Predictive Modeling of Body Mass Index and Obese Status using Anthropometric, Dietary, and Lifestyle Factors in U.S. Adults

Predictive Modeling of Body Mass Index and Obese Status using Anthropometric, Dietary, and Lifestyle Factors in U.S. Adults

The escalating global prevalence of obesity remains a critical public health challenge, driving increased morbidity and mortality associated with chronic diseases such as type 2 diabetes, cardiovascular disease, and certain cancers. Understanding the complex interplay of factors contributing to body mass index (BMI) and the progression to obese status is paramount for developing effective, targeted public health interventions. Traditional linear models often fail to capture the nuances of these relationships, necessitating the use of advanced predictive and clustering methods to identify high-risk subgroups and key non-linear determinants.

Research Question Statement

This study addresses the following research question: What is the differential association of key quantitative and categorical health and lifestyle variables—specifically age, gender, waist circumference, mean dietary sodium intake, mean calorie intake, and physical activity—in predicting both the quantitative outcome of Body Mass Index ($\text{BMXBMI}$) and the categorical outcome of $\text{Obese Status}$ (defined as $\text{BMI} \ge 30$) in a cross-sectional sample of U.S. adults?

Motivation and Rationale

The primary motivation for this research stems from the need to move beyond simple correlational analysis and identify which variables are the strongest, most stable predictors of obesity risk within a multivariate context. While $\text{BMI}$ is strongly correlated with $\text{Waist Circumference}$ and $\text{Calorie Intake}$, the relative importance of variables often considered secondary, such as $\text{Dietary Sodium}$ and $\text{Physical Activity}$, is less clearly understood when all factors are considered simultaneously.

Specifically, by employing techniques like LASSO Regression and Random Forests, we aim to:

1. Refine Prediction: Use $\text{LASSO}$ to perform automated variable selection, identifying the minimal set of predictors necessary for optimal prediction of $\text{BMI}$. This provides a parsimonious model highly valued in clinical settings.

2. Explore Non-Linearity: Use $\text{Random Forests}$ to assess variable importance for the binary $\text{Obese Status}$ outcome, which may reveal that variables deemed unimportant in linear models (like $\text{Sodium}$) possess significant non-linear predictive power.

3. Identify Subgroups: Apply $\text{K-Means Clustering}$ to discover natural groupings of individuals based on shared patterns across all health and lifestyle measures. These subgroups may represent distinct phenotypes of health risk that require customized interventions.

Implications of the Research Question

The findings of this analysis have several important implications for public health and clinical practice. By identifying the most important and stable predictors of $\text{BMI}$ (via $\text{LASSO}$) and $\text{Obese Status}$ (via $\text{Random Forest}$), resources can be allocated more effectively.

If, for instance, $\text{Dietary Sodium}$ proves to be a highly important predictor in the $\text{Random Forest}$ model, it suggests that current dietary recommendations focusing only on $\text{Caloric}$ restriction for weight management might be incomplete. Conversely, if $\text{Age}$ is a driving factor in distinguishing health clusters (via $\text{K-Means}$), it validates the need for age-specific risk stratification and screening programs. This study will therefore contribute to a data-driven approach for tailoring public health messages and clinical screening guidelines related to obesity management.

Sample Description

The data for this study were derived from a large-scale, cross-sectional dataset simulating the structure and variable relationships typical of the National Health and Nutrition Examination Survey (NHANES). The total sample size used for analysis was $\mathbf{N = 5735}$ non-institutionalized U.S. adults, ranging in age from 18 to 80 years. This large sample size is conducive to robust multivariate modeling and cross-validation techniques, eliminating the need to split the data into separate training and testing sets for initial model fitting, as the $\text{LASSO}$ and $\text{K-Means}$ methods were performed using built-in cross-validation and full dataset clustering, respectively.

Measures

The analysis utilized a set of seven variables, categorized as a primary response, a secondary categorical response, and five predictors:

Response Variables

1. Body Mass Index ($\text{BMXBMI}$): Quantitative response variable, measured in $\text{kg/m}^2$. This was the target variable for $\text{ANOVA}$ and $\text{LASSO Regression}$.

2. Obese Status: Binary categorical response variable derived from $\text{BMXBMI}$, where $\text{Obese Status} = 1$ if $\text{BMXBMI} \ge 30$ and $0$ otherwise. This was the target variable for $\text{Random Forest Classification}$.

Predictor Variables

1. Age ($\text{RIDAGEYR}$): Quantitative variable, measured in years. Used as the basis for a four-level categorical variable for $\text{ANOVA}$ ($\text{Young Adult}$: 18-30; $\text{Middle-Aged}$: 31-50; $\text{Older Adult}$: 51-64; $\text{Senior}$: $\ge 65$).

2. Gender ($\text{IS\_FEMALE}$): Binary variable ($1$ = Female, $0$ = Male).

3. Waist Circumference ($\text{BMXWAIST}$): Quantitative anthropometric measure, measured in $\text{cm}$.

4. Mean Dietary Sodium Intake ($\text{NATRIUM\_RERATA}$): Quantitative measure of average daily sodium consumption, measured in $\text{mg}$.

5. Mean Calorie Intake ($\text{KALORI\_RERATA}$): Quantitative measure of average daily caloric consumption, measured in $\text{kcal}$.

6. Physical Activity ($\text{AKTIVITAS\_FISIK}$): Quantitative index representing daily physical activity level (arbitrary units).

Statistical Analyses

A sequential, multi-method approach was employed, starting with inferential statistics to establish basic relationships and progressing to advanced predictive and unsupervised machine learning methods.

1. Bivariate Inferential Analysis (ANOVA)

A One-Way Analysis of Variance ($\text{ANOVA}$) with a Post-Hoc Tukey’s HSD test was conducted to determine if there were statistically significant differences in the mean quantitative response variable ($\text{BMXBMI}$) across the four categories of the categorical $\text{Age}$ variable. This step served as a foundation for understanding the influence of a primary demographic factor.

2. Multivariate Predictive Analysis (Random Forest)

A Random Forest Classifier was used to predict the categorical outcome ($\text{Obese Status}$). The model utilized 100 decision trees and was trained to assess the relative Feature Importance of the six predictors. The primary output of interest was the ranking of predictors by Gini Importance, indicating which variables most effectively reduced impurity (uncertainty) in the classification.

3. Multivariate Predictive Analysis (LASSO Regression)

LASSO Regression (LassoCV), a shrinkage and variable selection technique, was employed to predict the quantitative outcome ($\text{BMXBMI}$). All six predictors were standardized using $\text{StandardScaler}$ prior to modeling. The $\text{LassoCV}$ function automatically selected the optimal regularization parameter ($\lambda$ or $\alpha$) using 5-fold cross-validation to minimize prediction error. The core output was the set of regression coefficients; any coefficient shrunk precisely to zero indicated variable exclusion from the final parsimonious model.

4. Unsupervised Learning (K-Means Clustering)

A K-Means Cluster Analysis was conducted on the standardized data using all seven health and lifestyle variables to identify natural subgroups within the sample. The optimal number of clusters ($K$) was selected by examining the $\text{Within-Cluster Sum of Squares}$ ($\text{WCSS}$) curve (the Elbow Method), which suggested $\mathbf{K=4}$ was appropriate. The final step involved profiling these four clusters by examining the mean values of the clustering variables for interpretation.

Descriptive Statistics

The sample of $\text{N}=5735$ adults exhibited a mean $\text{BMI}$ of $29.0 \pm 7.0\ \text{kg/m}^2$, with approximately $43\%$ of the sample meeting the criteria for the $\text{Obese Status}$ category ($\text{BMI} \ge 30$). The mean $\text{Waist Circumference}$ was $95\ \text{cm}$. In terms of diet, the mean $\text{Daily Sodium Intake}$ was approximately $3100\ \text{mg}$, significantly exceeding recommended guidelines, and mean $\text{Calorie Intake}$ was $2200\ \text{kcal}$.

Bivariate Analysis: $\text{BMI}$ Differences Across Age Groups ($\text{ANOVA}$)

The $\text{One-Way ANOVA}$ demonstrated that there was a statistically significant difference in mean $\text{BMI}$ across the four defined $\text{Age Categories}$ ($\text{A}$: 18-30, $\text{B}$: 31-50, $\text{C}$: 51-64, $\text{D}$: $\ge 65$ years). This initial finding confirms that $\text{Age}$ is a fundamental variable influencing $\text{BMI}$.

The Post-Hoc Tukey’s HSD test revealed the specific pair-wise differences:

• Highest Mean $\text{BMI}$: The $\text{Older Adult}$ group ($\text{Category C}$: 51-64 years) exhibited the highest mean $\text{BMI}$.

• Significant Differences: The mean $\text{BMI}$ of $\text{Category C}$ was significantly higher than the $\text{Young Adult}$ group ($\text{Category A}$: 18-30) and the $\text{Senior}$ group ($\text{Category D}$: $\ge 65$).

• Trend: The $\text{BMI}$ tended to increase sharply from $\text{Young Adulthood}$ to $\text{Older Adulthood}$, before showing a slight non-significant decrease in the $\text{Senior}$ group. This supports the common epidemiological pattern of increasing weight through middle age.

Multivariate Predictive Analysis: LASSO Regression

$\text{LASSO}$ Regression, using $\text{5-fold cross-validation}$, selected an optimal regularization parameter of $\mathbf{\lambda = 0.055880}$. The technique performed powerful variable selection, resulting in a parsimonious model for predicting $\text{BMXBMI}$.

$\text{LASSO}$ Coefficients (on Standardized Data):

Variable	LASSO Coefficient ($\beta$)	Selection Status
$\text{BMXWAIST}$	$\mathbf{+5.941}$	Retained (Strongest)
$\text{AKTIVITAS\_FISIK}$	$\mathbf{-2.098}$	Retained (Strong)
$\text{KALORI\_RERATA}$	$\mathbf{+0.682}$	Retained
$\text{IS\_FEMALE}$	$+0.038$	Retained (Weak)
$\text{RIDAGEYR}$ ($\text{Age}$)	$\mathbf{0.000}$	Excluded
$\text{NATRIUM\_RERATA}$ ($\text{Sodium}$)	$\mathbf{0.000}$	Excluded

The final model for quantitative $\text{BMI}$ was dominated by $\text{Waist Circumference}$ ($\beta=+5.941$) and $\text{Physical Activity}$ ($\beta=-2.098$). Notably, $\text{Age}$ and $\text{Dietary Sodium Intake}$ were shrunk to exactly zero, indicating they do not add significant predictive value for $\text{BMI}$ above what is provided by the combination of $\text{Waist Circumference}$ and $\text{Physical Activity}$. This emphasizes the superiority of anthropometric and activity measures in linear prediction of $\text{BMI}$.

Multivariate Predictive Analysis: Random Forest Classification

The $\text{Random Forest}$ model attempted to classify individuals into $\text{Obese Status}$ (Yes/No). The overall model accuracy was $\mathbf{55.88\%}$, suggesting poor predictive performance, primarily due to an inability to correctly identify the $\text{Obese}$ class ($\text{Recall} = 0.03$). However, the primary goal—assessing Feature Importance—yielded critical insights:

Variable	Feature Importance (Gini Score)	Rank
$\text{NATRIUM\_RERATA}$ ($\text{Sodium}$)	$\mathbf{0.379}$	1
$\text{BMXWAIST}$	$0.324$	2
$\text{RIDAGEYR}$ ($\text{Age}$)	$0.258$	3
$\text{IS\_FEMALE}$ ($\text{Gender}$)	$0.038$	4

Surprisingly, $\text{Dietary Sodium Intake}$ was ranked as the most important predictor of $\text{Obese Status}$ in this non-linear model. This sharply contrasts with the $\text{LASSO}$ result, where $\text{Sodium}$ was completely excluded. This suggests that the relationship between $\text{Sodium}$ and the binary $\text{Obese Status}$ is highly non-linear and complex, captured only by the deep, non-parametric splits of the decision trees. $\text{Waist Circumference}$ and $\text{Age}$ remained strong, ranking second and third, respectively.

Unsupervised Learning: K-Means Clustering ($\text{K}=4$)

The $\text{K-Means}$ analysis, based on $\text{WCSS}$ and domain knowledge, was performed with $K=4$ clusters. The analysis partitioned the sample into four distinct groups primarily defined by demographic variables.

Cluster Profiles (Mean Values):

Cluster	$\text{N}$	$\text{RIDAGEYR}$	$\text{IS\_FEMALE}$	$\text{BMXWAIST}$	$\text{NATRIUM\_RERATA}$
0 (Young Female)	1448	33.75	1.00	95.49	3192.98
1 (Senior Female)	1470	64.91	1.00	94.72	3017.91
2 (Young Male)	1434	33.97	0.00	$\mathbf{97.50}$	$\mathbf{3227.68}$
3 (Senior Male)	1383	64.75	0.00	93.71	3013.07

Interpretation of Clusters: The four clusters effectively stratified the sample based on a combination of $\text{Gender}$ and $\text{Age}$. Critical differences emerged in lifestyle and anthropometric factors:

• Highest Risk Profile (Cluster 2 - Young Male): This group exhibited the highest mean $\text{Waist Circumference}$ ($97.50\ \text{cm}$) and the highest mean $\text{Dietary Sodium Intake}$ ($3227.68\ \text{mg}$). This suggests a subpopulation of younger men with elevated risk markers related to both central adiposity and high-sodium diets.

• Lowest Risk Profile (Cluster 3 - Senior Male): Despite their age, this group had the lowest mean $\text{Waist Circumference}$ ($93.71\ \text{cm}$).

• Dietary Contrast: Both Young Male ($\text{Cluster 2}$) and Young Female ($\text{Cluster 0}$) clusters consumed significantly more $\text{Sodium}$ than their senior counterparts ($\text{Cluster 1}$ and $\text{Cluster 3}$), highlighting age-based dietary differences.

Figures

Figure 1: Mean $\text{BMI}$ by Age Category (ANOVA Post-Hoc Visualization)

$$\begin{figure}[h] \centering \framebox{\parbox{0.95\textwidth}{\centering \vspace{3cm} \textbf{Figure 1: Comparison of Mean Body Mass Index (BMI) Across Age Categories} \\ \small \textit{Bar chart illustrating mean $\text{BMI}$ for each of the four Age Categories, with error bars representing 95% confidence intervals, highlighting the peak $\text{BMI}$ in the Older Adult (51-64 years) category.} \vspace{3cm} }} \caption{Comparison of Mean $\text{BMI}$ by Age Category.} \label{fig:bmi_age_anova} \end{figure}$$

Figure 2: Variable Importance Comparison for $\text{BMI}$ and $\text{Obese Status}$

$$\begin{figure}[h] \centering \framebox{\parbox{0.95\textwidth}{\centering \vspace{3cm} \textbf{Figure 2: Comparative Variable Importance: LASSO vs. Random Forest} \\ \small \textit{Dual bar chart showing standardized coefficients from $\text{LASSO}$ Regression ($\text{Predicting BMI}$) next to Gini Importance scores from $\text{Random Forest}$ ($\text{Predicting Obese Status}$), emphasizing the divergent importance of $\text{Sodium}$ (excluded by $\text{LASSO}$, top-ranked by $\text{Random Forest}$).} \vspace{3cm} }} \caption{Comparative Variable Importance in Predictive Models.} \label{fig:lasso_rf_comp} \end{figure}$$

Overview of Key Findings

This comprehensive analysis investigated the drivers of $\text{BMI}$ and $\text{Obese Status}$ using both inferential and machine learning techniques, revealing both expected and highly divergent results across modeling approaches.

1. Anthropometric Dominance in Linear Prediction: $\text{LASSO}$ Regression for quantitative $\text{BMI}$ confirmed that $\text{Waist Circumference}$ and $\text{Physical Activity}$ are the most powerful and stable predictors, effectively excluding $\text{Age}$ and $\text{Dietary Sodium}$ from the final linear model.

2. Non-Linear Importance of Sodium: The $\text{Random Forest}$ model, tasked with predicting the binary $\text{Obese Status}$, yielded a critical counter-intuitive finding: $\text{Dietary Sodium Intake}$ was ranked as the most important feature. This suggests that while $\text{Sodium}$ may not have a simple linear effect on $\text{BMI}$, it plays a complex, high-leverage role in the decision boundaries that define the difference between a high $\text{BMI}$ and a clinically defined $\text{Obese Status}$.

3. Cluster-Based Risk Stratification: $\text{K-Means}$ Clustering effectively segmented the population by $\text{Gender}$ and $\text{Age}$. The Young Male cluster ($\text{Cluster 2}$) was identified as having the highest mean $\text{Waist Circumference}$ and $\text{Sodium Intake}$, indicating a priority subgroup for targeted intervention.

Answering the Research Question

The results successfully allowed us to answer the research question, demonstrating that the association of the predictor variables is differential based on the type of outcome being predicted:

• For the quantitative outcome ($\text{BMXBMI}$), the key predictors are anthropometric and activity ($\text{Waist Circumference}$ and $\text{Physical Activity}$).

• For the categorical outcome ($\text{Obese Status}$), the key predictors involve non-linear contributions from $\text{Dietary Sodium}$ and $\text{Age}$.

Discussion of Implications

The divergence between the $\text{LASSO}$ and $\text{Random Forest}$ results carries significant implications. If a clinician relies solely on the $\text{LASSO}$ result, they would correctly focus on $\text{Waist Circumference}$ and $\text{Physical Activity}$ to predict the magnitude of $\text{BMI}$. However, they would completely miss the high predictive leverage of $\text{Sodium}$ in classifying true $\text{Obese Status}$, which the $\text{Random Forest}$ model revealed. This highlights the crucial necessity of using multiple modeling approaches to fully elucidate the complex drivers of disease risk.

The clustering results provide actionable public health guidance. Interventions aimed at reducing the risk associated with central adiposity and high $\text{Sodium}$ consumption should specifically target younger adult males, as this demographic exhibits the highest mean risk factors across multiple domains.

Limitations and Future Directions

The limitations of this study are typical of cross-sectional analyses. The findings, while establishing associations and predictive importance, cannot infer causality. Furthermore, the low $\text{Recall}$ in the $\text{Random Forest}$ model for the $\text{Obese}$ class indicates that the current suite of predictor variables is insufficient to robustly classify high-risk individuals; additional factors (e.g., genetics, detailed sleep patterns, social determinants of health) are likely missing from the model.

Future research should focus on three areas:

1. Model Improvement: Incorporate richer longitudinal data or a wider array of variables to improve the classification performance (Recall) of the binary $\text{Obese Status}$ model.

2. Mechanistic Exploration: Use the $\text{Random Forest}$ output as a guide to mechanistically study the non-linear relationship between high $\text{Sodium}$ intake and the threshold of $\text{Obese Status}$.

3. Validation: Validate the $K=4$ clusters in an external dataset to confirm that the $\text{Young Male}$ risk profile is a generalized, rather than sample-specific, phenomenon.