1. Introduction

With the global growth of obesity cases in the past few decades, there has been an exponential rise in morbidity and death rates. According to the World Health Organization (WHO), obesity is a condition where excess fat accumulates in the human body, which results in severe health issues (Obesity and Overweight, 2021). Obesity results in 2.8 million deaths worldwide (Global health risks: mortality and burden of disease attributable to selected major risks, 2009), making it one of the prevalent risk factors for mortality. Individuals diagnosed with this disease generally tend to suffer from other health ailments such as stroke, diabetes, and sometimes various types of cancers ^[3], impacting their life expectancy rates.

Various types of obesity are prevalent (Miklishanskaya, Solomasova, & Mazur, 2021). The driving forces for obesity are unhealthy food habits, hereditary factors, physical inactivity, and lack of awareness regarding health issues. Obesity has caused a widespread effect on individuals of all age groups, and the number of children affected by the disease is increasing at an alarming rate worldwide.

There has been a surge in the percentage of obese children aged 6-7 years (Obesity, 2021). According to reports, around 39 million children under 5 years old were affected by this disease in 2020 (Obesity and Overweight, 2021). Children suffering from obesity generally tend to become obese adults (Whitaker, Wright, Pepe, Seidel, & Dietz, 1997), and they become susceptible to the adverse effects of adult obesity (Khan, et al., 2014) (Mollard, et al., 2014) ( Wing, et al., 2013) (Wadden, et al., 2011). Thus, there is huge interest among researchers in formulating an effective artificial intelligence (AI)-based solution for monitoring obesity. In recent years, numerous AI solutions have been proposed for obesity classification. Besides diagnosing the disease (Devi, Bai, & Nagarajan, 2020), identifying the relevant features contributing to obesity can also help in the early monitoring of the disease. Much attention has recently been attracted to the features and factors ^[12] globally and regionally (Filho, Gon\c{c}alves, de Lira, Ribeiro, Eickmann, & Lima, 2022).

In this study, the following two research questions are addressed:

· Research Question 1 (RQ1): With what degree of accuracy can obesity be categorized using a feature vector (FS) with 16 features?

· Research Question 2 (RQ2): Can a smaller feature vector ($\mathrm{FSS}$) classify the disease with an accuracy of over 90% compared to the accuracy obtained in RQ1?

To determine the degree of accuracy that may be attained, RQ1 examines the viability of categorizing obesity using a complete collection of 16 criteria. RQ2 further investigates the possibility of improving classification accuracy by using a smaller subset of features (FSS) while keeping a threshold accuracy of 90%.

For answering RQ1 and RQ2, three machine learning (ML) algorithms have been studied based on their performance: logistic regression (LR), support vector machine (SVM), and XGBoost (XGB). A publicly available dataset has been used for the study. A very deep comparative analysis based on the performance of each feature subset has been done. We believe this study will be a benchmark for understanding the correlation between different features leading to obesity.

2. Related Study

The performance of related AI methods is presented based on different evaluation metrics in Table 2. The performance metrics used for the analysis were accuracy (Ac), precision (Pr), recall (Re), F1 score (Fs), and Area under the Receiver Operating Characteristic Curve (ROC-AUC) score (Ra). The performance results in Table 2 have been analyzed for model selection for RQ1. For example, Table 3 shows results from an analysis of literary proposals concerning SVM (Celik, Guney, & Dengiz, 2021). For LR, for the classification of obesity, it can be observed that the accuracy (Ferdowsy, Rahi, Jabiullah, & Habib, 2021) is better than the accuracies in other studies (Pang, Forrest, L\^{e}-Scherban, & Masino, 2021), (Zheng & Ruggiero, 2017), (Thamrin, Arsyad, Kuswanto, Lawi, & Nasir, 2021). It can also be observed that for all the performance evaluation parameters for proposals using LR, one method presented the best results (Ferdowsy, Rahi, Jabiullah, & Habib, 2021).

For DT, one method (Ferdowsy, Rahi, Jabiullah, & Habib, 2021) reflects the best results, whereas for RF, another method (Dugan, Mukhopadhyay , Carroll, & Downs, 2015) exhibits the best accuracy, but another method (Singh & Tawfik, 2020) presents the best Pr and Re values. Some authors (Zheng & Ruggiero, 2017) claimed the best accuracy in a set of proposals using KNN algorithms. For the classification of obesity using NN, one proposal (Celik, Guney, & Dengiz, 2021) claims the best accuracy.

XGB was previously used to classify obesity. This motivated us to implement XGB in this study. Although the method (Pang, Forrest, L\^{e}-Scherban, & Masino, 2021) does not exhibit very impressive accuracy or other performance metrics, the implementation of XGB presents better results in comparison to others. From the analysis result in Table 2, the best values of accuracies were generated by SVM (Celik, Guney, & Dengiz, 2021) (97.8% and 97.09%) (Ferdowsy, Rahi, Jabiullah, & Habib, 2021) for the implementation of LR. Hence, we used these two algorithms and the XGB model for this study.

3. RQ1

The detection framework for RQ1 is shown in Fig. 1. Pre-processing mechanisms were carried out to improve the accuracy and performance of the models. Pre-processing steps included cleaning and organizing data to increase the model performance.

Fig. 1. Obesity detection framework used in both FS and FSS.

The correlation among the numerical features of the raw dataset is shown in Fig. 2. The attributes with missing values in the dataset were filled using the median imputation method. The attributes with less than two categorical features were encoded using label encoder techniques, and the ones with more than two were encoded using one hot encoder. The outliers in the dataset were treated using three standard deviation techniques.

Fig. 2. Graphical representation of correlation of features using heat map.

After applying all data pre-processing techniques, before training the ML model, we split the dataset to increase the efficiency of the trained model. We split the data into training and testing sets using K-fold cross-validation, leave-P-out cross-validation, leave-one-out cross-validation, holdout cross-validation, and nested cross-validation, which ensured independent training and testing on the dataset for better model performance with unseen data. The XGB, SVC, and LR algorithms were trained on a complete feature set using the sci-kit-learn library (Pedregosa, et al., 2011).

In RQ2, we selected an $\mathrm{FSS}$ out of the 16 original $\mathrm{FS}$s. The $\mathrm{FSS}$ selection has been done by applying the SelectKBest algorithm on the FS and sorting the rankings. Table 3 presents the performance scores of each feature using the SelectKBest algorithm. This performance score was used to evaluate the influence of each feature on obesity individually. The complete workflow is presented in Fig. 3.

Fig. 3. High-level diagram for RQ2.

After sorting the 16 features' correlation importance values, the 8 highest-ranking features were selected for further processing. The feature identifier assigns an ``id'' to the respective feature for ease of analysis. The four FSSs were selected from Table 3 using Eq. (1).

where

It is interesting to note that Eq. (1) uses $FI_{id}$ with $1\leq id\leq 8,$ as discussed earlier, because the eight highest-ranking features have been used. The four $FSS$s generated are mentioned in Table 4 with their respective features.

5. Experimental Results

Table 5 presents the performance analysis results of the shortlisted ML algorithms in terms of standard implementations, hyper-parameter tuned implementations, and their comparison results. It can be observed that XGB presents the best accuracy among the three algorithms. The two ML algorithms, SVC and LR, exhibited the best results, whereas the XGB results were not very impressive. But, in this implementation, we observed that XGB produces better results than the other two ML algorithms. Moreover, the performance metrics are even better for all ML implementations, including XGB. We also present the increase or decrease in performance of the tuned versions for the standard implementation of the ML algorithms.

6. Conclusion

In this study, we have addressed two research questions about the effect of a complete feature vector generated from the parameters influencing obesity and a much smaller feature vector's influence on obesity. Hence, of all the parameters that are responsible for obesity, it can be concluded that concentrating on a smaller parameter set reduces obesity by about 91% than when all the parameters are taken into consideration.

The study’s objective was to investigate whether a smaller set of parameters can reduce or prevent obesity to the same extent as that provided by all the parameters. The experimental results reflect that a smaller subset of only four parameters like weight, age, gender, and SCC influences obesity levels by almost 91% compared to the influence subjected by the complete set of 16 parameters. The study targeted individuals who feel inert while addressing all related factors for obesity control and who may be more comfortable with addressing only a few factors to some extent.