of the user,
and split them by their anthropometric information
(above or below the median height and weight), with
each class having their specific ANN for Posture
classification.
The workflow of the classification optimization
process is shown in Figure 4. This process starts
with the user sitting on the chair prototype and the
pressure sensor acquisition. If the user is identified
in the computer interface, we just directly select the
specific Neural Network for Posture Classification,
based on the anthropometric features. If the user is
not identified, we need to detect which Neural
Network should be used, by using the best Pre-
Process Algorithm. The workflow then continues
with the Calibration and Data processing, finalizing
with the Posture Classification Process based on the
specific ANN.
The Best Pre-Process Algorithm (Classification
Trees with the Gini Index) for our specific problem,
gave an automatic separation score (97.8%), with an
overall separation of 97.8% for Height and 97.9%
for Weight.
Results showed that the best result for the
Posture Classification (using the ANN) was obtained
with Layer of 15 Neurons with an overall
classification of 90.2% for height and 89.8% for
weight, which translates into an overall optimization
of 9.3% (with the height) from the previously
reported result of 80.9% score for 12 standard sitting
postures (Pereira et al., 2015) and a 1.2% increase
over the previous optimization (using the sex of the
user) (Ribeiro et al., 2015) when the user is
identified with their anthropometric information.
Combining the automatic separation (when the
user is not identified), we use a pre-process
classification (based on Decision Trees) to
determine the specific Anthropometric Neural
Network, so by multiplying each specific result we
get an overall classification optimization of 88.3%
for the Height and 87.8% for the Weight, resulting in
an overall optimization of 7.4% over the normal
Posture Classification Algorithm and an increase of
1.2% over the previous optimization process.
Although using the Height optimization gave the
best results, we believe that combining all three
factors (Height, Weight and Sex) into a very
personalized Classification Algorithm will be our
best option to get scores higher than 90% and
optimize the sitting posture process, which will only
be achieved by increasing the participant’s database.
Figure 4: Workflow of the Posture Classification
Optimization Process.
The prototype is still undergoing a series of
operational trials in an office environment to
evaluate the classification algorithms to get realistic
statistical data from of daily postural habits. The
correct classification of different sitting postures is
necessary for the implementation of the posture
Optimization of Sitting Posture Classification based on Anthropometric Data
411
correction algorithms that hopefully will have a
societal impact of reducing the common back and
neck disorders.
ACKNOWLEDGEMENTS
This project (QREN 13330 – SYPEC) is supported
by FEDER, QREN – Quadro de Referência
Estratégico Nacional, Portugal 07/13 and
PORLisboa – Programa Operacional Regional de
Lisboa. The authors wish to thank Eng. Pedro
Duque, Eng. Rui Lucena, Eng. João Belo and Eng.
Marcelo Santos for the help provided in the
construction of the first prototype.
REFERENCES
Abeel, T., 2009. Java-ML: A Machine Learning Library.
Journal of Machine Learning Research, 10, pp.931–
934.
Adams, M. & Hutton, W., 1986. The effect of posture on
diffusion into lumbar intervertebral discs. Journal of
anatomy, 147, pp.121–34.
Ariëns, G.A. et al., 2001. Are neck flexion, neck rotation,
and sitting at work risk factors for neck pain? Results
of a prospective cohort study. Occupational and
environmental medicine, 58(3), pp.200–7.
Billy, G.G., Lemieux, S.K. & Chow, M.X., 2014. Changes
in Lumbar Disk Morphology Associated With
Prolonged Sitting Assessed by Magnetic Resonance
Imaging. PM&R. The journal of injury, function and
rehabilitation, 6(September), pp.790–795.
Boser, B.E. et al., 1992. A Training Algorithm for Optimal
Margin Classifiers. In COLT ’92 Proceedings of the
fifth annual workshop on Computational learning
theory. pp. 144–152.
Breiman, L. et al., 1984. Classification and Regression
Trees, Chapman and Hall/CRC.
Chau, J.Y. et al., 2010. Are workplace interventions to
reduce sitting effective? A systematic review.
Preventive Medicine, 51(5), pp.352–356.
Cyran, K.A. et al., 2013. Support Vector Machines in
Biomedical and Biometrical Applications. In
Emerging Paradigms in Machine Learning. pp. 379–
417.
Daian, I. et al., 2007. Sensitive Chair: A Force Sensing
Chair with Multimodal Real-Time Feedback via
Agent. In ECCE ’07 Proceedings of the 14th
European conference on Cognitive ergonomics:
invent! explore!. pp. 163–166.
Faudzi, A., Suzumori, K. & Wakimoto, S., 2010.
Development of an Intelligent Chair Tool System
Applying New Intelligent Pneumatic Actuators.
Advanced Robotics, 24(10), pp.1503–1528.
Forlizzi, J. et al., 2005. The SenseChair: The lounge chair
as an intelligent assistive device for elders. In DUX
’05 Proceedings of the 2005 conference on Designing
for User eXperience. p. Article No. 31.
Goossens, R.H.M., P, M. & Doelen, V. Der, 2012. An
office chair to influence the sitting behavior of office
workers. Work: A Journal of Prevention, Assessment
and Rehabilitation, 41(Supplement 1), pp.2086–2088.
Griffiths, E. & Saponas, T.S., 2014. Health Chair:
Implicitly Sensing Heart and Respiratory Rate. In
UbiComp ’14 Proceedings of the 2014 ACM
International Joint Conference on Pervasive and
Ubiquitous Computing. pp. 661–671.
Haller, M. et al., Finding the right way for interrupting
people improving their sitting posture. , pp.1–18.
Hartvigsen, J. et al., 2000. Is sitting-while-at-work
associated with low back pain? A systematic , critical
literature review. Scand J Public Health, 28(3),
pp.230–239.
Juul-Kristensen, B. et al., 2004. Computer users’ risk
factors for developing shoulder, elbow and back
symptoms. Scandinavian Journal of Work,
Environment & Health, 30(5), pp.390–398.
Kingma, I. et al., 2000. Monitoring water content in
deforming intervertebral disc tissue by finite element
analysis of MRI data. Magnetic Resonance in
Medicine, 44(4), pp.650–4.
Kotsiantis, S.B., 2007. Supervised Machine Learning: A
Review of Classification Techniques. Informatica,
31(3), pp.249–268.
Lis, A.M. et al., 2007. Association between sitting and
occupational LBP. European Spine Journal, 16(2),
pp.283–298.
Martins, L. et al., 2013. Intelligent Chair Sensor –
Classification and Correction of Sitting Posture. In
XIII Mediterranean Conference on Medical and
Biological Engineering and Computing 2013 IFMBE
Proceedings. pp. 1489–1492, Volume 41.
Martins, L. et al., 2014. Intelligent Chair Sensor:
Classification and Correction of Sitting Posture.
International Journal of System Dynamics
Applications, 3(2), pp.65–80.
Mutlu, B. et al., 2007. Robust, Low-cost , Non-intrusive
Sensing and Recognition of Seated Postures. In UIST
’07 Proceedings of the 20th annual ACM symposium
on User interface software and technology. pp. 149–
158.
Noble, W.S., 2003. Support vector machine applications in
computational biology. In Kernel Methods in
Computational Biology. pp. 71–92.
Noble, W.S., 2006. What is a support vector machine?
NATURE BIOTECHNOLOGY, 24(12), pp.1565–1567.
Owen, N. et al., 2014. Sedentary behaviour and health:
mapping environmental and social contexts to
underpin chronic disease prevention. British Journal
of Sports Medicine, 48(3), pp.174–177.
Owen, N. et al., 2010. Too Much Sitting: The Population-
Health Science of Sedentary Behavior. Exercise and
Sport Sciences Reviews, 38(3), pp.105–113.
HEALTHINF 2016 - 9th International Conference on Health Informatics
412
Paliwal, M. & Kumar, U. a., 2009. Neural networks and
statistical techniques: A review of applications. Expert
Systems with Applications, 36(1), pp.2–17.
Palumbo, F. et al., 2014. Sensor network infrastructure for
a home care monitoring system. Sensors (Basel,
Switzerland), 14(3), pp.3833–60.
Pereira, H. et al., 2015. System for Posture Evaluation and
Correction - Development of a second Prototype for an
Intelligent Chair. In In Proceedings of the 8th
International Conference on Biomedical Electronics
and Devices (BIODEVICES 2015), Lisbon, Portugal.
pp. 204–209.
Podgorelec, V. et al., 2002. Decision Trees: An Overview
and Their Use in Medicine. Journal of Medical
Systems, 26(5), pp.445–463.
Punnett, L. & Wegman, D.H., 2004. Work-related
musculoskeletal disorders: the epidemiologic evidence
and the debate. Journal of Electromyography and
Kinesiology, 14(1), pp.13–23.
Pynt, J., Higgs, J. & Mackey, M., 2001. Seeking the
optimal posture of the seated lumbar spine.
Physiotherapy Theory and Practice, 17(1), pp.5–21.
Ramdan, N.S.A. et al., 2014. On Lower-back Pain and Its
Consequence to Productivity. Journal of Industrial
and Intelligent Information, 2(2), pp.83–87.
Ribeiro, B., Pereira, H., et al., 2015. Optimization of
Sitting Posture Classification Based on User
Identification. In Bioengineering (ENBENG), 2015
IEEE 4th Portuguese Meeting on. pp. 1–6.
Ribeiro, B., Martins, L., et al., 2015. Sitting Posture
Detection using Fuzzy Logic Development of a
Neuro-Fuzzy Algorithm to classify postural transitions
in a sitting posture. In Proceedings of the 8th
International Conference on Health Informatics
(HEALTHINF 2015), Lisbon, Portugal.
Ritschard, G., 2006. Computing and using the deviance
with classification trees. In Compstat 2006 -
Proceedings in Computational Statistics. pp. 55–66.
Roffey, D. M. et al., 2010. Causal assessment of
occupational sitting and low back pain: results of a
systematic review. The spine journal: official journal
of the North American Spine Society, 10(3), pp.252–61.
Schrempf, A. et al., 2011. PostureCare - Towards a novel
system for posture monitoring and guidance. In 18th
World Congress of the International Federation of
Automatic Control (IFAC). pp. 593–598.
Singh, M. et al., 2014. Machine Perception in Biomedical
Applications: An Introduction and Review. J. Biol.
Engg. Res. & Rev., 1(2), pp.20–25.
Tan, H.Z. et al., 2001. A Sensing Chair Using Pressure
Distribution Sensors. IEEE/ASME TRANSACTIONS
ON MECHATRONICS, 6(3), pp.261–268.
Vergara, M. & Page, A., 2000. System to measure the use
of the backrest in sitting-posture office tasks. Applied
Ergonomics
, 31(3), pp.247–254.
Zheng, Y. & Morrell, J., 2010. A Vibrotactile Feedback
Approach to Posture Guidance. In IEEE Haptics
Symposium. pp. 351–358.
Zhu, M., Mart, A.M. & Tan, H.Z., 2003. Template-based
Recognition of Static Sitting Postures. In Proceedings
of The Workshop on Computer Vision and Pattern
Recognition for Human Computer Interaction, held at
the IEEE Conference on Computer Vision and Pattern
Recognition (CVPR03), IEEE Computer Society,
Madison, Wisconsin. pp. 1–6.
Optimization of Sitting Posture Classification based on Anthropometric Data
413