An Empirical Study Using Machine Learning to Analyze the
Relationship Between Musical Audio Features and Psychological Stress
Harini Anand
a
, Shalini Kammalam Srinivasan
b
, Hasika Venkata Boggarapu, Arti Arya
c
and Richa Sharma
d
Department of Computer Science and Engineering, PES University, Bangalore, India
Keywords:
Music, Psychological Stress, XAI, Classification, SHAP, Integrated Gradients, Feedforward Neural Network,
Audio Techniques, Danceability, Energy, Loudness, Mode, Speechiness, Acousticness, Instrumentalness,
Liveness, Tempo.
Abstract:
Music plays a vital role in regulating emotions and mental well-being, influencing brain function and stress
levels. This study leverages Explainable AI (XAI) techniques, specifically SHapley Additive exPlanations
(SHAP) and Integrated Gradients, to analyze the impact of scientifically backed audio features—such as
Danceability, Energy, Acousticness, etc on stress classification. Using a Feedforward Neural Network, we
achieved a 0.96 accuracy in categorizing music preferences into ”Stressed,” ”Not-stressed,” and ”Borderline”
states. The classifier operates effectively across languages and genres, enhancing its versatility for detecting
Psychological Stress by providing interpretable insights.
1 INTRODUCTION
Music significantly impacts brain function and struc-
ture, influencing areas related to emotion, motivation,
and anticipation (Vuust et al., 2022). It can modulate
heart rate and breathing, thereby affecting our stress
levels and overall mental state. Beyond its role as a
source of entertainment, music is an integral part of
our lives, playing a crucial role in mental health and
well-being by affecting emotions, moods, and other
such psychological states. Research has shown that
music preferences and listening strategies are linked
with the psychological welfare of listeners, as well
as stress and internalized symptomatology. However,
studies examining the time-varying nature of music
consumption in terms of acoustic content and its as-
sociation with users’ well-being, remain scarce. Mu-
sic’s power to shift and regulate mood makes it a use-
ful tool for managing emotions. For instance, dur-
ing periods of stress, individuals often rely on music
to impact their moods and alter affective states (Adi-
asto, 2022). Focusing on predictive accuracy over
model interpretability has a possibility of resulting in
a
https://orcid.org/0009-0002-1477-3197
b
https://orcid.org/0009-0000-0231-356X
c
https://orcid.org/0000-0002-4470-0311
d
https://orcid.org/0000-0002-4539-7051
a gap of transparency in the decision-making process,
which is important in crucial use cases such as health-
care. This underscores the extending need for ex-
plainable AI (XAI) approaches in psychological pre-
diction and diagnosis. The motivation of the research
is to explore the time-varying nature of music con-
sumption and its acoustic content in relation to users’
well-being, also understanding the potential of music
as a tool for emotion regulation and its implications
for mental health. Additionally, the study addresses
the challenge of bridging the gap between predictive
accuracy and interpretability in AI models used for
psychiatric diagnosis and prediction. Initial efforts re-
lied on a genre-based approach but lacked robustness
for non-English songs. By focusing on these univer-
sal audio features, the classifier is designed to operate
effectively across all languages and genres of music,
ensuring a broad, culturally inclusive application.
This research contributes to the field by conduct-
ing a empirical study to analyze the association be-
tween music consumption patterns and psychologi-
cal well-being, developing an explainable AI model
that balances predictive accuracy with interpretabil-
ity for detecting stress levels, and identifying specific
acoustic features of music that are linked to changes
in mood and mental state.
Anand, H., Srinivasan, S. K., Boggarapu, H. V., Arya, A. and Sharma, R.
An Empirical Study Using Machine Learning to Analyze the Relationship Between Musical Audio Features and Psychological Stress.
DOI: 10.5220/0013249900003890
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Conference on Agents and Artificial Intelligence (ICAART 2025) - Volume 3, pages 1031-1039
ISBN: 978-989-758-737-5; ISSN: 2184-433X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
1031

2 RELATED WORK
2.1 Literature Survey
Recent research underscores the significant role of
music in emotional regulation and mental well-being.
According to (Stewart et al., 2019), individuals
with tendencies toward depression often use music
for mood regulation, consciously selecting compo-
sitions that help manage their emotions. Music has
been shown to influence internalized symptomatol-
ogy and depression, indicating a strong connection
between listening strategies and psychological well-
being. This connection is especially pronounced
among young people, for whom music serves as a
critical emotional outlet (McFerran, 2014). These
findings highlight the potential of music as a non-
pharmacological intervention for mental health is-
sues, particularly in the context of emerging adult-
hood, which is characterized by significant life transi-
tions and stressors (Anderson et al., 2003).
Explainable AI (XAI) techniques are pivotal in
providing insights into how machine learning mod-
els make predictions, especially in complex domains
such as healthcare and psychology. These techniques
help decrease the gap between human understand-
ing and model predictions, enhancing trust and inter-
pretability. However, much of this research prioritizes
predictive accuracy over model interpretability, which
can be problematic in healthcare applications where
transparency is crucial (Lin, 2011).
Neuroscientific and clinical studies provide com-
pelling evidence for the therapeutic potential of mu-
sic. For example, (Juslin and Sloboda, 2010) dis-
cusses how music therapy can be an effective treat-
ment for different mental health conditions, includ-
ing depression, autism, schizophrenia, and dementia.
The therapeutic advantages of music are due to its
ability to affect tough neuro-biological processes in
the brain, thereby modulating emotions and allevi-
ating anxiety. Furthermore, the use of Music Infor-
mation Retrieval (MIR) algorithms in analyzing au-
dio features like tempo and rhythm offers a promis-
ing approach to understanding how different types
of music can impact stress levels and overall men-
tal health. This integration of technology and neu-
roscience paves the way for innovative interventions
that harness the power of music to improve psycho-
logical well-being (Saarikallio, 2007).
The study by (Gujar, 2023) investigates the corre-
lation between music and mood using machine learn-
ing techniques and reveals that certain musical ele-
ments like key, tempo, and mode are linked to spe-
cific moods, providing valuable insights for enhanc-
ing mental health through music. The study by
(Ahuja, 2019) analyses the mental stress among col-
lege students using machine learning algorithms to
evaluate the impact of exam pressure and internet us-
age on their well-being.
The study by (Erbay Dalli, 2023) elaborates on
how multiple-session music interventions can be em-
ployed as a nursing strategy to manage anxiety levels
in ICU patients. Hearing music on a regular basis,
prevents CUMS-induced oxidative stress in the hip-
pocampus, prefrontal cortex, and serum of mice. This
paper by (Gu, 2023) indicates that in mouse experi-
ments, hearing music reduces stress-induced anxiety
and depression-like behaviors. Music has the abil-
ity to regain preventing oxidative stress, neurotrophic
factor deficits, hypothalamus-pituitary-adrenal axis
homeostasis, and inflammation.
After an elaborate literature survey, as highlighted
in the previous section, the following research gaps
underscore the pressing need for further exploration
into the intricate correlation between audio features
in music and psychological stress:
• Existing studies focus primarily on genre based
stress detection and lack robustness for regional
and foreign songs.
• Insufficient exploration of how specific audio fea-
tures, such as tempo, liveness, and danceability,
influence stress psychopathology.
• A lack of research applying XAI (Explainable AI)
techniques in healthcare diagnosis and prediction,
which are needed to ensure transparency and in-
terpretability in AI models.
2.2 Our Contributions
Since the research gaps point to a substantial opportu-
nity to explore the potential for music, through these
individual audio features, we aim to implement cer-
tain key techniques, mentioned below that shall serve
as a preventive and therapeutic tool for reducing psy-
chological stress.
• Building a high performing classifier using Ma-
chine Learning and train it specifically on Music
Audio Features to detect Psychological Stress.
• Incorporate Explainable AI Techniques to demon-
strate interpretability.
• Validate the classifier on unseen data using differ-
ent types of linguistic music datasets.
3 PROPOSED APPROACH
The proposed approach is as depicted in Figure 1.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
1032

Figure 1: Proposed Approach to Establish a Relation between Music and Psychological Stress.
3.1 Dataset Description
The dataset is a comprehensive collection of 15,150
classic hits from 3,083 artists, spanning a century
of music history from 1923 to 2023 sourced from
Kaggle. This diverse dataset is divided into 19 dis-
tinct genres, showcasing the evolution of popular mu-
sic across different eras and styles. Each track in
the dataset is of songs from Spotify with audio fea-
tures such as Danceability, Energy, Loudness, Mode,
Speechiness, Acousticness, Instrumentalness, Live-
ness, and Tempo, offering detailed insights into the
acoustic properties, rhythm, tempo, and other musi-
cal characteristics of each track in the dataset.
3.2 Threshold Validation
To utilize the audio features in the dataset for our in-
tended objective of exploring the influence of music
on psychological stress, several clinical and research
based studies were used, where researchers explorato-
rily inferenced the connection between music features
and users’ preferred recovery-related feelings while
hearing and after hearing to self-selected music.
In one of their studies (Adiasto et al., 2023a), 470
participants took a survey where the users indicated
the type of music she or he would pick to de-stress
from a theoretically stressed situation. Using Data
analysis techniques such as split-sample procedures,
a k-medoid cluster analysis was held to identify au-
dio feature commonalities between songs that were
self-selected by the users. In addition to this, several
regression analyses were also done to cross check and
deduce the connection of musical audio features and
preferred recovery psychological states and emotions.
Analyses in (Adiasto et al., 2023b) revealed the
role played by positive emotions in the stress reduc-
tion process, it is safe to conclude that music’s au-
dio features have recovery potential under the condi-
tion that it draws out a favourable emotional response.
Based on this base theory, the expansive domain of
music emotion recognition (MER) hints which songs
are the most impactful for stress reduction and recov-
ery.
Studies on music emotion recognition (Duman
et al., 2022) have explored various combinations of
musical audio features, including tempo (the speed
of a song), pitch (the frequency of a particular note
or sound), and timbre (the overall quality or color of
a song). These features are analyzed in relation to
the valence (i.e., the positivity of an emotion, with
higher valence values indicating more positive emo-
tions) and arousal components of self-reported musi-
cal emotions. Compared to the baseline, participants’
preferred music choices to reduce stress, indicated
significantly higher levels of energy, danceability, and
acousticness.
In addition, tempo has been found to positively
correlate with both emotional valence and arousal,
demonstrating that songs with faster tempos tend to
be associated with emotions that are more positive
and have higher levels of arousal.
Table 1 lists the scientifically validated thresholds
using various audio features of Spotify songs, which
we make use of in our proposed approach.
3.3 Dataset Pre-Processing &
Augmentation
Valence as a feature was dropped due to low variabil-
ity across the dataset (i.e., most of its values were sim-
ilar), as it would contribute little information to the
model.
To handle class imbalance, rule-based synthetic
data generation (He et al., 2008) is used, where
we utilize a hybrid technique approach combining
boundary-based sampling (Liu et al., 2007) with
Gaussian noise injection (Zhang et al., 2018). This
is the most suitable approach when the dataset has
predefined class boundaries. The dataset distribution
after data augmentation is as seen in Table 2.
An Empirical Study Using Machine Learning to Analyze the Relationship Between Musical Audio Features and Psychological Stress
1033

Table 1: Threshold Distribution of Stress Level Class Labels Based on Music Audio Features.
Feature Not-Stressed (0) Stressed (1) Borderline (2)
Danceability 0.5 – 0.998 0 – 0.38 0.39 – 0.49
Energy 0.6 – 0.99 0.3 – 0.59 0 – 0.299
Mode 1 0 0
Speechiness 0.01 – 0.07 0.31 – 1 0.071 – 0.3
Acousticness 0.35 – 1 0 – 0.3 0.31 – 0.34
Instrumentalness 0 – 0.18 0.21 – 1 0.19 – 0.2
Liveness 0 – 0.1 0.2 – 1 0.11 – 0.2
Tempo (BPM) 60 – 130 181 – 220 130 – 180
Table 2: Class Distribution in Full, Training, and Testing
Datasets.
Dataset Stressed Borderline Not Stressed
Full Dataset 5602 5935 5848
Training Set 4482 4748 4678
Testing Set 1120 1187 1170
3.4 Model Architecture & Training
3.4.1 Why Feed-Forward Neural Network
(FFNN)
A Feed-Forward Neural Network (FFNN) was se-
lected for this classification problem due to its ef-
ficiency in modeling structured, non-sequential data
and capturing complex patterns. FFNNs are partic-
ularly effective in handling data like audio features,
where each feature represents a specific aspect of the
input without any inherent temporal or sequential de-
pendencies.
The FFNN architecture also allows for flexibil-
ity with activation functions and regularization tech-
niques, making it ideal for addressing potential over-
fitting, and ensuring the model generalizes well to un-
seen data. By using FFNN, we benefit from its abil-
ity to map high-dimensional input data to meaningful
output probabilities for each class. Other State of the
Art Models did not outperform the chosen architec-
ture of choice for our use case.
3.4.2 Model Architecture
The model was designed using three dense layers to
learn the complex relationships among the input fea-
tures. Each layer incorporates Dropout and Batch
Normalization techniques to eliminate overfitting and
improve generalization. Dropout with a rate of 0.7
was applied after each dense layer to randomly set a
section of the input units to 0 during training, effec-
tively preventing the model from relying too heavily
on any one specific feature. Batch normalization was
employed to normalize the activations in each layer,
stabilizing the learning process and improving con-
vergence speed. Softplus activation functions were
used in the intermediate layers due to their smooth
gradient properties, which facilitate stable training.
Finally, the output layer employs the softmax activa-
tion function to produce probabilities for each of the
three classes, making it suitable for a multi-class clas-
sification problem.
3.4.3 Training Configuration
Our model was compiled using the Adam optimizer
with a learning rate of 0.001, a commonly used opti-
mizer known for its efficient handling of sparse gra-
dients and adaptive learning rates. The categorical
cross-entropy loss function was used, as it is the stan-
dard for multi-class classification problems.
To further combat overfitting and ensure model
generalization, we applied early stopping, which
monitors the validation loss and halts training if it
does not improve for five consecutive epochs. This
prevents unnecessary training and helps preserve the
best weights. Additionally, the ReduceLROnPlateau
callback was implemented to reduce the learning rate
by a factor of 0.5 if the validation loss stagnates for
two epochs, promoting better convergence. These
training techniques, in combination with the model
architecture, lead to improved performance on the test
set while minimizing overfitting.
4 RESULTS & DISCUSSIONS
4.1 Model Performance
The feedforward neural network achieved a training
accuracy of 96% with a corresponding training loss of
0.2850 and demonstrated strong performance during
training and validation phases. Figure 2 depicts Train-
ing and Validation loss curves demonstrating model
performance, and Figure 3 suggests Training and Val-
idation accuracy curves demonstrating model perfor-
mance, thus proving it not only fits the training data
well but also generalizes excellently to unseen data.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
1034

Figure 2: Training and validation loss curves demonstrating
model performance.
Figure 3: Training and validation accuracy curves demon-
strating model performance.
4.2 Explainable AI (XAI)
Explainable AI (XAI) is very essential for interpret-
ing the predictions given by a feedforward neural net-
works (FFNNs) and addressing their ”black box” na-
ture. This helps the stakeholders to better understand,
trust and validate model predictions especially in do-
mains like healthcare.
SHapley Additive exPlanations (SHAP) and
Integrated Gradients are the two XAI techniques used
in our study to analyze the results of FFNN. These
methods help to interpret the influence of each feature
on the model’s output thereby providing insights into
the internal workings of the model.
SHapley Additive exPlanations (SHAP). SHAP
(Lundberg, 2017) assigns an importance value called
shapley value to each feature for individual predic-
tions. This importance value represents both the mag-
nitude and direction of a feature’s impact thus help-
ing us to interpret feature influence at a more gran-
ular and instance-specific level. Taking all possible
feature combinations into consideration, SHAP cal-
culates feature importance for each input feature.
The SHapley value for feature i, φ
i
( f ), is defined
as in Equation (1):
φ
i
( f ) =
∑
F⊆N\{i}
|F|! · (|N| − |F| − 1)!
|N|!
[ f (F ∪{i}) − f (F)]
(1)
Here:
- N is the set of all features.
- F is a coalition of features excluding feature i.
- |F| is the number of features in coalition F.
- f (F) is the model’s prediction for the instance with
only features in set F.
- f (F ∪ {i}) is the model’s prediction for the instance
with features in set F ∪ {i}.
- φ
i
( f ) is the SHapley value for feature i, capturing its
marginal contribution to the prediction in all possible
coalitions.
By computing the contribution of feature i across
all coalitions, SHAP provides a complete view of
how each feature influences the model’s output. The
interpretability it offers helps stakeholders pinpoint
which features are driving predictions in individual
cases bringing in transparency in FFNN predictions.
Integrated Gradients. Unlike SHAP which focuses
on instance-level feature contributions, Integrated
Gradients (Sundararajan, 2017) provide a global per-
spective on feature relevance by measuring the cumu-
lative effect of gradients along a path from a baseline
w
′
to the input w. The baseline is typically chosen as
a neutral input such as a black image for vision mod-
els or a zero vector for text models. By integrating
gradients along this path it capture the sensitivity of
the model’s output with respect to each feature help-
ing explain the model’s behavior over the entire input
space.
The Integrated Gradient for feature i, IG
i
(x), is de-
fined as in Equation (2):
IG
i
(w) = (w
i
− w
′
i
) ×
Z
1
α=0
∂G(w
′
+ α × (w − w
′
))
∂w
i
dα (2)
Here:
- w is the input for which we calculate attributions.
- w
′
(or w
0
) is the baseline input, serving as a
reference point for comparison.
- w
i
is the i-th feature value of the input x.
- w
′
i
is the i-th feature value of the baseline input w
′
.
- α is a scaling factor ranging from 0 to 1,
- G is the function representing the model (e.g., a
neural network), mapping inputs to outputs.
Integrated Gradients help explain how much each
feature contributes to a model’s output by capturing
An Empirical Study Using Machine Learning to Analyze the Relationship Between Musical Audio Features and Psychological Stress
1035

the average gradient along the path from baseline to
input. This provides insights into the overall influence
of each feature across the model, rather than individ-
ual predictions.
Together, SHAP and Integrated Gradients provide
complementary insights into Feed Forward Neural
Network predictions. While SHAP explains feature
impact on a case-by-case basis, Integrated Gradients
capture a broader view of feature relevance across
different instances.Using both methods can lead to a
deeper understanding of the model behavior, particu-
larly in sensitive applications like healthcare, where
interpretability is crucial for ethical considerations,
transparency, and trust. This XAI interpretability
framework provides stakeholders with granular de-
tails and global trends, fostering a more comprehen-
sive understanding of model decisions and enhancing
confidence in AI-assisted healthcare predictions.
4.2.1 Not Stressed Class
Figure 4: XAI SHAP Plot for Not Stressed Class.
For ’Not Stressed’ class as seen in Figure 4, Tempo
emerged as one of the most influential features ex-
hibiting a clear inverse relationship with the ”not
stressed” classification. Higher tempo values, repre-
sented by pink dots in the SHAP plot, consistently
showed negative SHAP values, indicating that faster-
paced music significantly decreases the likelihood of
a ”not stressed” classification. On the other hand,
lower tempo values (blue dots) displayed positive
SHAP values, suggesting that slower-paced music
contributes positively to the ”not stressed” prediction.
Danceability showed a notable pattern where
higher values (pink dots) on the positive SHAP value
side indicated that more danceable tracks are more
likely to be classified as ”not stressed”. This suggests
that songs with stronger rhythmic elements and reg-
ular patterns may contribute to a less stressful listen-
ing experience. Acousticness showed a more com-
plex distribution with moderate values (purple dots)
having a positive impact on the ”not stressed” clas-
sification. This indicates that songs with balanced
acoustic elements are more likely to be classified as
not-stressful compared to those at either extreme of
the acousticness spectrum.
Energy values clustered around the center of the
SHAP value range, with a slight positive skew for
higher energy levels, suggesting a positive association
with the ”not stressed” classification. Features such as
Speechiness and Instrumentalness exhibited minimal
impact on the model’s predictions, as evidenced by
their tight clustering around zero SHAP values. The
analysis also revealed that Liveness had a slight neg-
ative impact when at higher values, indicating that
highly live recordings are less frequently associated
with the ”not stressed” classification. This could po-
tentially be attributed to the more unpredictable and
dynamic nature of live performances.
Figure 5: Integrated Gradients for Not Stressed class.
The Integrated Gradients analysis for not stressed
class as in Figure 5 strongly validates our SHAP
value findings. Notably, Danceability shows the high-
est positive contribution, aligning with our SHAP in-
terpretation where higher danceability values posi-
tively influenced ”not stressed” predictions. Simi-
larly, Tempo displays the strongest negative contribu-
tion (-0.45), which perfectly corroborates our SHAP
analysis where higher tempo values pushed predic-
tions away from the ”not stressed” class. The moder-
ate positive contributions of Energy and Mode, along
with the negative impacts of Instrumentalness and
Liveness, also mirror the patterns observed in the
SHAP visualization, reinforcing the reliability of our
feature importance interpretations.
These findings provide valuable insights into the
musical characteristics that contribute to a song being
classified as ”not stressed,” with tempo and danceabil-
ity emerging as particularly significant predictors in
the model’s decision-making process.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
1036

4.2.2 Stressed Class
Figure 6: XAI SHAP Plot for Stressed Class.
For ’Stressed’ class as observed in Figure 6, Tempo
demonstrated the most substantial influence, showing
a strong positive correlation with stress classification.
The visualization indicates that higher tempo values
(pink dots) are predominantly positioned on the pos-
itive SHAP value side, suggesting that faster-paced
music significantly increases the likelihood of a song
being classified as ”stressed”. This aligns with the in-
tuitive understanding that rapid tempos may induce
heightened arousal states. Speechiness emerged as
the second most influential feature, with lower values
(blue dots) clustered on the negative side and higher
values (pink and purple dots) extending toward pos-
itive SHAP values. This suggests that songs with
greater vocal presence and spoken word content are
more likely to be classified as stressed. Liveness
showed a notable distribution pattern where higher
values (pink dots) extended into positive SHAP val-
ues, indicating that songs with stronger live perfor-
mance characteristics tend to be classified as more
stressful. This could be attributed to the increased am-
bient noise and audience participation typical in live
recordings.
Mode exhibited an interesting pattern with higher
values (pink dots) concentrated on the negative side,
suggesting that major mode songs are less likely to be
classified as stressed, while lower values (blue dots)
on the positive side indicate minor mode songs con-
tribute to stress classification. Instrumentalness dis-
played a scattered pattern with moderate to high val-
ues suggesting that highly instrumental tracks have
some association with stress classification.
Danceability showed an inverse relationship, with
higher values (pink dots) concentrated on the neg-
ative SHAP value side, indicating that more dance-
able tracks are less likely to be classified as stressed.
Acousticness and Energy demonstrated more modest
impacts, with relatively tight clustering around zero,
though both showed slight tendencies toward negative
SHAP values for higher feature values.
Figure 7: Integrated Gradients for Stressed Class.
The Integrated Gradients as seen in Figure 7 vali-
dates our SHAP analysis with Tempo emerging as the
strongest predictor (0.15) for stressed music. Speech-
iness and Liveness show notable positive contribu-
tions, aligning with their positive SHAP values. The
negative impacts of Danceability and Energy further
confirm our SHAP interpretations of their inverse re-
lationship with stress classification. The results sug-
gest that faster, speech-heavy, and live performance
elements are more strongly associated with stressed
music classification, while danceable and major mode
characteristics tend to oppose this classification.
4.2.3 Borderline Class
Figure 8: XAI SHAP Plot for Borderline Class.
In the ’Borderline’ class as in Figure 8, Mode
emerged as one of the most distinctive features, show-
ing a clear bimodal distribution. Lower values (blue
dots) appeared significantly on the positive SHAP
value side, while higher values (pink dots) were con-
centrated on the negative side. This suggests that mi-
nor mode songs (represented by lower values) have a
stronger association with borderline stress classifica-
tion, while major mode songs tend to push predictions
away from this category. Tempo displayed an interest-
ing distribution with both high (pink dots) and moder-
An Empirical Study Using Machine Learning to Analyze the Relationship Between Musical Audio Features and Psychological Stress
1037

ate (purple dots) values showing positive SHAP val-
ues, indicating that mid to higher-tempo songs con-
tribute to borderline stress classification. The pattern
suggests that tempo has a more nuanced impact com-
pared to its role in clear-cut stressed or not-stressed
classifications.
Speechiness demonstrated a pattern where lower
values (blue dots) showed positive SHAP values, sug-
gesting that songs with less vocal content are more
likely to be classified as borderline stressed. This
contrasts with its impact on definitive stress classifica-
tions, indicating a unique characteristic of borderline
cases.
Liveness and Danceability showed relatively mod-
est impacts, with slight clustering around zero but ex-
tending into both positive and negative SHAP values.
This suggests these features play a more subtle role in
borderline classification compared to their influence
on definitive stress categories. Acousticness exhibited
a pattern where moderate values (purple dots) showed
slight positive SHAP values, indicating that balanced
acoustic characteristics might contribute to borderline
classification.
Instrumentalness and Energy displayed relatively
concentrated distributions near zero, with Energy
showing a slight tendency toward positive SHAP val-
ues for moderate (purple) feature levels, suggesting
these features have minimal but consistent impacts on
borderline classification.
Figure 9: Integrated Gradients for Borderline Class.
The Integrated Gradients for borderline stress
class in Figure 9 shows Tempo with the highest pos-
itive contribution (0.3), matching our SHAP findings
where mid-to-high tempo values indicated borderline
stress. Danceability shows the strongest negative im-
pact (-0.13), while Mode and Acousticness display
moderate negative contributions. The results sug-
gest that borderline stress music often combines ele-
ments typically associated with both stressed and not-
stressed categories, creating a distinct musical profile
for this intermediate classification.
4.3 Model Validation on External
Datasets
4.3.1 Validation Technique and Its Importance
Validating an audio classifier using external datasets
from different regional languages is a crucial step
in assessing the robustness and generalization ca-
pabilities of the model. By testing the model on
datasets that it hasn’t seen before—especially those
that vary in linguistic, acoustic, and cultural charac-
teristics—the evaluation process goes beyond stan-
dard train-test splits. This technique helps ensure that
the classifier isn’t overfitting to the specific features of
the training data but instead is capable of accurately
categorizing diverse audio inputs. In this context, val-
idating with Tamil and Hindi music datasets allows
for a comprehensive evaluation, considering differ-
ent tonal qualities, rhythms, and cultural nuances in
music, ultimately providing confidence in the model’s
versatility and performance.
4.3.2 Performance on External Music Dataset
When evaluated on the Hindi music dataset, the au-
dio classifier achieved a high test accuracy of 98.63%
with a minimal test loss of 0.0478. This indicates that
the model effectively distinguishes between stressed,
not-stressed, and borderline music categories, even
when exposed to new, linguistically rich audio con-
tent. The high accuracy reflects the model’s ability to
generalize its learned thresholds for stress classifica-
tion beyond its training set, capturing the nuanced fea-
tures of Hindi music that may influence emotional and
stress responses. The model demonstrated an even
higher level of accuracy on the Tamil music dataset,
achieving a test accuracy of 99.43% and a remark-
ably low test loss of 0.0148. This impressive result
suggests that the classifier not only adapts well to the
phonetic and rhythmic distinctiveness of Tamil mu-
sic but also maintains consistency in identifying the
stress levels associated with different tracks.
5 CONCLUSION AND FUTURE
SCOPE
In conclusion, our study demonstrates that among the
wide range of music audio features such as Dancabil-
ity, Tempo, Energy, and Acousticness etc. are par-
ticularly effective in reducing psychological stress.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
1038

Leveraging these scientifically validated thresholds,
we developed a robust audio classifier using a Feed-
Forward Neural Network (FFNN) which efficiently
categorizes music across all languages and genres into
”Stressed,” ”Not-Stressed,” and ”Borderline” class la-
bels by analyzing key features, which are linked to
positive emotional valence and arousal. This uni-
versal audio feature-based approach surpasses genre-
specific limitations, offering accurate and culturally
inclusive stress classification across diverse languages
and music styles. For future work, we aim to explore
and incorporate advanced techniques that enhance the
interpretability of our model.
ACKNOWLEDGEMENTS
We extend our deepest gratitude to Dr Keya Das, De-
partment of Psychiatry at PES University Institute
of Medical Sciences and Research for her invaluable
contributions to this study. Her expertise, time, and
insightful feedback were crucial in validating our ap-
proach and ensuring the success of this research.
REFERENCES
Adiasto, K., B. D. G. J. v. H.-M. L. M.-R. K. . G. S. A. E.
(2022). Music listening and stress recovery in healthy
individuals: A systematic review with meta-analysis
of experimental studies. PLoS One, 17(6):e0270031.
Adiasto, K., van Hooff, M. L. M., Beckers, D. G. J., and
Geurts, S. A. E. (2023a). The sound of stress recovery:
an exploratory study of self-selected music listening
after stress. BMC Psychology, 11(1):40.
Adiasto, K., van Hooff, M. L. M., Beckers, D. G. J., and
Geurts, S. A. E. (2023b). The sound of stress recovery:
an exploratory study of self-selected music listening
after stress. BMC Psychology, 11(1):40.
Ahuja, R. & Banga, A. (2019). Mental stress detection
in university students using machine learning algo-
rithms. Procedia Computer Science, 152:349–353.
Anderson, C. A., Carnagey, N. L., and Eubanks, J. (2003).
Exposure to violent media: The effects of songs
with violent lyrics on aggressive thoughts and feel-
ings. Journal of Personality and Social Psychology,
84(5):960–971.
Duman, D., Neto, P., Mavrolampados, A., Toiviainen, P.,
and Luck, G. (2022). Music we move to: Spotify
audio features and reasons for listening. PLoS One,
17(9):e0275228.
Erbay Dalli,
¨
O., B. C. . Y. Y. (2023). The effectiveness
of music interventions on stress response in intensive
care patients: A systematic review and meta-analysis.
Journal of Clinical Nursing, 32(11-12):2827–2845.
Gu, Y.-Y., Z. L.-L. T. X.-Y. Y. F.-X. L. L.-M. G. Y.-S. L. H.
L. T.-Z. B. G.-Q. . F. Z.-Q. (2023). A spatial transcrip-
tome reference map of macaque brain states. Transla-
tional Psychiatry, 13:220.
Gujar, S. S. & Reha, A. Y. (2023). Exploring relationship
between music and mood through machine learning
technique. In Proceedings of the 5th International
Conference on Information Management & Machine
Intelligence, pages 1–6.
He, H., Bai, Y., and Garcia, E. A. (2008). Adasyn: Adaptive
synthetic sampling approach for imbalanced learning.
In Proceedings of the 2008 International Joint Con-
ference on Artificial Intelligence, pages 1322–1327.
Juslin, P. N. and Sloboda, J. A. (2010). Handbook of Music
and Emotion: Theory, Research, Applications. Oxford
University Press, Oxford, UK.
Lin, S. T., Y. P. L. C. Y.-S. Y. Y.-Y. Y. C.-H. M. F.-. C. C. C.
(2011). Mental health implications of music: Insight
from neuroscientific and clinical studies. Harvard Re-
view of Psychiatry, 19(1):34–46.
Liu, X., Wu, J., and Zhou, Z. (2007). Generative oversam-
pling for mining imbalanced datasets. Proceedings of
the 7th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, pages 9–18.
Lundberg, S. M. & Lee, S. I. (2017). A unified approach to
interpreting model predictions. In Advances in Neural
Information Processing Systems, volume 30.
McFerran, K. S., G. S. H. C. . M. K. (2014). Examining the
relationship between self-reported mood management
and music preferences of australian teenagers. Nordic
Journal of Music Therapy, 24(3):187–203.
Saarikallio, S. & Erkkil
¨
a, J. (2007). The role of music in
adolescents’ mood regulation. Psychology of Music,
35(1):88–109.
Stewart, J., Garrido, S., Hense, C., and McFerran, K.
(2019). Music use for mood regulation: Self-
awareness and conscious listening choices in young
people with tendencies to depression. Frontiers in
Psychology, 10:1199.
Sundararajan, M., T. A. . Y. Q. (2017). Axiomatic attribu-
tion for deep networks. In International Conference
on Machine Learning, pages 3319–3328. PMLR.
Vuust, P., Heggli, O. A., Friston, K. J., et al. (2022). Music
in the brain. Nature Reviews Neuroscience, 23:287–
305.
Zhang, Y., Chen, J., Liao, L., and Yao, K. (2018). Ran-
dom noise injection-based adversarial training for ro-
bust speech recognition. In Proceedings of the 2018
IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), pages 3156–3160.
An Empirical Study Using Machine Learning to Analyze the Relationship Between Musical Audio Features and Psychological Stress
1039