Machine-Learning-Based Prediction of Obesity

Siyuan Chen

International Business School, Henan University, Zhengzhou, China

Keywords: Machine Learning, Obesity Prediction, Neural Network.

Abstract: Obesity is a common phenomenon today. It is a chronic metabolic disease caused by excessive fat

accumulation and is the result of the interaction of multiple factors such as genetics and environment. As

machine learning is widely used in various fields, obesity data is processed by using machine learning methods

to obtain fitting models, to realize the prediction of obesity and to determine the main causes of obesity. The

main contents of this paper include: (1) using the obesity data provided by the UCI data set as the research

object, using a series of preprocessing data in Python language. (2) Establish a machine learning model and

import the data to generate bar charts and other graphs that reflect the relevant results. (3) Results analysis, to

Determine which factors are more closely related to obesity and evaluate the performances of the models. The

comprehensive analysis of accuracy, recall rate, precision rate and other indicators finally obtains the best

prediction effect of the GBDT algorithm, which can effectively predict obesity.

INTRODUCTION

Nowadays, people's living standards are constantly

improving with the development of society, and their

diet is becoming more and more diversified. While

enjoying these conveniences, there are also some

potential risks. The pace of society is accelerating,

and people are under pressure from all sides. With this

situation, people's health constantly produces various

problems. Obesity is a common phenomenon, it is

caused by genetic and environmental factors such as

chronic metabolic diseases, the Chinese Residents

Nutrition and Chronic Condition Report (2020),

according to more than half of adults overweight /

obesity, 6~17, children under 6 and adolescents

overweight / obesity rate reached 19.0% and 10.4%

respectively (Liu, 2021). Obesity not only affects the

external appearance but also leads to various other

diseases. At the same time, to prevent and treat

obesity, people need to spend more time and money.

Therefore, the prevention and treatment of obesity is

of great significance to prevent chronic diseases and

reduce the personal financial burden. With artificial

intelligence developing so quickly, machine learning

has become widely applied in many different fields,

the medical field is also a big aspect of machine

learning application, through the use of the machine

https://orcid.org/0009-0006-3168-8559

learning function of existing data analysis, which can

conclude that the relevant data about obesity to

determine what are the main causes of obesity.

From the point of the current development

situation, although there have been some machine

learning methods applied to disease prediction, for the

prediction of obesity is relatively lack of

comprehensive and perfect research, so this paper

based on machine learning from a variety of

algorithms to find out the good prediction of obesity

prediction model, to realize the timely management

and treatment of obesity.

2 METHOD

This chapter introduces six machine learning models

of logical regression, decision tree, random forest,

GBDT, XGBoost and deep learning network (DNN),

and presents the experimental data set and the

unbalanced learning methods, and expounds the main

content and direction of this experiment.

2.1 Logistic Regression

Among the linear regression models is logistic

regression, which uses data from independent

Chen, S.

Machine-Learning-Based Prediction of Obesity.

DOI: 10.5220/0012916000004508

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2024), pages 153-158

ISBN: 978-989-758-713-9

153

variables as input to forecast the chance of a desired

result. Logistic regression has similarities to the

principle of multiple linear regression, which first

determines the best-fitting regression line to represent

the connection between the independent variable (x)

and the dependent variable (y), the regression line

(Wang, 2022). The model form is related to the linear

equation. If the independent variable is a data set, the

equation can be shown as a matrix:

baxy +=

bxaxaxaxaz

+⋅++⋅+⋅+⋅=

)()()3()3()2()2()1()1(



(1)

Logistic regression functions the linear equation

corresponding to a state p, determining the size of the

dependent variable based on the values of p and 1-p.

The dependent variables of logistic regression can be

dichotomous or multiple classifications, and the

independent variables can be continuous or discrete.

There are three types of logistic regression:

ordinal, multinomial, and binary.

The results of this paper are divided into obesity

and non-obesity after data processing, so binomial

logistic regression is adopted. The dependent variable

of binomial logistic regression is essentially a

dichotomy method, that is, there are only 0 or 1 results,

and the probability distribution is as follows:

xYP

⋅⋅

⋅

)|0(

)|1( ，

(2)

2.2 Decision Tree

The decision tree technique is a prediction technique

for creating target variables or a categorization

scheme based on several variables. This algorithm

can effectively handle large data sets. Common uses

of the decision tree model include variable selection,

evaluation of variable importance, processing of

missing values, and prediction (Song, 2015). The

duality of the results enables a good application of

decision trees for obesity prediction in this

experimental study.

The main parts of the decision tree model are the

nodes and branches and the construction of the model

includes splitting, stopping and pruning (Song, 2015).

The nodes of the decision tree can be divided into

three types: (1) root nodes, also called decision nodes.

(2) Internal nodes. (3) The decision tree's ultimate

outcome is represented by the leaf node, sometimes

referred to as the end node.

The decision tree is a continuous model that

combines a series of tests and compares the feature

values in each test to the threshold value(Navada,

Ansari, Patil, 2011). Each node in the decision tree

corresponds to an analysis of data properties, The

decision tree model links the dataset's observations to

the conclusions(Sharma, Kumar, 2016) of the

pertinent target values, with each branch denoting the

analysis's findings.

2.3 Random Forest

The classification and regression tree model is further

improved by the random forest technique, which is

composed of a large number of decision trees created

by randomization, which can be used for prediction

once constructed. Because the validity of the decision

tree for binary classification applies to the prediction

of obesity, the random forest model was adopted by

this experiment. The average of the outputs of a

random forest with several decision trees is

aggregated into a single output of reference (Rigatti,

2017). Formally, a model made up of several random

base regression trees is called a random forest {rn (x,

Θ m, Dn), m 1}, where Θ 1, Θ 2,... Is the output of the

random variable, Θ. Combining these random trees

forms an aggregate regression estimate of research

(Biau, 2016 ). The formula is as follows:

)()(

[]

nnn

DXDX ,,, ΘΕ=

(3)

Based on each decision tree produces a result

based on the input data; and the final output of the

random forest is the paper obtained after integrating

the multiple results (Liu, 2014).

The advantage of random forest is that it can find

the interaction between the predicted variables and

the non-linear relationship, but it is difficult to judge

which variables have a greater impact on the

prediction results.

2.4 GBDT Model

GBDT, whose full name is a gradient-lifting decision

tree, is an iterative decision tree algorithm that is also

applicable to the binary characteristics of this study.

To obtain the ultimate prediction outcome, the model

aggregates the outcomes of several decision trees.

The next weak classifier fits the residual function of

the predicted value, which is the difference between

the predicted value and the true value, and the

principle adds the results of all the weak classifiers to

the predicted value. The decision tree is the common

learner in the GBDT model, which is an integrated

learning model (Zhang, 2021 ).

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

154

2.5 XGBoost Model

XGBoost is an improved model of the gradient

boosting algorithm and a machine learning algorithm

implemented under the Gradient Boosting framework.

The basic component of the XGBoost model is the

decision tree, which has good results for dichotomous

data analysis, so it is applied to the prediction of

obesity in this experiment. There is an order between

them: the latter decision tree will be combined with

the prediction results of the previous decision tree,

that is, take the analysis error of the previous tree into

account, which increases the proportion of the error

between the previous samples in the subsequent

prediction, thus improving the accuracy and

scientificity of the model prediction. According to the

model principle, the prediction calculation formula is

obtained as follows:

)(

)1()(

fyy

−

∧∧

(4)

That is the predicted value of the first tree for the

sample t = the predicted value of the first a-1 tree for

the first t-1 sample + the first tree for the sample t.

The objective function of the model prediction result

is obtained as follows:

)()

)(

Ω+=Ι



∧

(5)

2.6 Deep Neural Network

The DNN is designed to predict the data set. The

model construction of the neural network includes

input, output and three hidden layers. Through the

excitation function tanh () in the hidden layer, the

nonlinearity of the data set can expand the expression

ability of the neural network (Figure 1). The tanh

function is centered on zero, the gradient is steeper

and not limited to one direction, and it is overall

superior to the sigmoid function (Sharma, 2017).

Figure 1: Deep Neural network model (Photo/Picture credit :

Original).

Deep neural networks have the good nonlinear

fitting ability but require large-scale datasets for

training, otherwise overfitting may occur.

3 RESULTS AND DISCUSSION

3.1 Pro-Processing

3.1.1 Distribution of the Data Sets

As shown in pie figure 2, the proportion of obesity in

the surveyed subjects in this experimental dataset is

not balanced compared with other resultPs, so it is

speculated in the dataset.

Figure 2: Schematic representation of the obesity ratio

among the survey subjects (Photo/Picture credit : Original).

3.1.2 Unbalanced Processing Algorithm

The SMOTE algorithm, namely the Synthetic

Minority Oversampling Technique, also known as the

synthetic minority oversampling technology, is also

an improved random oversampling method that

generates new artificial minority instances by

interpolating in several instances located together

(Yang, 2021). It is based on sampling data from a few

classes by connecting random data points (Elreedy,

2019). SMOTE can alleviate the overfitting

phenomenon of random oversampling by inserting

synthetic instances in a new position, but it still has

two disadvantages, one is that it can spread the noise,

and the other is that in the SMOTE algorithm, all

instances share the same global neighborhood

parameters, which will ignore the distribution

characteristics, resulting in poor classification effect.

The SMOTE algorithm is very robust, but in the case

of chaotic sample data and some scattered samples, it

will mechanically generate new points, which will

Machine-Learning-Based Prediction of Obesity

155

become noise points and affect the classification

performance of the model (Yang, 2021).

3.1.3 Hyperparameter Setting

The experiment code was run using Python language

through Pycharm compilation software. Where the

main hyperparameter package Epochs, batch_size.

Epochs refer to the number of times the training

set is fully trained when training the neural network,

also called the process of neural network forward

propagation and backpropagation.

The amount of samples that the neural network

chooses to use in a single training session is known as

the batch_size.

In this experiment, the results of different

parameters were analyzed for multiple times, so that

the set epochs value is 200 and the base _ size value

is 70, and the comparison results between models

obtained with the above values are the most balanced.

3.2 Evaluation Indicators

(1) Accuracy

Model accuracy (Accuracy) is one of the main

indicators to evaluate the model performance, and the

following is its calculating formula:

FNFPTNTP

TNTP

Accuracy

+++

(6)

(2) Precision Rate

Accuracy is the proportion of samples with positive

model prediction results in all positive samples. The

following is the calculating formula:

FPTP

ecision

=Pr

(7)

(3) Recall

The percentage of samples that the model predicts is

referred to as the recall rate. The following is the

calculating formula:

FNTP

call

=Re

(8)

When evaluating the model with recall rate, a

model's recall rate is low; when the precision rate is

low, its recall rate is relatively high.

(4) F1-score

The F1-score, also known as the F1 score, is an

indicator that combines precision and recall rate to

produce average results. The formula is:

RecallPrecisi

RecallPrecisi2

××

(9)

Using the F1 score can avoid the problem that the

data value is relatively close and it is difficult to

choose and can find the model with better

performance more efficiently.

(5) AUC

The meaning of AUC is the area under the curve and

the axis, here the curve is the ROC curve. The area

under the curve can be more intuitively judged by the

model performance.

3.3 Analysis of the Indicators

(1) Precision (accuracy)

First, observe the prediction effect of each model on

the data set through accuracy, and the comparison

figure is as follows (Figure 3 and Table 1):

Figure 3: The prediction accuracy of each model

(Photo/Picture credit : Original).

Table 1: The prediction accuracy of each model.

Metho

accurac

istic Re

ression 0.87

Decision Tree 0.91

Random Fores

0.915

GBDT 0.919

XGBoos

0.92

DNN 0.897

As illustrated in Figure 3, the logistic regression

model has the lowest prediction accuracy, only

around 87%, followed by DNN, and the models with

higher accuracy were GBDT and XGBoost.

Because the result of the data prediction is binary,

the model based on the tree model prediction accuracy

is high, the GBDT in predicting new samples, each

tree will produce an output value, the output value

superposition, to get the final prediction value, for

each tree training is the difference is the true result of

the next tree prediction, for the experimental data set

its data difference is large, so the test GBDT model

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

156

prediction accuracy is high. The XGBoost model is

similar to the GBDT principle, and the result of the

previous tree during training affects the generation of

the latter tree, so the prediction accuracy of this data

set is relatively high.

(2) Accuracy Rate (precision) and Recall Rate (recall)

The precision and recall of each experimental data set

are shown in the figure below (Figure 4 and Table 2

Table 3):

Figure

Prediction accuracy and recall rate

each

model(Photo/Picture credit : Original).

Table 2: Prediction accuracy of each model.

Metho

Precision0

istic Re

ression 0.723

Decision Tree 0.822

Random Fores

0.806

GBDT 0.817

XGBoos

0.818

DNN 0.84

Table 3: Recall rate of each model.

Metho

Recall0

istic Re

ression 0.824

Decision Tree 0.84

Random Fores

0.89

GBDT 0.905

XGBoos

0.904

DNN 0.75

As demonstrated in Figure 4, the relationship

between precision and recall rate. Among them, the

comparison of the DNN model is the most obvious,

with the unbalanced phenomenon of the data set itself.

Therefore, the two indicators of the model that cannot

make gradient adjustments in the prediction are quite

different, while the models that can make gradient

adjustments, such as GBDT and XGBoost, are

relatively balanced.

(3) F1-Score

Considering that the analysis results combining

precision and recall rate are not intuitive enough, the

F1 score is used to reflect the performance of each

model more intuitively. The pairs between models are

such as the following Figure 5:

Figure 5: The predicted F1 scores for each model

(Photo/Picture credit: Original).

Table 4: The predicted F1 scores for each model.

Metho

F1-score

istic Re

ression 0.909

Decision Tree 0.939

Random Fores

0.941

GBDT 0.946

XGBoos

0.95

DNN 0.93

Can intuitively seen from Figure 5, the performance

of the GBDT and XGBoost models is better than other

models, the reason speculated that there may be two:

one because the two models’ training mode is more

superior, and can constantly adjust data gradient

because the experimental data set itself is unbalanced,

the two models can more effectively improve the

imbalance phenomenon caused by bad results (Table

4). Therefore, the GBDT and XGBoost models have

a better performance.

Machine-Learning-Based Prediction of Obesity

157

4 CONCLUSION

This study through the UCI data on obesity prediction

data research model, then through the model of the

data visualization analysis, and by combining the

model performance evaluation through the

parameters of accuracy, precision, recall rate, and F1

score, finds more accurate obesity prediction model,

the results show that GBDT and XGBoost model in

the data prediction related indicators are high, the

fitting effect is good, the future can through the two

machine learning algorithms to predict obesity and

related diseases. At the same time can be obtained

from the characteristics of age, family whether

someone with obesity, whether often eats calorie food

and two meals between other food frequencies these

four characteristics are associated with obesity, the

doctor in determining whether obese patients can be

according to these characteristics to further determine

whether obesity.

However, there are certain limitations and areas

for improvement in this study. The relative lack of

data volume in this experiment leads to insufficient

analysis of other features. If a large amount of

relevant data can be obtained, it will be more

favorable for the model prediction.

The experimental data acquisition range is small,

and the prediction results are less applicable. The

survey results of the data set come from a small range

of acquisitions, and it is controversial in its

universality. The scope of data collection should be

increased to make it have better universality.

REFERENCES

Biau G. (2012). Analysis of a random forests model. The

Journal of Machine Learning Research, 13: 1063-1095.

Elreedy D, Atiya A F. (2019). A comprehensive analysis of

synthetic minority oversampling technique (SMOTE)

for handling class imbalance. Information Sciences, 505:

32-64.

Liu Y. The Report on Nutrition and Chronic Diseases of

Chinese Residents (2020) was released. Agricultural

Products Market Weekly, 2021 (2): 58-59.

Liu Y. (2014). Random forest algorithm in big data

environment. Computer modelling & new technologies,

18(12A): 147-151.

Navada A, Ansari A N, Patil S, et al. (2011). Overview of

use of decision tree algorithms in machine learning,

2011 IEEE control and system graduate research

colloquium, 2011: 37-42.

Rigatti S J. (2017). Random forest. Journal of Insurance

Medicine, 47(1): 31-39.

Sharma S, Sharma S, Athaiya A. (2017). Activation

functions in neural networks. Towards Data Sci, 6(12):

310-316.

Sharma H, Kumar S. (2016). A survey on decision tree

algorithms of classification in data mining.

International Journal of Science and Research (IJSR),

5(4): 2094-2097.

Song YY, Lu Y. (2015). Decision tree methods:

applications for classification and prediction. Shanghai

Arch Psychiatry, 27(2):130-5.

Wang X. (2023). Machine learning based prediction model

for heart disease. Southwestern University.

Yang D. (2021). Research and application of unbalanced

data processing algorithms. Jiangsu: Jiangsu University

of Science and Technology.

Yu C S, Lin Y J, Lin C H, et al. (2020). Predicting metabolic

syndrome with machine learning models using a

decision tree algorithm: Retrospective cohort study.

JMIR medical informatics, 8(3): e17110.

Zhang W, Yu J, Zhao A, et al. (2021). Predictive model of

cooling load for ice storage air-conditioning system by

using GBDT. Energy Reports, 7: 1588-1597.

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

158