Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1962

Explainable AI (XAI) for Obesity Prediction:

An Optimized MLP Approach with SHAP
Interpretability on Lifestyle and
Behavioral Data
Darren Kevin T. Nguemdjom1; Alidor M. Mbayandjambe2; Grevi B. Nkwimi3;
Fiston Oshasha4; Célestin Muluba5; Héritier I. Mbengandji6; Ibsen G. Bazie7
1,2,7
International School, Vietnam National University, Hanoi, Vietnam
2,4,5
University of Kinshasa, Faculty of Sciences and Technology, Kinshasa, DR Congo
2,3
University of Kinshasa, Faculty of Economic and Management Sciences, Kinshasa, DR Congo
6
Department of Letters and Humanities, Institut Supérieur Pédagogique Du Sud Banga, Ilebo, DR Congo

Publication Date: 2025/05/12

Abstract: Obesity represents a major public health challenge, requiring accurate and interpretable predictive tools. This study
proposes an approach based on a Multilayer Perceptron (MLP) optimized to predict obesity levels from lifestyle data, eating habits, and
physiological characteristics, using a comprehensive Kaggle dataset combining real and synthetic samples. After rigorous preprocessing,
including normalization and class rebalancing, we compare the performance of the MLP with four classical algorithms (Logistic
Regression, KNN, Random Forest, and XGBoost) using comprehensive metrics (accuracy, precision, recall, F1-score, AUC-ROC). The
results demonstrate the superiority of the optimized MLP (98.4% accuracy, F1-score of 0.97) over the other models, with a significant
improvement from hyperparameter optimization through GridSearchCV. The XAI analysis via SHAP identifies weight, gender, height,
and physical activity as the most determinant factors, providing crucial transparent explanations for clinical applications. This
combination of high predictive performance and interpretability makes the MLP a valuable tool for obesity prevention and diagnosis in
public health.

Keywords: Obesity Prediction, Machine Learning, Lifestyle Data, SHAP, AUC-ROC, GridSearchCV, Interpretable AI, MLP,
Health Analytics.

How to Cite: Darren Kevin T. Nguemdjom; Alidor M. Mbayandjambe; Grevi B. Nkwimi; Fiston Oshasha; Célestin Muluba;
Héritier I. Mbengandji; Ibsen G. BAZIE; (2025). Explainable AI (XAI) for Obesity Prediction: An Optimized MLP
Approach with SHAP Interpretability on Lifestyle and Behavioral Data. International Journal of Innovative
Science and Research Technology, 10 (4), 3192-3200. https://doi.org/10.38124/ijisrt/25apr1962

I. INTRODUCTION systematic review of ML-based obesity prediction

approaches, highlighting the best combinations of dietary,
Obesity is a chronic multifactorial disease that today demographic, and behavioral variables. Meanwhile, Lin et al.
represents a major global public health issue. According to (2023) successfully applied several supervised models to
the World Health Organization, more than 650 million adults predict childhood obesity based on data collected from
were affected by obesity in 2016, a number that continues to parents and children. These studies confirm that leveraging
grow each year (World Health Organization, 2021). This lifestyle data can reliably predict weight status. In parallel,
condition is associated with increased risks of type 2 deep learning (DL) techniques, notably multilayer
diabetes, hypertension, cardiovascular diseases, and certain perceptrons (MLP) and recurrent neural networks (LSTM),
cancers (Guh et al., 2009). In response to this issue, research have been used to capture complex nonlinear relationships
is focused on developing predictive tools to identify between health variables (Mahmut et al., 2023). However,
individuals at risk before the appearance of clinical the use of these models in medical contexts raises the crucial
symptoms. In recent years, artificial intelligence (AI) models question of their explainability. To address this issue,
have shown their effectiveness in this field. In particular, explainable AI methods such as SHAP (SHapley Additive
machine learning (ML) approaches such as Random Forest, Explanations) (Lundberg & Lee, 2017) or LIME techniques
Support Vector Machine, and XGBoost have achieved high have been developed to interpret the decisions of "black box"
accuracy in classifying obesity levels (Maria et al., 2023; Lin models. In this work, we propose a hybrid approach that
et al., 2023). The study by Maria et al. (2023) provides a combines traditional machine learning algorithms and deep

IJISRT25APR1962 www.ijisrt.com 3192

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1962
learning models, applied to a Kaggle dataset on obesity. We Finally, the issue of explainability remains largely
evaluate and compare the performance of several models underexplored. Tjoa and Guan (2021) emphasize that the
(Logistic Regression, KNN, Random Forest, XGBoost, integration of explainable AI (XAI) tools like SHAP or
MLP), optimize them via GridSearchCV, and analyze their LIME remains marginal in health-related studies. This lack of
behavior using global indicators (accuracy, F1-score, AUC- transparency poses a significant barrier to clinical adoption,
ROC) and explainable methods (SHAP). The objective of where it is essential to justify predictions, especially for
this study is to provide a high-performance, interpretable, and critical medical decisions.
reproducible approach for the personalized prediction of
obesity risk. Recent studies, such as that by Lin et al. (2023), have
demonstrated that the use of SHAP allows for identifying key
II. LITERATURE REVIEW predictive factors of obesity, such as family history and
eating habits, providing a better understanding of model
The prediction of obesity using artificial intelligence decisions. Additionally, research by Du et al. (2024)
(AI) methods has garnered increasing interest, particularly highlighted the effectiveness of an obesity risk prediction
with the rise of machine learning (ML) and deep learning system based on XGBoost and SHAP, enabling personalized
(DL) techniques. However, despite notable advances, several healthcare management. However, despite these
methodological limitations persist in the current literature. advancements, challenges remain, particularly regarding the
computational complexity of SHAP for large datasets and
Traditionally, logistic regression has been one of the complex models, as well as the need for thorough clinical
most widely used tools in epidemiology for predicting health validation to ensure the reliability of interpretations provided
conditions due to its simplicity and interpretability. Cheng by these tools.
and Rai (2021) emphasize that this model is well-suited for
simple linear relationships. However, its ability to handle III. PROPOSED METHODOLOGY
complex interactions between behavioral and biological
variables remains limited, making it unsuitable for A. Dataset
multifactorial phenomena like obesity. The dataset used in this study is the "Obesity Level
Dataset," made available on Kaggle by Mehrparvar (2023). It
To address these shortcomings, machine learning-based consists of 2,111 individuals, each described by 17
approaches have been proposed. Lin et al. (2023) explored explanatory variables related to dietary habits (e.g.,
algorithms such as Random Forest and SVM for predicting frequency of fast food consumption, hydration), physical
childhood obesity. Although their results showed significant activity, family history of overweight, and certain
improvement compared to traditional statistical methods, demographic characteristics such as gender and age. The
their study did not address the critical issue of class target variable is multiclass and categorizes individuals into
imbalance nor the impact of hyperparameter optimization on seven obesity levels: [0] Underweight, [1] Normal weight,
model robustness. Furthermore, their approach remained [2] Overweight level I, [3] Overweight level II, [4] Obesity
limited to the simple application of algorithms without type I, [5] Obesity type II, and [6] Obesity type III. This
explicit systematic optimization. granularity is essential for modeling the progression of body
weight states, as emphasized in recent works by Begum et al.
On the other hand, Saxena et al. (2023) highlighted the (2024), who stress the importance of a finer classification in
effectiveness of boosting techniques, particularly XGBoost. clinical predictions related to obesity.
However, this review also notes that most studies focus
solely on raw performance (accuracy), neglecting metrics B. Data Preprocessing
more suitable for imbalanced medical contexts, such as F1- Before training the models, several preprocessing steps
score or AUC-ROC. Moreover, few studies account for the were applied to improve the quality of the inputs. First,
need to make these models understandable for healthcare categorical variables such as Gender,
professionals. family_history_with_overweight, SMOKE, or MTRANS
were converted into numerical values using the LabelEncoder
The rise of deep learning has led to the adoption of method, which is commonly used in similar studies to handle
neural networks such as Multilayer Perceptron (MLP) for non-numeric attributes (Alsareii et al., 2023). Continuous
predicting obesity, as demonstrated by Mahmut et al. (2023). numeric variables such as Weight, Height, FCVC, FAF, or
While these architectures capture complex nonlinear CH2O were normalized using StandardScaler to ensure scale
relationships, they are often seen as "black boxes." The lack homogeneity. This preprocessing is essential for models
of explainability tools in these studies severely limits their sensitive to Euclidean distances, particularly KNN and MLP,
applicability in clinical contexts, where understanding as highlighted by Helforoush et al. (2024) in their
algorithmic decisions is crucial. Another major weakness comparative study on machine learning-based obesity
identified by Sadiku et al. (2019) is the lack of classification. The target variable was encoded as a one-hot
hyperparameter optimization. Most studies rely on default vector for deep learning models, a common practice in
configurations without using techniques like GridSearchCV architectures with a softmax output. The dataset was then
or RandomizedSearchCV, compromising the generalization stratified and split into 80% for training and 20% for testing,
of models outside of training datasets. while maintaining the class proportions. This procedure aims
to avoid distribution bias and is particularly recommended
when certain classes are underrepresented, as observed in the
study by Genc et al. (2025) on similar models.

IJISRT25APR1962 www.ijisrt.com 3193

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1962
C. Model Training and Evaluation metric, combining the classic Euclidean distance with a
We evaluated and compared five supervised clinical similarity measure, in accordance with the
classification algorithms applied to predicting obesity levels. recommendations of Zhang et al. (2022). This approach
The choice of models is based on both their demonstrated allows the integration of explicit medical knowledge in the
effectiveness in medical and nutritional contexts, as well as distance calculation by over-weighting critical variables such
their complementarity in terms of complexity. as BMI or family history, thus improving the relevance of
neighborhoods in a biomedical context.
 Multinomial Logistic Regression,
Was used as a baseline model for the classification of E. Multilayer Perceptron (MLP)
obesity levels. This linear model was optimized using the We designed a Multilayer Perceptron (MLP)
SAGA solver, compatible with L2 regularization, to specifically tailored to model the complex nonlinear
minimize the regularized logistic loss function. A relationships between behavioral, demographic, and
GridSearchCV was conducted to identify the best biomedical variables involved in classifying obesity levels.
hyperparameters: C = 10, penalty = 'l2', solver = 'saga' Our architecture includes four hidden layers (256, 128, 64, 32
(Hosmer et al., 2000). neurons) combined with advanced regularization techniques
such as Dropout (Srivastava et al., 2014) and Batch
 Random Forest, Normalization (Ioffe et al., 2014), ensuring better
Proposed by Breiman (2001), was chosen for its generalization and accelerated convergence. The use of these
robustness to noisy data and its ability to model complex methods follows recent recommendations by Raghu et al.
non-linear interactions. Each tree is built using bagging, and (2019), who emphasize their effectiveness in neural networks
the final prediction is obtained through majority voting. applied to health data.
Hyperparameter optimization (n_estimators = 500,
max_depth = 15, min_samples_split = 5, max_features = The model optimization was performed using the Adam
'sqrt') via GridSearchCV ensured excellent stability without algorithm (Kingma & Ba, 2014), which is particularly suited
overfitting. for heterogeneous and moderately sized datasets due to its
dynamic adjustment of learning rates. An early stopping
 XGBoost, mechanism was integrated to prevent overfitting, in line with
Developed by Chen & Guestrin (2016), was adopted for best practices outlined by Prechelt et al. (2012), by halting
its ability to efficiently handle complex multiclass training when the validation loss stopped improving after 10
classification problems using gradient boosting. It was iterations.
configured through rigorous hyperparameter optimization
using GridSearchCV with stratified 5-fold cross-validation, F. Framework for Explainability (SHAP)
ensuring an optimal balance between bias and variance. Key To ensure the transparency and interpretability of our
parameters adjusted include n_estimators, max_depth, predictive models applied to obesity level classification, we
learning_rate, and subsample, ensuring effective integrated a systematic approach based on the SHAP
regularization to prevent overfitting. To enhance (SHapley Additive exPlanations) method, developed by
explainability, SHAP analysis was integrated to identify the Lundberg and Lee (2017). This technique, based on game
most influential variables in the model’s decisions, in line theory, assigns each variable a precise contribution in the
with the recommendations of Lundberg & Lee (2017). This decision of each prediction, making complex models such as
methodological approach combining performance and Random Forest, XGBoost, and Multilayer Perceptron (MLP)
transparency meets the requirements of sensitive medical interpretable.
applications (Chen et al., 2016).
Our methodology relied on the use of various
D. Clustering and Segmentation explainers provided by the SHAP library, utilizing
TreeExplainer for tree-based models (Random Forest and
 K-Nearest Neighbors (KNN), XGBoost) to ensure quick and accurate SHAP value
Algorithm was implemented as a non-parametric computation tailored to these structures, as well as
classification method based on the principle of local DeepExplainer for MLP, which efficiently estimates the
similarity. The methodological optimization was based on contributions of variables within neural networks. For each
several approaches to adapt this classifier to the specificities model, we conducted both global and local analyses using
of heterogeneous biomedical data. To overcome the inherent summary plots, which provide an overview of the average
limitations of the KNN algorithm, particularly those related importance of variables across all predictions, and
to the curse of dimensionality and the hubness phenomenon dependence plots, which are used to examine the marginal
effect of key variables while highlighting potential nonlinear
interactions.

described by Radovanović et al. (2009), we

incorporated two major methodological improvements. First,
an adaptive weighting of the votes was implemented,
following the approach of Chaoyu et al. (2023), to reduce the
disproportionate influence of majority classes by dynamically
adjusting the weight assigned to each neighbor based on the
local density of the data. Second, we implemented a hybrid

IJISRT25APR1962 www.ijisrt.com 3194

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1962

Fig 1 Pipeline of the SHAP Explain Ability Methodology for Obesity Classification Models

IV. INTERPRETATION OF EXPERIMENTAL analyzing each model's ability to correctly classify samples
RESULTS while accounting for class imbalances, which is particularly
important in a medical context where prediction accuracy and
A. Global Performance Comparison reliability are crucial. The results obtained for each model
To assess the comparative performance of the different were compared to identify the most effective ones in the
approaches implemented, we summarized the results obtained context of our study, with a particular focus on each model's
by each model according to standard metrics in multiclass ability to handle complex interactions between behavioral,
classification: accuracy, precision, recall, macro F1-score, and demographic, and biomedical variables.
average AUC-ROC when applicable. These metrics allow for

Table 1 Comparative Performance Table of ML/MLP Models (Default Settings)

Modèle Accuracy (%) Précision (%) Recall (%) F1-score (%) AUC-ROC moyen
Logistic Regression 87.2 87.2 87.5 87.4 ~0.89
K-Nearest Neighbors 80.9 81.5 80.8 80 ~0.91
Random Forest 95.3 96.7 95.6 95 ~0.99
XGBoost 95.7 96.9 96.4 96 ~0.995
MLP 98.4 97.8 97.3 97.1 N/A

Our study demonstrates that MLP and XGBoost are the  Impact of Hyperparameter Optimization on the Accuracy of
two most performant models, with respective accuracies of Machine Learning Models
98.4% and 95.8%. Random Forest remains highly
competitive (95.1%), while Logistic Regression and KNN This table highlights the improvement in accuracy
show lower, yet acceptable, performance after optimization. achieved after hyperparameter optimization for each machine
The performance gain after hyperparameter tuning exceeds learning model.
+6% for certain models (Logistic Regression, KNN),
highlighting the importance of systematic optimization.

Table 2 Accuracy Gain After Hyper Parameter Optimization for ML Models

Models Accuracy Accuracy (optim) Gain (%)
Logistic Regression 0.871 0.936 6.50
K-Nearest Neighbors 0.813 0.872 5.90
Random Forest 0.949 0.951 0.20
XGBoost 0.912 0.969 5.70

IJISRT25APR1962 www.ijisrt.com 3196

3195
Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1962
Table 2 shows that Logistic Regression and K-Nearest B. Detailed Model Analysis
Neighbors benefit from the most significant gains, with
improvements of +6.5% and +5.9% in accuracy, respectively,  One-vs-Rest AUC-ROC Curves
highlighting their sensitivity to fine-tuning. The XGBoost A detailed analysis of the performance of the RF,
model also shows a notable improvement of +5.7%, XGBoost, KNN, and LR models was conducted through the
confirming the importance of parameter tuning in boosting interpretation of the One-vs-Rest AUC-ROC curves to assess
algorithms. their discriminative capacity for each class.

Fig 2 Multi-Class AUC-ROC Curves for ML Models

Figure 2 illustrates the multi-class AUC-ROC curves for highlighting its limitations in handling complex relationships
the different evaluated models. It can be observed that and class imbalances.
XGBoost (a) and Random Forest (b) demonstrate exceptional
performance, with AUCs close to or equal to 1 for all classes,  The learning curves for the MLP model (accuracy & loss),
indicating excellent discriminative ability. The K-Nearest Show the progressive convergence of the model, which
Neighbors (c) model yields satisfactory but more variable minimizes categorical cross-entropy loss using the Adam
results, with a noticeable performance drop for class 1 (AUC = optimizer (learning_rate=0.0005), with significant
0.82). In contrast, Logistic Regression (d) displays curves less improvements observed over iterations.
close to the upper left corner, with overall lower AUCs,
particularly for the minority classes (Classes 5 and 6),

Fig 3 Learning curves (accuracy & loss)

IJISRT25APR1962 www.ijisrt.com 3196

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1962
The Figure 3 demonstrate significant stability between 0.9716, and a macro recall of 0.9731, confirming the model's
the training and validation sets. The final evaluation on the test effectiveness in predicting obesity levels.
set reveals exceptional performance with an accuracy of
98.49%, a macro F1-score of 0.9714, a macro precision of  KNN Validation Curve

Fig 4 Validation curve Showing the Evolution of Accuracy and F1-score as a Function of the
Number of Neighbors k (k ∈ [1, 30])

Figure 4 illustrates the impact of the parameter k on the C. Feature Importance and Interpretability
performance of the K-Nearest Neighbors classifier. It is The comparative analysis of feature importance across
observed that both accuracy and macro F1-score are models represents a critical step in identifying consistent
maximized for k=1, reaching approximately 85.5% and biomedical determinants and potential algorithmic biases.
84.9%, respectively. Beyond this point, performance gradually
declines, suggesting a loss of precision in class discrimination.

Figure 4: Relative Importance of Features Across Models

Figure 4 presents the normalized feature importance in D. Explainability Insights (SHAP Analysis)
our study, highlighting Weight, Height, and Frequency of
Vegetable Consumption (FCVC) as the most influential  SHAP Summary Plots,
variables across all models. However, notable differences are This section discusses the dominant variables identified
observed: XGBoost and Random Forest assign greater through SHAP analysis, highlighting key factors such as
importance to Age and Gender, whereas Logistic Regression Weight, Height, Age, and Frequency of Vegetable
and K-Nearest Neighbors place more emphasis on factors such Consumption (FCVC) as the most influential in predicting
as Weight and Height. obesity levels across the evaluated models.

IJISRT25APR1962 www.ijisrt.com 3197

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1962

Fig 5 Relative Importance of Variables Across Models

Figure 5 illustrates the impact of variables on model  SHAP Stacked Bar Charts for Multi-Class (MLP)
predictions for K-Nearest Neighbors (a), Logistic Regression A comparative analysis of feature influence across
(b), and XGBoost (c) using SHAP values. The color gradient predicted classes, highlighting how each variable contributes
represents the feature values, with blue indicating low values differently depending on the obesity level classification.
and red indicating high values, providing a clear visualization
of how each variable influences the classification outcomes.

Figure 6: SHAP summary plots

IJISRT25APR1962 www.ijisrt.com 3198

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1962
Figure 6 visualizes the average feature importance for Logistic regression, while useful, remains limited in
each target class (from Underweight to Obesity Type III). modeling complex interactions, and KNN exhibits
The chart highlights a clear dominance of certain behavioral weaknesses in handling class imbalances. The integration of
and biometric variables across classifications. SHAP (SHapley Additive exPlanations) enhanced model
interpretability by highlighting the importance of key
V. DISCUSSION variables such as weight, height, and age. Our hybrid
approach, combining high performance, transparency, and
The results of our study demonstrate that advanced
generalizability, provides a robust and reproducible pipeline
models such as XGBoost and MLP outperform traditional
applicable in clinical settings.
approaches like Logistic Regression and KNN, achieving
accuracies of 95.7% and 98.4%, respectively. These models
REFERENCES
effectively capture complex non-linear relationships,
consistent with the findings of Sajid et al. (2022) and Esteva
[1]. World Health Organization. (2021). Obesity and
et al. (2019). However, KNN remains limited by the curse of
overweight. https://www.who.int/news-room/fact-
dimensionality and class imbalance issues, as noted by
sheets/detail/obesity-and-overweight
Radovanović et al. (2009). While hyperparameter
[2]. Guh, D. P., et al. (2009). The incidence of co-
optimization through GridSearchCV significantly enhanced
morbidities related to obesity and overweight: A
model performance, generalizing these results to other
systematic review and meta-analysis. BMC Public
populations remains challenging (Brown et al., 2024). The
Health, 9, 88. https://doi.org/10.1186/1471-2458-9-88
integration of SHAP improved model interpretability by
[3]. Maria, A.S & Ramasamy, Sunder & Kumar,
identifying key variables such as weight, height, and age,
R.Satheesh. (2023). Obesity Risk Prediction Using
aligning with the recommendations of Lundberg and Lee
Machine Learning Approach. 1-7. 10.1109/I
(2017) and Samek et al. (2017). Nevertheless, SHAP’s
CNWC57852.2023.10127434.
computational complexity on large datasets restricts its
[4]. Lin W, Shi S, Huang H, Wen J, Chen G. Predicting
applicability in real-time clinical environments, as
risk of obesity in overweight adults using interpretable
highlighted by Tjoa and Guan (2021).
machine learning algorithms. Front Endocrinol
(Lausanne). 2023 Nov 17;14:1292167. doi:
VI. LIMITATIONS AND FUTURE RESEARCH
10.3389/fendo.2023.1292167. PMID: 38047114;
PMCID: PMC10693451.
Although this study demonstrated the effectiveness of [5]. M. Dirik, “Application of machine learning
advanced models in predicting obesity risk, several techniques for obesity prediction: a comparative
limitations must be considered. The use of a single dataset study,” Journal of Complexity in Health Sciences,
from Kaggle may introduce selection bias, particularly due to Vol. 6, No. 2, pp. 16–34, Oct. 2023,
limited sample diversity. Furthermore, while hyperparameter https://doi.org/10.21595/chs.2023.23193
optimization improved model performance, the optimal [6]. Lundberg, Scott & Lee, Su-In. (2017). A Unified
configuration may not be generalizable to other populations. Approach to Interpreting Model Predictions.
The integration of SHAP (SHapley Additive exPlanations) 10.48550/arXiv.1705.07874.
for model interpretability presents computational challenges, [7]. Saxena, A.; Mathur, N.; Pathak, P.; Tiwari, P.;
especially for complex models such as XGBoost and MLP, Mathur, S.K. Machine Learning Model Based on
limiting its real-time application in clinical settings. Finally, Insulin Resistance Metagenes Underpins Genetic
the clinical acceptability of AI-driven decision support Basis of Type 2 Diabetes. Biomolecules 2023, 13,
systems remains a major challenge, particularly for critical 432. https://doi.org/10.3390/biom13030432
healthcare decisions. [8]. Sadiku, Matthew & Eze, Kelechi & Musa, Sarhan.
(2018). Data Mining in Healthcare. International
Future research should focus on improving Journal of Advances in Scientific Research and
interpretability, incorporating temporal models such as Engineering. 4. 90-92. 10.31695/IJASRE.2018.32881.
LSTMs or Transformers, and expanding datasets to include [9]. E. Tjoa and C. Guan, "A Survey on Explainable
greater demographic and geographic diversity. Additionally, Artificial Intelligence (XAI): Toward Medical XAI,"
integrating IoT (Internet of Things) sensors for real-time data in IEEE Transactions on Neural Networks and
collection could enhance prediction personalization. Learning Systems, vol. 32, no. 11, pp. 4793-4813,
Nov. 2021, doi: 10.1109/TNNLS.2020.3027314.
VII. CONCLUSION [10]. Mehrparvar, F. (2023). Obesity Level Dataset.
Kaggle. https://www.kaggle.com/datasets/fatemehm
This study explores the use of artificial intelligence for ehrparvar/obesity-levels.
obesity prevention, a rapidly growing global health concern. [11]. Begum, N. B. M., et al. (2024). Predicting Obesity
By comparing traditional models (logistic regression, KNN, Levels: Machine Learning Approaches Analyzing
Random Forest) with advanced approaches (XGBoost, Eating Habits and Lifestyle Factors.
MLP), we demonstrated that deep neural networks NeuroQuantology, 22(4), 305-313.https://www.n
outperform classical methods in multiclass classification. All euroquantology.com/open-
models were optimized using GridSearchCV, significantly access/PREDICTING%2BOBESITY%2BLEVELS.D
improving their performance, with XGBoost achieving an OI Number: 10.48047/nq.2024.22.4.nq24035.
accuracy of 96.9% and our optimized MLP reaching 98.05%.

IJISRT25APR1962 www.ijisrt.com 3199

Volume 10, Issue 4, April – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25apr1962
[12]. Alsareii, Saeed & Awais, Muhammad & Alamri, [24]. Kingma, Diederik & Ba, Jimmy. (2014). Adam: A
Abdulrahman & Alasmari, Mansour & Irfan, Method for Stochastic Optimization. International
Muhammad & Raza, Mohsin & Manzoor, Umer. Conference on Learning Representations.
(2023). Machine-Learning-Enabled Obesity Level [25]. Prechelt, Lutz. (2000). Early Stopping - But When?.
Prediction Through Electronic Health Records. Lecture Notes in Computer Science. 10.1007/3-540-
Computer Systems Science and Engineering. 46. 49430-8_3.
3715-3728. 10.32604/csse.2023.035687. [26]. M. Sajid et al., Deep learning approaches for
[13]. Helforoush Z, Sayyad H. Prediction and classification predicting obesity using dietary and physical activity
of obesity risk based on a hybrid metaheuristic patterns. Computers in Biology and Medicine, vol.
machine learning approach. Front Big Data. 2024 Sep 149, 105962, 2022. DOI: 10.48047/nq.2
30;7:1469981. doi: 10.3389/fdata.2024.1469981. 024.22.4.nq24035.
PMID: 39403430; PMCID: PMC11471553. [27]. Esteva A, Robicquet A, Ramsundar B, Kuleshov V,
[14]. Genc, A. C., & Arıcan, E. (2025). Obesity DePristo M, Chou K, Cui C, Corrado G, Thrun S,
classification: a comparative study of machine Dean J. A guide to deep learning in healthcare. Nat
learning models excluding weight and height data. Med. 2019 Jan;25(1):24-29. doi: 10.1038/s41591-018-
Revista da Associação Médica Brasileira, 71(1), 0316-z. Epub 2019 Jan 7. PMID: 30617335.
e20241282. https://doi.org/10.1590/1806-9282.2024 [28]. Brown, Williams & Noah, Asher & John, Ada.
1282. (2024). Optimizing Hyperparameters in Machine
[15]. Hosmer Jr, D. W., Lemeshow, S., & Sturdivant, R. X. Learning Models: Techniques and Best Practices.
(2013). Applied logistic regression. John Wiley & [29]. Samek, Wojciech & Wiegand, Thomas & Müller,
Sons. Klaus-Robert. (2017). Explainable Artificial
[16]. Breiman, L. (2001). Random forests. Machine Intelligence: Understanding, Visualizing and
Learning, 45(1), 5-32. https://doi.org/10.1023/A:1010 Interpreting Deep Learning Models. ITU Journal: ICT
933404324 Discoveries - Special Issue 1 - The Impact of
[17]. Chen, T., & Guestrin, C. (2016). XGBoost: A Artificial Intelligence (AI) on Communication
Scalable Tree Boosting System. Proceedings of the Networks and Services. 1. 1-10. 10.48550/ar
22nd ACM SIGKDD International Conference on Xiv.1708.08296.
Knowledge Discovery and Data Mining, 785–794.
https://doi.org/10.1145/2939672.2939785.
[18]. Radovanovic, Milos & Nanopoulos, Alexandros &
Ivanovic, Mirjana. (2009). Nearest neighbors in high-
dimensional data: The emergence and influence of
hubs. Proceedings of the 26th International
Conference On Machine Learning, ICML 2009. 382.
109. 10.1145/1553374.1553485.
[19]. Chaoyu Gong, Zhi-gang Su, Xinyi Zhang, Yang You,
Adaptive evidential K-NN classification: Integrating
neighborhood search and feature weighting,
Information Sciences, Volume 648, 2023,
119620,ISSN 0020-0255, https://doi.org/10.1016/j.ins.
2023.119620.
[20]. Zhang, Chao & Zhong, Peisi & Liu, Mei & Song,
Qingjun & Liang, Zhongyuan & Wang, Xiao. (2022).
Hybrid Metric K-Nearest Neighbor Algorithm and
Applications. Mathematical Problems in Engineering.
2022. 1-15. 10.1155/2022/8212546.
[21]. Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky,
Ilya Sutskever, and Ruslan Salakhutdinov. 2014.
Dropout: a simple way to prevent neural networks
from overfitting. J. Mach. Learn. Res. 15, 1 (January
2014), 1929–1958.
[22]. Sergey Ioffe and Christian Szegedy. 2015. Batch
normalization: accelerating deep network training by
reducing internal covariate shift. In Proceedings of the
32nd International Conference on International
Conference on Machine Learning - Volume 37
(ICML'15). JMLR.org, 448–456.
[23]. Raghu, Maithra & Zhang, Chiyuan & Kleinberg, Jon
& Bengio, Samy. (2019). Transfusion: Understanding
Transfer Learning with Applications to Medical
Imaging. 10.48550/arXiv.1902.07208.

IJISRT25APR1962 www.ijisrt.com 3200