Biomedical Signal Processing and Control 95 (2024) 106306

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

journal homepage: www.elsevier.com/locate/bspc

A method for complex posture recognition during long-term sitting using

neural networks and pressure mapping systems
Isaac Morales-Nolasco a, Sandra Arias-Guzman b, Laura Garay-Jiménez a, *
a
Instituto Politécnico Nacional, Interdisciplinary Professional Unit in Engineering and Advanced Technologies, 2580 IPN Ave, Laguna Ticomán, Gustavo A. Madero,
07340, Mexico City, Mexico
b
EC Service, Corp., 915 South Frontage Rd, Centerville, UT 84014

A R T I C L E I N F O A B S T R A C T

Keywords: This study aimed to identify correct sitting postures during prolonged periods using a pressure mapping system
Posture recognition on the seat along with Convolutional Neural Networks (CNN). The study was conducted in three stages. In the
Sitting first stage, twenty-two volunteers participated to obtain a dataset of pressure maps of three correct postures (flat,
Pressure distribution
short lordosis, and long lordosis) using two pressure mapping systems and validating with two methods. This
Neural network
Transfer learning
involved angle measurement through an Inertial Measurement Unit (IMU) system and image recognition. In the
second stage, a CNN model was trained using data from the first stage that represented each posture, and then
Transfer Learning was implemented using a different pressure mapping system. In the third stage, the system was
used for long-term monitoring based on the model, and feedback on the number of correct postures was provided
to two healthy individuals during prolonged sitting periods for at least 2 h. The system provided feedback on
posture recognition to the participants. Additionally, the participants evaluated their experience with the system
using a Likert-5 questionnaire. The results showed that the system accurately identified the three postures with
an accuracy of 0.854, precision of 0.856, and recall of 0.854. The study found that the data collection meth­
odology was non-invasive and unobtrusive. Moreover, the participants found the system’s feedback under­
standable and helpful in improving their posture. The approach presented in this study has the potential to
facilitate further research on sitting posture since it can be easily adapted to various pressure mapping systems
without altering the methodology.

1. Introduction because they are less expensive, require low computational costs, and do
not need a controlled environment [6]. Many studies have used pressure
The analysis of sitting posture is a widely researched topic due to its sensor arrays or individual sensors placed strategically to provide
common occurrence in daily activities [1–3]. Various methods have posture information [5]. Researchers have successfully classified sitting
been used to detect posture, including video cameras [4], individual postures using pressure sensor data in controlled laboratory conditions
pressure sensors, and pressure sensor arrays [5–7]. [5–7].
W. Albarracín et al. conducted a study on sitting posture analysis Identifying correct postures has also been achieved by using pressure
using the Kinect system. The system’s RGB camera and depth sensor sensors and neural networks. Wang et al. developed a posture recogni­
allowed for capturing 3D motion from fragmented video frames or im­ tion system utilizing Spiking Neural Networks (SNN) and a liquid state
ages. Their findings showed that the system determined a person’s machine. They placed Force Sensitive Resistors (FSRs) on the seat and
posture angles while seated and evaluated whether they were correct or backrest of an office chair to provide real-time feedback. Their algorithm
incorrect. However, direct visual access to the participant was necessary imitated a neuron’s response, considering immediate and previous im­
for detection, resulting in high computational costs and a specific pulses. Wang et al. concluded that their system using FSRs provided
environment setting [4]. sufficient feedback to distinguish postures and that SNN could accu­
Recently, pressure sensors have become the preferred method for rately classify postures [6]. The study only took place in a laboratory and
analyzing sitting posture over computer vision systems [8]. This is lacked user feedback, so actual feedback is required to assess its

* Corresponding autor.
E-mail address: [email protected] (L. Garay-Jiménez).

https://doi.org/10.1016/j.bspc.2024.106306
Received 5 December 2023; Received in revised form 24 February 2024; Accepted 8 April 2024
Available online 17 April 2024
1746-8094/© 2024 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
I. Morales-Nolasco et al. Biomedical Signal Processing and Control 95 (2024) 106306

feasibility under practical conditions. 2.1. Study design

On the other hand, Kim et al. conducted a study to analyze children’s
postures using pressure sensors. They compared traditional machine This study was carried out in three stages. In Stage 1, a systematic
learning methods, such as decision tree (DT), Naive Bayes (NB) classi­ process was proposed to obtain the dataset of pressure maps of three
fier, neural network (NN), and support vector machines (SVM) with the correct postures: flat, short lordosis, and long lordosis. As visual analysis
deep learning Convolutional Neural Network (CNN) method. The results alone cannot accurately differentiate the correct posture, a double
showed that the CNN method was the most accurate for posture classi­ validation method was used to determine the correct postures, which
fication. The study also revealed that children weighing less than 20 kg involved angle measurement through the inertial measurement unit
had lower accuracy measurements than children weighing more than (IMU) system and image recognition. In Stage 2, a CNN model was
40 kg due to the low pressure they exert on the sensors because of their trained using data from Stage 1 that represented each posture, and then
weight [7]. Thus, the research suggested that neural networks were an Transfer Learning was implemented using a different pressure mapping
effective solution for posture classification. system. In Stage 3, based on the model, a system was utilized for long-
C. Ma et al.’s 2017 study aimed to detect different postures using a term monitoring, and feedback to the user was provided on the num­
wheelchair cushion sensor and compare different classification methods ber of correct postures. The study schematic is shown in Fig. 1.
to find the optimal one. They also considered the Body Mass Index (BMI)
to observe the influence on the classification. They classified using five 2.2. The pressure acquisition systems
supervised machine learning algorithms using J48 decision tree, Naive
Bayes, multilayer perceptron (MLP), support vector machine, and k- The acquisition process was carried out with two different sensor
Nearest Neighbors. They reported that the most effective method was systems. The first system utilized was the XSENSOR® interface pressure
J48, with an accuracy of 99.48 %, and MLP obtained an accuracy of mapping system (XSENSOR Technology Corp, Canada). This system
95.50 %. They also used the backward selection method to determine comprises 26 x 64 flexible capacitive pressure sensors, and the data was
which sensors provided the most helpful information for posture recorded using the specialized software X3 PRO V7. Each record con­
detection. They concluded that BMI was not a representative parameter tains the reading of 1664 sensors numbered according to their position
in their classification algorithm but may influence other populations in the seat. The recorded data was exported in.txt format.
with elevated BMI [5]. The second system was the BodiTrak Seat® high-resolution interface
Research on detecting postures has mainly focused on short periods, pressure mapping system (Vista Medical, Canada), which contains 32 x
but assessing postures for longer durations is also important. Sitting with 32 flexible pressure sensors. The system software recorded the data, and
poor posture for extended periods can cause discomfort and pain [9,10], each register contains the reading of 1024 sensors numbered according
particularly among office workers who lead sedentary lifestyles [9]. to their sitting position. The recorded data was exported to.xlsx format.
Research has shown that discomfort in the lower back, thighs, hips, and Pressure distribution maps were obtained from the information
glutes can occur after just two hours of office work [11], and it is rec­ saved in the.xlsx and.txt files, and the neural training and posture
ommended to take breaks and change posture regularly [12–14]. identification were performed utilizing the computer programming
However, there is currently no research on posture in healthy in­ language Python. Table 1 summarizes the technical information of the
dividuals who remain seated for extended periods, such as during sensors used.
workdays. Understanding sitting postures over time could be beneficial
in promoting correct posture and preventing injuries or discomfort. 2.3. Dataset
Therefore, providing tools to aid in these types of studies is essential. In
addition, identifying the best method for posture pattern recognition can According to Kendall [17], the ideal posture relates to the minimum
be used for various user feedback applications and for researchers or stress and pressure conducive to maximal efficiency in using the body.
physicians interested in preventing long-term injuries. For sitting posture, there are different proposed positions [18] that
The aim of this study is to use a pressure mapping system on the seat, reduce stress and pressure on the backbone (see Fig. 2). The most
along with a Convolutional Neural Networks (CNN) model, to identify representative sitting postures in healthy adults are as follows:
correct sitting postures during long periods of sitting. CNNs were chosen 1) Flat (F): In this position, the angles in the thoracic-lumbar and
due to their ability to extract low-level features from data representa­ lumbar regions are vertically aligned.
tion, making them a type of deep learning algorithm. These networks 2) Long Lordosis (LL): This is defined by the angles in the lordosis in
consist of multiple convolution blocks, each with a convolutional layer the thoracic-lumbar and lumbar region.
and a sub-sampling (pooling) layer, and they may also include dropout 3) Short Lordosis (SL): In this position, the flat thoracic-lumbar re­
layers [15]. The final layers are fully connected layers (FC) that extract gion and the lumbar angles in lordosis are similar to the curves of the
high-level features. CNNs are widely used because they can automati­ spine when standing.
cally detect significant features without needing preliminary feature
extraction or human supervision [16]. The proposed method’s robust­ 2.3.1. Study group characteristics
ness was evaluated using transfer learning and two mapping systems. The inclusion criteria were healthy adults. The exclusion criteria
This paper makes three significant contributions: were pregnancy and any musculoskeletal disorders relating to posture
and/or the vertebral column and coccyx. Participants signed an
• A method focusing only on the sitting area pressures for classifying informed consent following the Helsinki Declaration format after the
complex sitting postures. procedure was explained to them.
• Double-checked tagged database of three complex sitting postures For data obtained from XSENSOR®, ten healthy volunteers, two
• Transfer Learning approach to improve the performance of pressure women, and eight men participated. The age was 23 ± 1.5 years, the
maps of different systems. weight was 72.7 ± 9.9 kg, the height was 171 ± 9.9 cm, and the BMI was
24.74 ± 3.2 kg/m2.
2. Methods Then, for data obtained from BodiTrak®, another twelve healthy
volunteers participated, three women and nine men, with age of 27 ±
This study aims to identify when and how long complex postures 10 years, weight of 72.3 ± 11.4 kg, height of 171 ± 7.7 cm, and BMI of
were presented by volunteers during long-term monitoring while sitting 24.46 ± 2.9 kg/m2.
and working on a computer. Finally, in Stage 3, two volunteers participated. These subjects were
both 23 years old, weighed 76 kg, had heights of 170 cm and 178 cm,

2
I. Morales-Nolasco et al. Biomedical Signal Processing and Control 95 (2024) 106306

Fig. 1. Flowchart of the study protocol.

30 min to avoid problems related to excessive data. The first participant

Table 1
session lasted 2 h, and the second lasted three and a half hours.
Technical characteristics of the recording systems used in the experiments
XSENSOR® and BodiTrak®.
XSENSOR® BodiTrak Seat® 2.4. Experiments
Max sampling frequency [Hz] 53 100
Sensing range [mmHg] 5 to 200 1 to 200 2.4.1. Verification of the correct postures
Sensing area per sensor 3.048 x 3.048 1.27 x 1.27 A double verification method was used to ensure accurate posture
[cm x cm]
Total sensing area 81.2, 203.2 46.5, 46.5
identification. This involved using an AMS and image analysis to check if
[cm x cm] the subject’s set of back angles was in accordance with the established
literature on correct posture. This allowed for the collection of pressure
pattern information and the subsequent generation of a posture-type tag
and had BMIs of 26.3 kg/m2 and 24 kg/m2, respectively. for classification. Virtual markers were placed on the subjects’ backs
using Tracker® software based on the original video and AMS reference.
2.3.2. Data acquisition process The position of these markers was displayed in each frame, allowing for
In order to accurately measure different postures, a private area with angle calculation (Eq. (1) formed by three points. Finally, vectors a and
a bed was set up to serve as a distraction-free place to record pressure b were obtained by measuring the markers’ distance from the gold/
distributions. The area was also used for long-term monitoring. The brown lines, as shown in Fig. 2.
volunteer was given instructions on how to wear a harness, which acted ⎛ ⎞
as a restraint system for the Angle Measurement System (AMS). The
⎜a •b ⎟
volunteer was shown three postures to adopt. The AMS was placed on θ=⎜ ⎟ (1)
⎝⃒⃒ ⃒⃒⃒⃒ ⃒⃒ ⎠
the volunteer’s back, over their clothes, at the T2, T7, and L1 vertebrae. ⃒a ⃒⃒b ⃒
A camera was positioned to monitor the curve of their back as an
additional safety measure. As an example, the angle generated by T10 (blue), L3 (pink), and S2
The volunteer was asked to maintain each posture for one minute (purple) from Fig. 2, is calculated as follows:
with a 30-second rest period between each sequence. A total of 5 min per
posture per participant was obtained, with a sampling frequency of 4 Hz. 1. Position of T10, L3, and S2 by the pixels of the image.
Using XSENSOR®, 1200 samples of pressure distribution maps were 2. ais the resulting vector between the coordinates of T10-L3, bis the
taken for each of the ten participants for each of the three postures (Flat, resulting vector between the coordinates of S2-L3.
Long Lordosis, and Short Lordosis), resulting in a total of 12,000 maps. 3. The internal angle between the three points: inverse cosine of the dot
Using BodiTrak®, the sampling time was reduced to 2 min per
product of vectors a and b, divided by the product of their
posture per volunteer, and a total of 4800 samples of pressure distri­
magnitudes.
bution maps were taken for another 12 participants. Furthermore, a
4. The resulting angle is subtracted from 180◦ to obtain the comple­
video recording was made during both stages of the sample collection
mentary angle.
process to help identify the posture.
In the third stage, the sampling rate was 1 sample per second in two
The AMS was built on an IMU and saved the angle on a Secure Digital
long sessions with two volunteers. The entire session was divided into
(SD) flash memory card during movement. The AMS was programmed to

3
I. Morales-Nolasco et al. Biomedical Signal Processing and Control 95 (2024) 106306

Fig. 2. Setup to obtain angles from video of different postures. On the left, the Flat (F) posture is shown; in the center, the Long Lordosis (LL) posture is shown; on
the right, the proposed Short Lordosis (SL) posture is shown. Backbone identification is made with colored points: red T1, green T5, blue T10, pink L3, and purple S2;
An example of an angle generated between three points is shown in gold/brown lines.

capture 30 samples per second. The analysis was conducted on a 50-sec­ kernel size (p × p) [16]. If overfitting was observed during training, a
ond segment while the participants maintained their posture to elimi­ dropout layer was recommended. This layer randomly deactivated a
nate any noise associated with finding the required posture. percentage of the layer’s neurons, accelerating the training process and
The angle datasets obtained from the video and IMU with the 22 reducing overfitting. In the final part of the CNN, the FC layers received
volunteers in Stage 1 were compared to previously reported angles ob­ the features and created a high-level abstraction, similar to a multilayer
tained with the posture angles defined by Claus [18]. perceptron network (MLP) [15]. Finally, the classification was gener­
Analysis of Variance (ANOVA) was conducted to compare the angle ated using an ending layer such as the SoftMax function. Once the
datasets; significance was defined at α = 0.05. The analysis involved training stage was completed, the score represented the probability of a
comparing the range and the mean value of each angle that defines a new instance belonging to a specific class [18].
posture: Thoracic, Thoracic Lumbar, and Lumbar. For sitting posture recognition, we proposed a classification model
based on two-dimensional CNN (2D CNN) to train and classify the three
2.4.2. Convolutional Neural network implementation postures. The network architecture is shown in Fig. 3, and the param­
In the implemented CNN model, the input x was defined by the eters corresponding to the 2D CNN model structure are shown in
image’s height, width, and depth (m × m × r). Then, k kernels were Table 2. The CNN model has two convolutional layers, two max-pooling
applied in each convolutional layer with similar dimensions to the input layers, one dropout layer, and one fully connected layer. The size of the
x ,( n × n × q ) with n < m and r = q; the kernels were the basis set convolution kernel is (3,3), the Stride is two, and the activation func­
associated with the local connections, and hk were the k feature maps tions of the convolutional layer were the Rectified Linear Unit (ReLU)
with similar weight Wk and bias bk parameters and size of p = m − n − 1. function.
These maps, also called filters, were convolved with the input presented The network input data size is 26 × 26 × 3, and it had a data
in Eq. (2) transformation by rotation in a range of 20 and a width and height shift
of 0.2. Before feeding the images into the network, they were pre­
hk = f (W k x + bk ) (2) processed using VGG16 in Phyton. The VGG16 preprocessing involved
converting the images from RGB to BGR and then zero-centering each
In the convolution layer, a dot product was performed between the
color channel is zero-centered with respect to the ImageNet dataset,
weights generated by the map and the input. After that, the filter shifts
without scaling. The convolutional layers extract the features, and the
over the input matrix to compute the next dot product. The number of
classification is made using the dropout layer and max-pooling layers on
pixels shifts is called Stride. Finally, the output was obtained by applying
the ReLU dense layer.
a nonlinearity or activation function.
The open-source TensorFlow framework and Python algorithms
After the convolution and activation operations in the CNN, a pool­
running on NVIDIA GTX960M were implemented to train the CNN. The
ing function was used to down-sample the feature maps to reduce the
hyperparameters are defined as learning rate η = 0.0001, 200 Epochs,
number of network parameters. This pooling function was either a
verbose = 2, loss = categorical cross entropy, using an ADAM optimizer,
maximum or average function applied to each adjacent area with a

4
I. Morales-Nolasco et al. Biomedical Signal Processing and Control 95 (2024) 106306

were extracted by the convolutional layers, dropout layer, and max-

pooling layers, and the classification takes part in the ReLU dense
layer. The input image size of the network was set at 26 × 26 × 3 for
both systems. The parameter values within the shaded area of Table 3
were utilized as initial values, and the final layers of the proposed
network were defined as per Table 3.
The model was trained using 80 % of the data from one of the sen­
sors, and the remaining 20 % was used to evaluate the classification
performance. The performance metrics of the proposed model are ac­
curacy, precision, and recall.

2.4.4. User feedback

In Stage 3, after generating a model capable of identifying the pro­
posed postures, it was implemented to analyze sitting postures over an
extended period for two volunteers.
In this stage of the experiment, the subjects did not wear the AMS,
and the video was not analyzed. The subjects worked on their routine
chores using a computer while having the pressure system in their seats.
The pressure maps were saved every 30 min to prevent data loss.
After two hours, the dataset was evaluated and classified with the
Transfer learning CNN. The analysis was shown to the volunteers in a
dashboard as feedback to improve their posture.
The dashboard was designed to display the user’s sitting posture
throughout their workday, categorizing it as Flat, Long Lordosis, or
Fig. 3. The structure of the Convolutional Neural Network. Short Lordosis. Any posture that deviated from these categories was
considered incorrect. To evaluate the usability of the dashboard, we
asked users to complete a self-reported Likert-5 type survey. The Likert-5
Table 2 questionnaire is comprised of six subjective questions on a scale of 1
The structure of the stand-alone Convolutional Neural Network. (totally disagreed) to 5 (totally agreed).
Layer Kernel Filters Stride Activation Function

Conv2D (3,3) 16 − ReLU

2.4.4.1. Ethics considerations. The study was conducted in accordance
Max-pooling2D (2,2) − 2 − with the provisions of the Declaration of Helsinki. Informed consent was
Dropout − − − − obtained from all volunteers after a comprehensive explanation of the
Conv2D (3,3) 32 − ReLU study. There was no economic retribution for the participation in this
Max-pooling2D (2,2) 2
study.
− −
Flatten − − − −
Dense 3 − − Softmax
3. Results

and batches of 20. 3.1. Dataset

In order to verify whether the model could predict an individual’s
posture using new data, the data of nine out of ten volunteers was used Table 4 shows the summary of the acquired data. Fig. 4 shows an
to train the model, and the remaining volunteer’s data was used to test example of the pressure map images of the three postures obtained with
it. BodiTrak®. Fig. 5 shows a batch of the original pressure maps and their
In order to validate the model, a train/test split method was repeated transformations that were used to train the model from both systems.
seven times into each shuffled dataset to perform a cross-validation. The
training dataset consisted of 80 % of images associated with the three
3.2. Tag assignment results
AMS-tagged postures in the first stage of the protocol. This dataset was
used to train a CNN, and the remaining 20 % of the data was used for
The postures obtained using the system based on the IMU and those
validation. The evaluation metrics of the proposed model are accuracy,
obtained using video were compared with those reported by A. P. Claus
precision, and recall.

Table 3
2.4.3. Transfer learning approach
Structure of Transfer Learning method applied in the CNN model.
Transfer learning (TL) is a method that involves leveraging knowl­
edge gained from one task to improve performance on another related Layer Kernel Filters Stride Activation Phase
Function
task. In transfer learning, a model that has been trained with a large
dataset is used as a starting point. This study conducted an experiment to Conv2D (3,3) 16 − ReLU Pre- trained
evaluate the benefits of transfer learning when using datasets from Max- (2,2) − 2 − network.
pooling2D
different pressure mapping systems, from data gathered from XSEN­ Dropout − − − −
SOR® to BodiTrak® and from BodiTrak® to XSENSOR® to compare the Conv2D (3,3) 32 − ReLU
performance of the proposed classification method. Then, the TL model Max- (2,2) − 2 −
was compared with the stand-alone CNN model for each dataset of the pooling2D
Conv2D (3,3) 16 ReLU Trainable
pressure mapping systems.
−
Max- (2,2) − 2 − parameters
The image size was reduced from 32 x 32 pixels to 26 x 26 pixels to pooling2D
ensure uniformity. The dataset also had the same data transformation, Dropout − − − −
by rotation in a range of 20 and width and height shift of 0.2, and a Flatten − − − −
VGG16 image preprocessing was used to resize the input. The features Dense 3 − − Softmax

5
I. Morales-Nolasco et al. Biomedical Signal Processing and Control 95 (2024) 106306

Table 4 [18]. Each posture, Flat, Long Lordosis, and Short Lordosis, consists of
Data set definitions and volunteer information. three spine-analysis angles: Thoracic, thoracic lumbar, and lumbar.
Sensor Participants Postures Results Fig. 6 is an example of the correspondence in the distribution of the
angle comparison sumarized in Table 5.
XSENSOR® 10 volunteers Flat 12,000 pressure maps per According to the ANOVA results, the Flat position’s angles were
23–27 years old. Long posture similar to what Claus et al. reported. However, there was some variation
av. weight: 72.7 Lordosis 10 database of measurements in the thoracic and lumbar angles of the Long Lordosis. Despite this, the
kg Short of angles of the back while collected data was within the range reported in their study. The Short
av. high: 1.71 m Lordosis sitting.
av. BMI: 24.74 10 videos while sitting
Lordosis posture was discovered to be the most difficult to reproduce.
kg/m2 Overall, the distribution of angle data had features similar to the mean
BodiTrak® 12 volunteers Flat 5760 pressure maps per and range reported by Claus et al [18].
27 ± 10 years old. Long posture
av. weight: 72.3 Lordosis 12 database of measurements
kg ± 11.4 kg Short of angles of the back while
3.3. Classification results
av. high: 1.71 m Lordosis sitting.
± 7.7 cm 12 videos while sitting The CNN network trained with XSENSOR® dataset presents an
av. BMI: 24.46 average accuracy of 84.7 %, The CNN network trained with Boditrak®
kg/m2
dataset presents an average accuracy of 98.4 %. Transfer Learning using
BodiTrak® 2 volunteers Flat 2 h of recordings of pressure
Long maps previous CNN layers of Boditrak® and trained with XSENSOR® shows
Lordosis And 3.5 h of recordings of an average accuracy of 66.7 %. Transfer Learning using previous CNN
Short pressure maps layers of XSENSOR® and trained with Boditrak® shows an average ac­
Lordosis curacy of 85.4 %.
Incorrect
Table 6 summarizes the mean (x) and standard deviation (σ ) of the
postures metrics showing the performance of the classifiers with the two TL
methods. The metrics were obtained by cross-validation on the same
dataset and with the same hyperparameters (N = 7 times each).

Fig. 4. Example of pressure images associated to each posture. Images are obtained from the same subject with BodiTrak®.

Fig. 5. Example of a training batch. The first row shows a batch of original XSENSOR® pressure maps. The second row shows a batch of rotation, shifting, and
vgg16 preprocessed XSENSOR® pressure maps. The third row shows a batch of original BodiTrak® pressure maps. The fourth row shows a batch of rotation, shifting,
and vgg16 preprocessed BodiTrak® pressure maps.

6
I. Morales-Nolasco et al. Biomedical Signal Processing and Control 95 (2024) 106306

Fig. 6. Comparison of the angle distributions. The Angle Measurement System (AMS) and video data distribution are compared to the adjusted distribution of the
results reported by Claus for the Flat thoracic angle.

Additionally, the study results were compared with the results of other help improve their overall health and quality of life [13]. This research
classification methods used to detect various sitting postures, as pre­ introduced a reliable approach to detecting and categorizing good
sented in Table 7. sitting postures, which can reduce pain and discomfort, as indicated by
Additionally, after training the CNN model with the data of nine out previous studies.
of ten volunteers and testing with the remaining volunteers’ data, the In this study, the pressure mappings were converted to images and
accuracy during the training was 86.4 %, while in the validation, it was the CNN because this technique is considered a well-established tech­
34.9 %. This indicates that although pre-training with data from other nique as automatic image feature extractor technique; in recent years,
subjects can be helpful, personalized fine-tuning of the CNN input is still using CNN as a feature extractor has become useful to obtain Yoga poses
necessary. from images [19], and even the basic characteristics of sitting postures
such as back, forward, left and right-leaning into FSR mapping [14,20].
In evaluating sitting posture, some studies have focused on wheel­
3.4. Feedback results
chair users or long-term drivers. However, to the best of the authors’
knowledge, no specific dataset is available related to the long-term
The scores from the Likert-5 questionnaire were as follows: for
sitting of healthy individuals. Therefore, the use of Transfer Learning
question 1 (sensor perception) was 5, for question 2 (no distraction) was
(TL) has been helpful in demonstrating the strength and reliability of the
5, for question 3 (posture improvement) was 3.5, question 4 (under­
model used in this study, which was able to effectively utilize two
standable graphs) was 4.5, for question 5 (prevalence graph usefulness)
different pressure mapping systems. The XSENSOR® system has less
was 4, and for question 6 (total time sitting in different posture) was 4.5.
resolution than the BodiTrak® sensor. On a stand-alone version, the test
These individual question ratings collectively contributed to an overall
presents a relationship between sensor resolution and the performance
score of 4.25, reflecting the general system acceptance.
metrics (accuracy 0.847 ± 0.011 vs. 0.984 ± 0.005, respectively).
Participants agreed that they did not perceive the pressure mapping
Since deep learning methods have shown better results with large
system BodiTrak® during extended work hours. They also reported that
datasets, augmentation methods is a common practice, and transfer
having a screen with the FSA program for data collection did not distract
learning is recommended when transferring from larger to smaller
them from their work. These findings suggest that the data collection
datasets. The TL version obtained a resolution from the larger to smaller
method was non-invasive and went unnoticed by the participants. Par­
dataset (0.854 ± 0.007), similar to the stand-alone CNN with the
ticipants slightly agreed that the posture report helps improve posture,
XSENSOR performance (0.847 ± 0.011). However, in the reversal TL
indicating that the report could be improved, and one comment sug­
version, the performance decreased from the smaller to the larger
gested that continuous feedback would be more helpful. The partici­
dataset (0.667 ± 0.005). It was expected because the bigger version was
pants perceived the graphs as understandable and useful for posture
used to train the feature extraction and the last part to tune-fine the
improvement.
classification. In the opposite way, the TL application means that the
smaller dataset was used to train to extract the features and a low-
4. Discussion resolution dataset for fine-tuning. The models learned useful features
and patterns relevant to similar tasks in all tested cases, even with
Sitting with poor posture for extended periods can lead to discomfort different sensor characteristics. So, fine-tuning or additional training
and injuries [3]. To reduce such risks, sitting properly or limiting was performed on the new task-specific dataset.
abnormal postures is crucial [11,12]. A smart system that identifies Other studies by J. Wang, Y. Min, and C. Ma examined various
irregular sitting postures and reminds individuals to correct them can

7
I. Morales-Nolasco et al. Biomedical Signal Processing and Control 95 (2024) 106306

Table 5 Table 6
Statistics of the angle measurement with the AMS and image systems associated Summary of performance of the different training methods with both datasets.
with the three postures: Flat, Long, and Short Lordosis. Method Accuracy Precision Recall F1-score
Position Angle (◦ ) Video AMS Claus (x ± σ) (N (x ± σ) (N (x ± σ) (x ± σ)
= 7) = 7) (N = 7) (N = 7)
FLAT Thoracic Mean:17.1 Mean:17.9 Mean: 19.0
Range: Range: Range: (15.0, CNN (XSENSOR dataset) 0.847 ± 0.849 ± 0.847 ± 0.846 ±
(7.9,16.3) (16.1,19.7) 23.0) 0.011 0.012 0.011 0.012
ANOVA: ANOVA: CNN (BodiTrak dataset) 0.984 ± 0.984 ± 0.984 ± 0.984 ±
F = 13.61 F = 0.85 0.005 0.005 0.005 0.005
P value = P value = CNN (XSENSOR dataset) + 0.854 ± 0.856 ± 0.854 ± 0.854 ±
0.0003 0.350 Transfer Learning 0.007 007 0.008 0.008
Thoracic Mean: 5.1 Mean: 5.5 Mean: 4.6 (BodiTrak dataset)
Lumbar Range: Range: Range: (0.0, CNN (BodiTrak dataset) + 0.667 ± 0.672 ± 0.669 ± 0.667 ±
(4.3,6.0) (4.2,6.8) 9.2) Transfer Learning 0.005 0.004 0.004 0.005
ANOVA: ANOVA: (XSENSOR dataset)
F = 2.20 F = 2.57
P value = P value =
0.140 0.130 In contrast, our research used more complex postures that might not
Lumbar Mean: − 2.7 Mean: − 7.2 Mean: − 1.5 be visually identified and are being reported as beneficial in preventing
Range: Range: (− 8.6, Range: (− 5.2,
back pain. Because of this complexity, the double-validation method to
(− 1.9,3.4) − 5.9) 3.7)
ANOVA: ANOVA: detect a correct posture by measuring the angles in the back using an
F = 5.37 F = 53.61 IMU system and by image recognition used in the study helped train the
P value = 0.02 P value = 1x CNN model effectively with data representing each posture.
10− 10 It is worth noting that Wang and Ma achieved 95 % accuracy with the
LONG Thoracic Mean:16.6 Mean:18.5 Mean: 16.2
aid of back sensors, and we want to reduce the sensors needed, so we
LORDOSIS Range: (15.7, Range: (16.7, Range:
17.5) 20.2) (12.2,20.2) only used a seat pressure distribution sensor to distinguish between
ANOVA: ANOVA: positions.
F = 0.77 F = 5.85 It is worth noting that our model’s metrics might seem to be lower
P value = 0.38 P value = 0.01
compared to other models, as shown in Table 7. However, it is essential
Thoracic Mean: − 4.1 Mean: − 6.3 Mean: − 2.6
Lumbar Range: (− 4.8, Range: (− 7.9, Range:
to consider that the pressure maps generated during complex postures,
− 3.4) − 4.8) (− 7.0,1.8) such as long lordosis, short lordosis, and flat, were more complicated
ANOVA: ANOVA: than those generated when a person is leaning toward one direction.
F = 7.62 F = 16.51 This indicates that our model can capture the inherent complexity of
P value = P value =
these postures and is capable of recognizing a higher level of complexity.
0.006 0.0001
Lumbar Mean: − 10.1 Mean: − 9.3 Mean: − 9.2 Moreover, our findings demonstrate the reliability of our model and
Range: (− 11.3, Range: (− 10.5, Range: (− 12.4, suggest the potential for further investigation into posture dynamics
− 8.9) − 8.1) − 6.0) over extended periods.
ANOVA: ANOVA:
The robustness of the posture classification methodology was eval­
F = 2.01 F = 0.04
P value = 0.15 P value = 0.83
uated using two different pressure mapping systems and preprocessing
SHORT Thoracic Mean:19.4 Mean:16.7 Mean: 18.9 the batch input with augmentation methods.. First, the XSENSOR®
LORDOSIS Range: Range: Range: system was used, and then it was replaced with the BodiTrak® system to
(18.4,20.3) (15.5,18.0) (14.0,23.0) evaluate the posture. Transfer learning tests were implemented to
ANOVA: ANOVA:
confirm the feasibility and versatility of the algorithms to use different
F = 0.39 F = 8.61
P value = P value = pressure distribution systems. This work created the infrastructure for
0.531 0.004 further studies of sitting posture, which can be adapted without
Thoracic Mean: 5.3 Mean: 5.2 Mean: 3.1 changing the methodology. Moreover, based on the participant feed­
Lumbar Range: (4.5, Range: (4.1, Range:
back, the graphs were found to be easy to understand and helpful for
6.1) 6.3) (0.6,5.6)
ANOVA: ANOVA:
improving posture. This enables the user to become more aware of their
F = 25.66 F = 13.61 posture and ultimately improve their well-being.
P value = 1x P value =
10− 6 0.0004 5. Conclusions
Lumbar Mean: − 8.3 Mean: − 7.6 Mean: − 4.1
Range: (− 10.0, Range:(− 8.7, Range: (− 7.6-
− 6.6) − 6.4) ,0.6) This study provided the algorithms and methodology to detect cor­
ANOVA: ANOVA: rect sitting postures in long-term seating periods. The study demon­
F = 21.88 F = 26.73 strated that transfer learning could be used in limited datasets provided
P value = 9x P value = 1x by different sensor map characteristics. The interface pressure mapping
10− 6 10− 6
system along with the CNN-based recognition system, showed great
potential for obtaining objective data about sitting posture.
visually distinguishable postures, including sitting while leaning left or The results of this study demonstrated the robustness of the posture
right, sitting on the edge of a chair, and crossing feet. All proposed classification methodology and the feasibility and versatility of the al­
postures seem to be visually identified. Their study used a force-sensitive gorithms to be used with different pressure distribution systems through
resistor (FSR) inserted in the backrest, as well as the chair’s seat pan. the implementation of Transfer Learning (0.854 ± 0.007). However, it is
Their main goal was to compare different classification methods, such as relevant to highlight that the sample size and sensor resolution in pre-
MLP, CNN, and Spike neural networks, with and without preprocessing training could have an impact on the final performance, as demon­
methods of complex patterns associated with the FSR maps, not evaluate strated in this study. CNN trained with the BodiTrak® dataset presented
their prevalence or type of sitting posture. This study inspires our pro­ an accuracy of 0.984 ± 0.005 because the sensor resolution was higher
posal using Convolutional Neural Networks. than the XSENSOR® dataset with an accuracy of 0.847 ± 0.011 when
they were stand-alone.

8
I. Morales-Nolasco et al. Biomedical Signal Processing and Control 95 (2024) 106306

Table 7
Comparison of precision, accuracy, and recall of different classification methods to detect different sitting postures.
Author Sensor type and sensor position Participants Detected postures Classification method Evaluation
(#) condition

J. Wang et al. FSR Matrix (9x9) in the seat and FSR 19 volunteers (1) seating Upright (SU) SNN + LR (logistic Precision: 0.88
2020 [6] Matrix (10x9) in the backrest (4) Leaning right (LR), left (LL), regression) Recall: 0.86
forward (LF) and back (LB)
(10 combination)
SU & left leg crossed over the right leg
(LC)
SU & rigth leg crossed over the left leg
(RC)
LC & LB
RC & LB
SU& left ankle resting on the leg (LA)
SU& right ankle resting on the leg(RA)
LA & LB &
RA & LB
Slouchy back down
Sitting on the leading edge
Kim, Y. M et al. FSR Matrix (8x8) in the seat 10 volunteers (1)sitting straight CNN: LeNet-5 Accuracy:
2018 [7] (2) lean left and right 0.942
(1) sitting at the front of the chair
(1) sitting crossed-legged
Ma et al. 7 FSR in the seat and 12 volunteers User seated correctly on the MLP Accuracy:
2017 [5] 5 FSR in the backrest wheelchair 0.955
Lean Left (LL) Precision:0.926
Lean Right (LR) Recall: 0.926
Lean Forward (LF)
Lean Backward (LB)
This study XSENSOR® (26x26) 10 volunteers Long lordosis CNN + Transfer Accuracy:
BodiTrak® (32x32) (XSENSOR dataset) Short lordosis Learning 0.854
Both in the seat. 12 volunteers (Boditrak Flat Precision:0.856
dataset) Recall: 0.854

The approach utilized in this study has the potential to facilitate analysis, Conceptualization. Laura Garay-Jiménez: Writing – review &
additional research on sitting posture since it is designed to be versatile editing, Visualization, Validation, Supervision, Resources, Project
and can easily be adapted to various pressure mapping systems without administration, Methodology, Investigation, Funding acquisition,
altering the methodology. Moreover, this methodology can serve as a Conceptualization.
foundation for further exploration into the dynamics of long-term
sitting. The insights gained from such research could be invaluable in
helping individuals maintain good posture, leading to improved health Declaration of competing interest
and overall well-being.
The authors declare that they have no known competing financial
CRediT authorship contribution statement interests or personal relationships that could have appeared to influence
the work reported in this paper.
Isaac Morales-Nolasco: Writing – original draft, Visualization,
Software, Methodology, Investigation, Formal analysis, Data curation, Data availability
Conceptualization. Sandra Arias-Guzman: Writing – review & editing,
Writing – original draft, Validation, Supervision, Methodology, Formal Data will be made available on request.

9
I. Morales-Nolasco et al. Biomedical Signal Processing and Control 95 (2024) 106306

Annex.

.
Annex1. Screenshot of the feedback dashboard.

References [10] K. O’Sullivan, L. O’Sullivan, P. O’Sullivan, W. Dankaerts, Investigating the effect of

real-time spinal postural biofeedback on seated discomfort in people with non-
specific chronic low back pain, Ergonomics 56 (2013) 1315–1325.
[1] A. Gonzales, J.-H. Lin, J.S. Cha, Physical activity changes among office workers
[11] S.B. Mahdavi, R. Riahi, B. Vahdatpour, R. Kelishadi, Association between sedentary
during the COVID-19 pandemic lockdown and the agreement between objective
behavior and low back pain; A systematic review and meta-analysis, Health
and subjective physical activity metrics, Applied Ergonomics 105 (2022) 103845.
Promotion Perspectives 11 (2021) 393.
[2] A. Moretti, F. Menna, M. Aulicino, M. Paoletta, S. Liguori, G. Iolascon,
[12] C. Bontrup, W.R. Taylor, M. Fliesser, R. Visscher, T. Green, P.-M. Wippert, et al.,
Characterization of home working population during COVID-19 emergency: a
Low back pain and its relationship with sitting behaviour among sedentary office
cross-sectional analysis, International Journal of Environmental Research and
workers, Applied Ergonomics 81 (2019) 102894.
Public Health 17 (2020) 6284.
[13] F. Arippa, A. Nguyen, M. Pau, C. Harris-Adamson, Postural strategies among office
[3] G.F. Papalia, G. Petrucci, F. Russo, L. Ambrosio, G. Vadalà, S. Iavicoli, et al.,
workers during a prolonged sitting bout, Applied Ergonomics 102 (2022) 103723.
COVID-19 pandemic increases the impact of low back pain: a systematic review
[14] X. Zhang, J. Fan, T. Peng, P. Zheng, X. Zhang, R. Tang, Multimodal data-based deep
and metanalysis, International Journal of Environmental Research and Public
learning model for sitting posture recognition toward office workers’ health
Health 19 (2022) 4599.
promotion, Sensors and Actuators a: Physical 350 (2023) 114150.
[4] W. Albarracín, F. Jurado, L. Chuquitarco, I.C. Proaño, Análisis y corrección de la
[15] L. Alzubaidi, J. Zhang, A.J. Humaidi, A. Al-Dujaili, Y. Duan, O. Al-Shamma, et al.,
postura empleando la tecnología de la cámara KINECT, Revista Ibérica De Sistemas
Review of deep learning: Concepts, CNN architectures, challenges, applications,
e Tecnologias De Informação (2020) 227–236.
future directions, Journal of Big Data 8 (2021) 1–74.
[5] C. Ma, W. Li, R. Gravina, G. Fortino, Posture detection based on smart cushion for
[16] Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, “A survey of convolutional neural
wheelchair users, Sensors 17 (2017) 719.
networks: analysis, applications, and prospects,” IEEE transactions on neural
[6] J. Wang, B. Hafidh, H. Dong, A. El Saddik, Sitting posture recognition using a
networks and learning systems, 2021.
spiking neural network, IEEE Sensors Journal 21 (2020) 1779–1786.
[17] K. Ho, F. Kendall, D. Boynton, Posture and pain, Williams and Wilkins, Baltimore,
[7] Y.M. Kim, Y. Son, W. Kim, B. Jin, M.H. Yun, Classification of children’s sitting
1952, pp. 95–97.
postures using machine learning algorithms, Applied Sciences 8 (2018) 1280.
[8] R. Zemp, W.R. Taylor, S. Lorenzetti, Seat pan and backrest pressure distribution
while sitting in office chairs, Applied Ergonomics 53 (2016) 1–9.
[9] R. Baker, P. Coenen, E. Howie, A. Williamson, L. Straker, The short term
musculoskeletal and cognitive effects of prolonged sitting during office computer
work, International Journal of Environmental Research and Public Health 15
(2018) 1678.

10
I. Morales-Nolasco et al. Biomedical Signal Processing and Control 95 (2024) 106306

[18] A.P. Claus, J.A. Hides, G.L. Moseley, P.W. Hodges, Is ‘ideal’sitting posture real?: [20] Z. Fan, X. Hu, W.-M. Chen, D.-W. Zhang, X. Ma, A deep learning based 2-dimen­
Measurement of spinal curves in four sitting postures, Manual Therapy 14 (2009) sional hip pressure signals analysis method for sitting posture recognition,
404–408. Biomedical Signal Processing and Control 73 (2022) 103432.
[19] S. Garg, A. Saxena, R. Gupta, “Yoga pose classification: a CNN and MediaPipe
inspired deep learning approach for real-world application,” Journal of Ambient
Intelligence and Humanized, Computing (2022) 1–12.

11