Non Contact Stress Assessment Based on Deep Tabular Method
Urmila
a
and Avantika Singh
b
Dr. Shyama Prasad Mukherjee International Institute of Technology, Naya Raipur, Raipur, Chhattisgarh, India
Keywords:
Stress Detection, Remote Photoplethysmography (rPPG), Physiological Features.
Abstract:
In today’s competitive world, stress is the major factor that influences human health negatively. In the long
term, stress can lead to serious health problems such as diabetes, depression, anxiety, and various heart dis-
eases. Thus, timely stress recognition is important for efficient stress management. Currently, for stress as-
sessment various wearable devices are used to capture physiological signals. However, these devices although
accurate are cost-sensitive and requires direct physical contact which may lead to discomfort in long run. In
this work we have introduced a tabular based deep learning architecture for detecting stress by analyzing phys-
iological features. The architecture extracts physiological features from remote photoplethysmography (rPPG)
signals computed from facial videos. The proposed architecture is validated on publicly available UBFC-Phys
dataset for two sets of experiments (i) Stress task classification and (ii) Multi-level stress classification. For
both set of experiments the proposed methodology outperforms the current state-of-art method. The code is
available at https://github.com/Heeya2205/Deep_tabular_methods.
1 INTRODUCTION
Stress is a physical, emotional, or mental response
to external pressures or demands, which can arise
from various factors. Life situations like work chal-
lenges, family issues, and environmental changes of-
ten arouse stress conditions in an individual. It is a
natural response of the body to perceived threats and
can manifest as tension, anxiety or other psycholog-
ical or physical reactions (Giannakakis et al., 2019).
Stress causes the body to release hormones like corti-
sol and adrenaline, which raise your heart rate, blood
pressure, and blood sugar to help you deal with chal-
lenges. However, if stress lasts a long time, it can lead
to serious health problems like heart disease, diabetes,
anxiety, and depression (Fink, 2010).
In human body autonomic nervous system (ANS)
is responsible for controlling the number of heart-
beats per minute (Barazi et al., 2021). ANS can be
further divided into two sub-parts: (a) Sympathetic
nervous system, (b) Parasympathetic nervous system.
Sympathetic nervous system is active when an indi-
vidual encounters situation like stress, anxiety, fear or
undergoes laborious exercises, as a result it increases
the heart rate. On the other hand parasympathetic ner-
vous system is active when an individual is in calm
a
https://orcid.org/0009-0003-8642-8525
b
https://orcid.org/0000-0002-2606-5959
state of mind particularly when an individual feels
compassion or love and thus, it decreases the heart-
rate. As mentioned above heart rate is directly influ-
enced by ANS which further depends upon our state-
of-mind. Thus in this work we aim to estimate stress
based on heart-rate (HR) and its related factors like
Heart rate variability (HRV), Peak detection etc.
Traditionally, heart rate can be measured by var-
ious methods like electrocardiography (ECG), elec-
tromyography (EMG), and photoplethysmography
(PPG). However, these methods although accurate
are cost-sensitive and requires direct physical contact.
Henceforth, in this work we explore, non-invasive
method of heart-rate estimation by utilizing remote
photoplethysmography (rPPG) signals.
Here, in this work for stress assessment, firstly
rPPG signals are extracted from recorded video
frames. Later, physiological features like HR, HRV
etc are calculated from observed rPPG signals. Fur-
ther, the extracted physiological features are encoded
for learning discriminative stress levels by employ-
ing deep tabular data learning architecture (Arik and
Pfister, 2021) based on attention mechanism. Main
contributions of this work are as follows:
• Efficient Tabular Learning with TabNet: The
proposed method harnesses TabNet network
(Arik and Pfister, 2021) as the foundational back-
bone to directly process physiological features ex-
Urmila, and Singh, A.
Non Contact Stress Assessment Based on Deep Tabular Method.
DOI: 10.5220/0013150300003905
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 14th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2025), pages 597-604
ISBN: 978-989-758-730-6; ISSN: 2184-4313
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
597

tracted from rPPG signals, incorporating built-
in feature prioritization and sparsity mechanisms.
This eliminates the need for extensive preprocess-
ing while enhancing model interpretability and
performance.
• Improved Generalization Across Classification
Tasks: The method demonstrates lower variance
in accuracy across binary and multi-class stress
classification tasks, highlighting superior gener-
alization compared to existing approaches, which
often exhibit significant performance variability.
• Robust Feature Extraction: To effectively man-
age the dynamic motion of faces across video
frames and improve the precision of facial re-
gion extraction, the proposed method integrates
the Haar Cascade (Choi et al., ), and Mediapipe
library (Lugaresi et al., 2019).This combination
facilitates robust face detection and accurate lo-
calization of facial landmarks.
2 RELATED WORK
Several works have been reported in the literature for
rPPG signal analysis (Das et al., 2023) (Speth et al.,
2024) but the work done in the field of stress analy-
sis using rPPG signal is still limited. In one of the
notable work (Ziaratnia et al., 2024) author’s pro-
posed deep learning-based method that utilize Com-
pact Convolutional Transformers (CCT) for feature
extraction and Long Short-Term Memory (LSTM) for
temporal pattern recognition. This study presents a
non-contact approach for recognizing stress by uti-
lizing rPPG signals derived from RGB facial videos.
In another work (Xu et al., 2024) a multi-task atten-
tional convolutional neural network (MTASR) is de-
ployed, which integrates peak detection and heart rate
estimation to enhance stress recognition. By employ-
ing peak detection as a physiological parameter, the
researchers trained their network to identify more re-
liable physiological indicators. In another work au-
thors (Casado et al., 2023) recognized depression
based on a pipeline that extract rPPG signals in a
full unsupervised manner, and calculate 60 statisti-
cal, geometrical, and physiological features. These
extracted features are further used to train several
machine learning regressors to recognize different
level of depressions. In (Ntalampiras, 2023) authors
proposed a non-intrusive, low-cost, and automatic
stress monitoring framework. This framework ex-
tracts multi-domain speech features to reveal comple-
mentary stress-related characteristics. In (Pan et al.,
2024) a deep network for stress assessment is pro-
posed that focuses on facial features such as expres-
sions, movements, and specific areas like the eyes,
nasolabial folds, and jaw, which are related to depres-
sion. The proposed model emphasizes dynamic facial
features through its attention mechanism.
3 METHODOLOGY
This section conceptualizes our proposed approach.
Figure 1 gives a generic overview of the proposed sys-
tem which consists of mainly five parts: (i) ROI selec-
tion and facial landmark detection (ii) Color space ex-
traction from cropped facial regions (iii) rPPG signal
extraction (iv) Physiological features extraction from
rPPG signals (v) Tabular learning on physiological
features.
3.1 ROI Selection and Facial Landmark
Detection
In the pre-processing phase, video frames are ex-
tracted from facial video sequences. Specifically, we
utilize a frame rate of 30 fps for temporal segmen-
tation of the video. Each extracted frame is main-
tained at a resolution of 513 × 513 pixels. Following
the frame extraction, the next step involves generat-
ing a bounding box around the facial region of inter-
est. To achieve this, we implement the Haar Cascade
algorithm (Choi et al., ), which operates by detecting
facial features through the application of rectangular
filters that compute the intensity differences between
adjacent regions in the image. Once the bounding
box is delineated, the subsequent step is the detection
of facial landmarks. For this, we employ the Medi-
apipe Face Mesh (Lugaresi et al., 2019) framework,
which facilitates the identification of 478 facial land-
marks (Ziaratnia et al., 2024).
By integrating both techniques in our proposed
method, we address the challenges inherent in extract-
ing the facial region of interest. Figures 2a and 2b
demonstrate these challenges and highlight the im-
provements achieved through our approach.
Furthermore, convex hull based masking tech-
nique has been employed to isolate and emphasize the
facial region. Mathematically, the convex hull is the
smallest convex shape that can enclose all the points
representing the facial landmarks. For this we have
first created a binary mask with the same resolution
as our extracted video frame resolution (513 × 513).
Formally, binary mask is described as:
mask(x, y) =
(
255 if (x, y) ∈ convex hull
0 otherwise
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
598

Figure 1: Proposed architecture comprising of five main parts: (i) ROI selection and facial landmark detection (ii) Color space
extraction (iii) rPPG signal extraction (iv) Physiological features extraction (v) Tabular learning on extracted physiological
features.
where (x, y) denotes the pixel coordinates. The last
step is to generate cropped facial region, which is gen-
erated as:
F
Cropped
= F
Mediapipe
∧ mask
where, F
Mediapipe
is the output representing 478 facial
landmarks. Here, the bitwise AND operation is used
to retain only the pixel values of the facial region that
fall within the convex hull, effectively setting all non-
facial pixels to black. This selective masking tech-
nique highlights the facial features by suppressing ir-
relevant background elements.
3.2 Color Space Extraction from
Cropped Facial Regions
Inspired by the prior study (Kim et al., 2021) we apply
YCgCr color space (F
YCgCr
) from F
Cropped
(cropped
facial region). By successfully decoupling the lu-
minance component (Y) from the chrominance com-
ponents (Cg) and (Cr), this color space facilitates
the identification of subtle color changes that signify
physiological conditions. By concentrating on the lu-
minance channel and working with various skin tones,
this color space lessens lighting variations (Panigrahi
and Sharma, 2022).
3.3 rPPG Signal Extraction
For extracting rPPG signals from F
YCgCr
we have
utilized state-of-the-art POS (Plane-Orthogonal-to-
Skin) (Wang et al., 2016) algorithm. Algorithm 1 il-
lustrates POS detailed steps.
Inspired by the prior study (Xu et al., 2024) to
normalize extracted rPPG signals (S
pos
) we have fil-
tered it with Butterworth filter (Selesnick and Burrus,
(a)
(b)
Figure 2: Preprocessing steps. (a) Without proper prepro-
cessing . (b) After preprocessing. The image is taken from
UBFC-Phys dataset.
1998). While filtering low cut-off frequency was set
as 0.7 Hz and high cut-off frequency was set as 2.5
Hz. Once the signals are normalized, we divide them
into 2-second windows (30 frames per second) with a
10% overlap between segments to account for tempo-
ral differences across frames.
Non Contact Stress Assessment Based on Deep Tabular Method
599

Algorithm 1: rPPG Signal Extraction.
1: Input: F
YCgCr
2: Output: S
pos
(extracted rPPG Signal)
3: Step 1: Compute Color Signals
4: C
1
← Y −Cg ▷ Color Signal 1
5: C
2
← Cr −
Y +Cg
2
▷ Color Signal 2
6: Step 2: Compute Means of Color Signals
7: µ
1
←
1
n
∑
n
i=1
C
1i
▷ Mean of Color Signal 1
8: µ
2
←
1
n
∑
n
i=1
C
2i
▷ Mean of Color Signal 2
9: Step 3: Adjust Color Signals by Subtracting
the Mean
10: S
1
← C
1
− µ
1
▷ Adjusted Signal 1
11: S
2
← C
2
− µ
2
▷ Adjusted Signal 2
12: Step 4: Compute Final rPPG Signal
13: S
pos
← S
1
+ S
2
▷ rPPG Signal
14: Return S
pos
Figure 3: Single window depicting signal peak and interval
detection.
3.4 Physiological Features Extraction
From the normalized rPPG signals (S
npos
) we have
extracted signal peaks as shown in Figure 3. From
the extracted peaks, signal peak intervals are com-
puted as depicted in Figure 3. Later on the basis of
aforementioned signal intervals 7 physiological fea-
tures are computed as follows:
(i) Heart Rate (HR):
HR =
60
mean interval between peaks (seconds)
where intervals are the time differences between con-
secutive S
npos
peaks.
(ii) Heart Rate Variability (HRV):
HRV = Standard deviation of intervals (seconds)
which represents the variation in time between heart-
beats.
(iii) Standard Deviation of Heart Rate
(SD HR):
SD HR = Standard deviation of computed heart rates
(bpm) calculated for each interval.
(iv) Root Mean Square of Successive Differ-
ences (RMSSD):
RMSSD =
s
1
N − 1
N−1
∑
i=1
(interval
i+1
− interval
i
)
2
where N is the number of intervals, indicating short-
term HRV.
(v) Ups Count: In a sympathetic activity in Au-
tonomic nervous system(ANS) this ups count refer to
the number of times S
npos
signal transit from a lower
value to higher value. It indicates the onset of a heart-
beat.
(vi) Downs Count: It is the moment when the
heart finishes pumping blood and start to relax, which
is captured as a downward transition in the S
npos
sig-
nal.
(vii) Cycles: Mathematically, it is defined as the
minimum of Ups Count and Downs Count.
3.5 Tabular Learning on Physiological
Features
The extracted physiological features are numeric val-
ues; thus, we store them in a tabular structure. To
extract meaningful information from it, we have in-
corporated the state-of-the-art tabular learning archi-
tecture, TabNet (Arik and Pfister, 2021). In our pro-
posed methodology, we have extracted discriminative
information from physiological features by utilizing
the TabNet encoder architecture, as depicted in Fig-
ure 4. This tabular-based deep learning architecture
operates through a series of sequential decision steps,
where each step refines its focus based on the infor-
mation processed from the previous step. Each deci-
sion step employs non-linear transformations to refine
the feature representations. Furthermore, it includes a
sparsity regularization mechanism to control the num-
ber of features selected at each step.
As depicted in Figure 4, this architecture mainly
comprises three modules: (i) Feature Transformer,
(ii) Attentive Transformer, and (iii) Feature Selection
Mask. All these are described below:
Feature Transformer: This module is responsi-
ble for transforming the input physiological features
into meaningful representations. These representa-
tions are later divided into two parts using a split
block. The first part, denoted as d[i] (as depicted in
Figure 4), contributes to the current decision output,
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
600

Figure 4: Encoder architecture for extracting discriminative information from physiological features. This architecture op-
erates through a series of sequential decision steps mainly comprising of three parts (i) Feature Transformer (ii) Attentive
Transformer (iii) Feature Selection Mask. Here, a[i] serves as the input for the attention transformer in the next decision step
while d[i], denotes the current decision output.
while the other part, denoted as a[i] (as depicted in
Figure 4), serves as the input for the attentive trans-
former in the next decision step.
Attentive Transformer: This module generates a
trainable mask that identifies the most important fea-
tures for each decision step using a sparsemax func-
tion, which creates a sparse, interpretable feature se-
lection mechanism. A prior scale term is also em-
ployed to regulate how often each feature is selected
across decision steps.
Feature Selection Mask: This module is used for
selecting globally important feature attributions.
To develop a deep tabular learning model capa-
ble of robustly identifying and prioritizing specific
features derived from rPPG signals that play a criti-
cal role in our analysis, we leverage the TabNet ar-
chitecture. This approach eliminates the need for
preprocessing techniques by allowing the model to
adaptively focus on essential features. If a feature is
deemed less supportive at any stage, the model ad-
vances to another level of evaluation to reassess its
importance.
4 EXPERIMENT AND RESULTS
4.1 Dataset Description
To validate the performance of proposed approach
in this work we have worked on UBFC-Phy
dataset (Sabour et al., 2021). This dataset is pub-
licly available and consists of 56 subjects in total.
Out of these 56 subjects 47 subjects are female and
rest are male. This dataset contains facial videos of
enrolled subjects. Each subject underwent three dif-
ferent tasks: resting task (T1), speech task (T2) and
arithmetic task (T3) resulting in generation of 168
videos in total. Furthermore, this dataset is divided
into two groups based on the challenging nature of
the task namely: control group and test group. In
control group subjects faced less challenging task as
compared to subjects in test group.
4.2 Model Validation and Experimental
Setup
For validating our proposed approach we have con-
ducted two set of experimentation as conducted in
previous state of the art work (Ziaratnia et al., 2024)
namely: (i) Stress task classification, (ii) Multi-level
stress classification.
In stress task classification subjects are classified
on the basis of tasks performed. Under this we have
performed both binary (T 1 vs. T 2, T 2 vs. T 3, and T 1
vs. T 3) as well as ternary classification (T1 vs. T 2
vs. T 3). While in case of multi-level stress classifi-
cation subjects are classified on the basis of the chal-
lenging nature of the task. Out of the three tasks T 3
(arithmetic task) is considered as the most demand-
ing task (Ziaratnia et al., 2024). Thus, this set of
experimentation is focused towards classifying stress
into three levels: (i) no stress ( T 1 task for both
control and test group), (ii) low stress ( T 3 control
group), and (iii) high stress ( T 3 test group). For stress
task classification seven-fold cross-validation tech-
nique and for multi-level stress classification stratified
five-fold cross-validation technique has been adopted
as in (Ziaratnia et al., 2024).
All the experiments have been conducted on a sys-
tem with specifications as: Intel(R) Xeon(R) Silver
4114 CPU, 64 GB of RAM, 4 TB SSD, NVIDIA
GeForce GTX 1080 Ti GPU. For the software envi-
ronment, we have utilized libraries such as OpenCV
(Mordvintsev and Abid, 2017), PyTorch (Imambi
Non Contact Stress Assessment Based on Deep Tabular Method
601

et al., 2021), and Mediapipie (Lugaresi et al., 2019).
Table 1: Stress task 7-fold cross-validation classification
results. Binary classification T1 (resting task) vs. T2
(speech task).
K-folds Accuracy Precision Recall F1 score
Fold-1 0.750 0.833 0.625 0.714
Fold-2 0.681 0.615 1.000 0.761
Fold-3 0.812 0.727 1.000 0.842
Fold-4 1.000 1.000 1.000 1.000
Fold-5 0.937 0.888 1.000 0.941
Fold-6 0.937 0.888 1.000 0.941
Fold-7 0.875 1.000 0.750 0.857
7-Fold Mean 0.857 0.850 0.9107 0.865
Table 2: Stress task 7-fold cross-validation classification re-
sults. Binary classification T1 (resting task) vs. T3 (arith-
metic task).
K-folds Accuracy Precision Recall F1 score
Fold-1 0.812 0.857 0.750 0.800
Fold-2 1.000 1.000 1.000 1.000
Fold-3 0.937 1.000 0.875 0.933
Fold-4 0.930 1.000 0.875 0.933
Fold-5 1.000 1.000 1.000 1.000
Fold-6 1.000 1.000 1.000 1.000
Fold-7 0.937 0.888 1.000 0.941
7-Fold Mean 0.946 0.963 0.928 0.944
Table 3: Stress task 7-fold cross-validation classification re-
sults. Binary classification T2 (speech task) vs. T3 (arith-
metic task).
K-folds Accuracy Precision Recall F1 score
Fold-1 0.750 0.833 0.625 0.714
Fold-2 0.875 0.875 0.875 0.875
Fold-3 0.812 0.857 0.750 0.800
Fold-4 0.875 1.000 0.750 0.857
Fold-5 0.812 0.857 0.750 0.800
Fold-6 0.875 1.000 0.750 0.857
Fold-7 0.937 0.888 1.000 0.941
7-Fold Mean 0.848 0.901 0.785 0.835
Table 4: Stress task 7-fold cross-validation classification
results. Ternary classification T1 (resting task) vs. T2
(speech task) vs. T3 (arithmetic task).
K-folds Accuracy Precision Recall F1 score
Fold-1 0.750 0.757 0.750 0.752
Fold-2 0.750 0.795 0.750 0.752
Fold-3 0.875 0.891 0.875 0.873
Fold-4 0.916 0.933 0.916 0.918
Fold-5 0.833 0.838 0.833 0.829
Fold-6 0.916 0.933 0.916 0.918
Fold-7 0.916 0.933 0.916 0.918
7-Fold Mean 0.851 0.869 0.851 0.851
4.3 Performance Metrics
The proposed framework’s performance is evaluated
using common classification evaluation criteria, in-
cluding accuracy, precision, recall, and F1 score. Ac-
curacy is calculated as the ratio of correctly predicted
instances to total instances, while precision is the ratio
Table 5: Multi-Level stress 5-fold cross-validation classifi-
cation results. Ternary classification: no stress (T1 task
for both control and test group), low stress (T3 control
group) and high stress (T3 test group).
K-folds Accuracy Precision Recall F1 score
Fold-1 0.826 0.896 0.826 0.816
Fold-2 0.869 0.864 0.869 0.864
Fold-3 0.863 0.877 0.863 0.857
Fold-4 0.954 0.961 0.954 0.953
Fold-5 0.863 0.865 0.863 0.861
5-Fold Mean 0.875 0.882 0.875 0.870
Table 6: Multi-Level stress 5-fold cross-validation classi-
fication results. Mean value (± standard deviations) for
three class classification individually: no stress (T1 task
for both control and test group), low stress (T3 control
group) and high stress (T3 test group) .
Class Accuracy Precision Recall F1 score
No Stress 0.981(±0.040) 0.910(±0.092) 0.981(±0.040) 0.943(±0.061)
Low Stress 0.866(±0.139) 0.812(±0.059) 0.866(±0.139) 0.835(±0.088)
High Stress 0.653(±0.086) 0.910(±0.124) 0.653(±0.086) 0.756(±0.081)
of accurately predicted positive cases to all instances
predicted as positive. Recall is the proportion of accu-
rately predicted positive instances relative to all actual
positive instances, including those incorrectly classi-
fied as negative. F1 Score is calculated as the har-
monic mean of precision and recall.
4.4 Stress Task Classification
Performance
As mentioned in section 4.2 for stress task classifi-
cation experimentation we have computed results for
both binary as well as ternary classification. Table 1,
Table 2, and Table 3 illustrates binary stress classi-
fication results for T1 (resting task) vs. T2 (speech
task), T1 (resting task) vs. T3 (arithmetic task), and
T2 (speech task) vs. T3 (arithmetic task) tasks respec-
tively. Table 4 illustrates ternary stress task classifica-
tion results for T1 (resting task) vs. T2 (speech task)
vs. T3 (arithmetic task). Major observations from Ta-
ble 1, Table 2, Table 3, and Table 4 are as follows:
• Highest accuracy is observed while differentiat-
ing subjects undergoing resting task versus arith-
metic task as depicted in Table 2. As stated in
(Ziaratnia et al., 2024) arithmetic task is the most
challenging task and our proposed approach ef-
fectively discriminates it which depicts our model
superiority.
• Lowest recall is observed while differentiating
subjects undergoing speech task versus arithmetic
task as depicted in Table 3. This phenomenon is
quite obvious because speech task that involves
expressing views in front of others and solving
arithmetic problems both involved inducing stress
in individuals.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
602

Table 7: Experimental results comparing our approach to other cutting-edge techniques for stress task categorisation on the
UBFC-Phys dataset. The best findings are bolded in the table, which displays the mean values (± standard deviations) of the
7 cross-validation. This table’s values are derived from (Ziaratnia et al., 2024).
Methods T1 vs. T2 T1 vs. T3 T1 vs. T2 vs. T3
Accuracy F1 Score Accuracy F1 Score Accuracy F1 Score
MLP (Dolmans et al., 2021) 0.709(±0.061) 0.706(±0.064) 0.599(±0.040) 0.587(±0.037) 0.440(±0.028) 0.434(±0.031)
LIT (Dolmans et al., 2021) 0.701(±0.063) 0.699(±0.066) 0.625(±0.042) 0.622(±0.052) 0.447(±0.027) 0.443(±0.031)
DFAF (Gao et al., 2019) 0.758(±0.035) 0.756(±0.033) 0.689(±0.040) 0.686(±0.037) 0.478(±0.034) 0.477(±0.036)
CAM (Praveen et al., 2021) 0.726(±0.060) 0.722(±0.063) 0.650(±0.052) 0.645(±0.054) 0.494(±0.028) 0.487(±0.033)
MFN (Yu et al., 2021) 0.769(±0.035) 0.763(±0.043) 0.684(±0.050) 0.670(±0.049) 0.501(±0.038) 0.500(±0.036)
BCSA (Zhang et al., 2023) 0.818(±0.063) 0.817(±0.063) 0.723(±0.039) 0.722(±0.039) 0.558(±0.052) 0.552(±0.048)
Multimodal CCT-LSTM (Ziaratnia et al., 2024) 0.981(±0.016) 0.981(±0.016) 0.924(±0.037) 0.924(±0.037) 0.832(±0.058) 0.834(±0.056)
Proposed Approach 0.857(±0.104) 0.865(±0.095) 0.946(±0.062) 0.943(±0.065) 0.851(±0.069) 0.851(±0.070)
• In case of ternary stress task classification task
as depicted in Table 4 the mean accuracy, pre-
cision and recall across 7-folds are almost same.
This phenomenon indicates fairness of our model
which focuses not only on minimizing false neg-
atives but also focuses on minimizing false posi-
tives.
4.5 Multi-Level Stress Classification
Performance
Table 5 and Table 6 depicts results for multi-level
stress classification. According to state-of-the-art re-
search, we employed a 5-fold stratified cross valida-
tion strategy in this experiment because the number
of subjects in the control and test groups was unbal-
anced (Ziaratnia et al., 2024). As depicted in Table 5
the highest accuracy reported across all folds was in
fold-4 while the mean accuracy achieved was 87.5%.
Furthermore, as depicted in Table 6 the highest and
lowest multi-level stress accuracy was achieved in
No-Stress and High-Stress class respectively.
4.6 Comparative Analysis
To validate the efficacy of proposed approach we have
compared our computed results with other state-of-
the-art methods working on same dataset (UBFC-
Phy) as ours. It is evident from Table 7 that, with
one exception (T1 vs. T2), our suggested methodol-
ogy performs better than any previous state-of-the-art
effort. When compared to the most recent state-of-
the-art work (Ziaratnia et al., 2024), the findings for
(i) T1 vs. T3 and (ii) T1 vs. T2 vs. T3 demon-
strate an accuracy improvement of 2.2% and 1.9%,
respectively. For our proposed approach the results
obtained in case of T1 vs. T2 case is less as compared
to the available state-of-the-art approach. It should be
noted that in case of current available state-of-the-art
work (Ziaratnia et al., 2024) there is variation of about
15% in accuracy (as reported in Table 7) while con-
sidering various stress task classification cases. In this
work our objective is not only to develop an approach
that achieves high accuracy but that also generalizes
well to different cases. Henceforth, we have focused
towards reducing the accuracy variation across differ-
ent stress task classification cases. For our proposed
approach the variation in accuracy across different
stress task classification cases was reported around
9% which is approximately 6% less than the current
state-of-the-art work (Ziaratnia et al., 2024).
For multi-level stress classification task we have
compared our proposed approach with two state-of-
the-art approaches (Ziaratnia et al., 2024), (Xu et al.,
2024). In (Ziaratnia et al., 2024) authors have com-
puted results by using 5 fold stratified cross validation
approach for computing results. Using the same strat-
egy in our case we have achieved accuracy of 0.875
with F1-score as 0.870 while in (Ziaratnia et al.,
2024) computed accuracy was 0.805 with F1-score
as 0.803. Clearly, we have achieved an improvement
of 7% in accuracy as compared to (Ziaratnia et al.,
2024). In (Xu et al., 2024), authors used 10-fold
cross-validation for the categorization of low stress
(T2) vs. high stress (T3) and achieved an accuracy
of 83.83% . Following the same strategy, for our pro-
posed architecture we have achieved comparable ac-
curacy of 83.50%.
5 CONCLUSION AND FUTURE
SCOPE
Stress assessment is crucial for applications such as
driver condition monitoring, workplace productivity
analysis, and tailored healthcare. The development
of precise and economical stress task and level clas-
sification techniques is urgently needed. Therefore,
in this work we have presented a deep tabular non-
contact stress measurement technique. The suggested
method concentrates on extracting the physiological
parameters from the rPPG signal and improving ROI
selection while dynamically moving the face inside
the video frame. The suggested architecture per-
Non Contact Stress Assessment Based on Deep Tabular Method
603

formed better than earlier approaches on the UBFC-
Phy dataset in both sets of experiments. Future re-
search may examine the possibilities of combining
speech and eye gaze data with rPPG signals to assess
stress.
REFERENCES
Arik, S.
¨
O. and Pfister, T. (2021). Tabnet: Attentive inter-
pretable tabular learning. In Proceedings of the AAAI
conference on artificial intelligence, volume 35, pages
6679–6687.
Barazi, N., Polidovitch, N., Debi, R., Yakobov, S., Lakin,
R., and Backx, P. H. (2021). Dissecting the roles of
the autonomic nervous system and physical activity
on circadian heart rate fluctuations in mice. Frontiers
in physiology, 12:692247.
Casado, C.
´
A., Ca
˜
nellas, M. L., and L
´
opez, M. B. (2023).
Depression recognition using remote photoplethys-
mography from facial videos. IEEE Transactions on
Affective Computing, 14(4):3305–3316.
Choi, C.-H., Kim, J., Hyun, J., Kim, Y., and Moon, B. Face
detection using haar cascade classifiers based on ver-
tical component calibration.
Das, M., Bhuyan, M. K., and Sharma, L. N. (2023).
Time–frequency learning framework for rppg signal
estimation using scalogram-based feature map of fa-
cial video data. IEEE Transactions on Instrumenta-
tion and Measurement, 72:1–10.
Dolmans, T. C., Poel, M., van’t Klooster, J.-W. J., and
Veldkamp, B. P. (2021). Perceived mental work-
load classification using intermediate fusion multi-
modal deep learning. Frontiers in human neuro-
science, 14:609096.
Fink, G. (2010). Stress: Definition and history. Stress sci-
ence: neuroendocrinology, 3(9):3–14.
Gao, P., Jiang, Z., You, H., Lu, P., Hoi, S. C., Wang, X., and
Li, H. (2019). Dynamic fusion with intra-and inter-
modality attention flow for visual question answer-
ing. In Proceedings of the IEEE/CVF conference on
computer vision and pattern recognition, pages 6639–
6648.
Giannakakis, G., Grigoriadis, D., Giannakaki, K., Simanti-
raki, O., Roniotis, A., and Tsiknakis, M. (2019). Re-
view on psychological stress detection using biosig-
nals. IEEE transactions on affective computing,
13(1):440–460.
Imambi, S., Prakash, K. B., and Kanagachidambaresan, G.
(2021). Pytorch. Programming with TensorFlow: so-
lution for edge computing applications, pages 87–104.
Kim, N. H., Yu, S.-G., Kim, S.-E., and Lee, E. C. (2021).
Non-contact oxygen saturation measurement using
ycgcr color space with an rgb camera. Sensors,
21(18):6120.
Lugaresi, C., Tang, J., Nash, H., McClanahan, C., Uboweja,
E., Hays, M., Zhang, F., Chang, C.-L., Yong, M. G.,
Lee, J., et al. (2019). Mediapipe: A framework
for building perception pipelines. arXiv preprint
arXiv:1906.08172.
Mordvintsev, A. and Abid, K. (2017). Opencv-python tuto-
rials documentation.
Ntalampiras, S. (2023). Model ensemble for predicting
heart and respiration rate from speech. IEEE Internet
Computing, 27(3):15–20.
Pan, Y., Shang, Y., Liu, T., Shao, Z., Guo, G., Ding, H., and
Hu, Q. (2024). Spatial–temporal attention network for
depression recognition from facial videos. Expert Sys-
tems with Applications, 237:121410.
Panigrahi, A. and Sharma, H. (2022). Non-contact hr ex-
traction from different color spaces using rgb camera.
In National Conference on Communications (NCC),
pages 332–337.
Praveen, R. G., Granger, E., and Cardinal, P. (2021). Cross
attentional audio-visual fusion for dimensional emo-
tion recognition. In 16th IEEE International Con-
ference on Automatic Face and Gesture Recognition,
pages 1–8.
Sabour, R. M., Benezeth, Y., De Oliveira, P., Chappe, J., and
Yang, F. (2021). Ubfc-phys: A multimodal database
for psychophysiological studies of social stress. IEEE
Transactions on Affective Computing, 14(1):622–636.
Selesnick, I. W. and Burrus, C. S. (1998). Generalized dig-
ital butterworth filter design. IEEE Transactions on
signal processing, 46(6):1688–1694.
Speth, J., Vance, N., Sporrer, B., Niu, L., Flynn, P., and
Czajka, A. (2024). Mspm: A multi-site physio-
logical monitoring dataset for remote pulse, respira-
tion, and blood pressure estimation. arXiv preprint
arXiv:2402.02224.
Wang, W., Den Brinker, A. C., Stuijk, S., and De Haan,
G. (2016). Algorithmic principles of remote
ppg. IEEE Transactions on Biomedical Engineering,
64(7):1479–1491.
Xu, J., Song, C., Yue, Z., and Ding, S. (2024). Facial video-
based non-contact stress recognition utilizing multi-
task learning with peak attention. IEEE Journal of
Biomedical and Health Informatics.
Yu, H., Vaessen, T., Myin-Germeys, I., and Sano, A. (2021).
Modality fusion network and personalized attention in
momentary stress detection in the wild. In 9th Inter-
national Conference on Affective Computing and In-
telligent Interaction (ACII), pages 1–8.
Zhang, X., Wei, X., Zhou, Z., Zhao, Q., Zhang, S., Yang,
Y., Li, R., and Hu, B. (2023). Dynamic alignment and
fusion of multimodal physiological patterns for stress
recognition. IEEE Transactions on Affective Comput-
ing.
Ziaratnia, S., Laohakangvalvit, T., Sugaya, M., and Srip-
ian, P. (2024). Multimodal deep learning for remote
stress estimation using cct-lstm. In Proceedings of
the IEEE/CVF Winter Conference on Applications of
Computer Vision, pages 8336–8344.
ICPRAM 2025 - 14th International Conference on Pattern Recognition Applications and Methods
604