Modeling the Relation Between Non-Communicable Diseases and the Health Habits of the Mexican Working Population: A Hybrid Modeling Approach

Abstract

The impact that Non-Communicable Diseases (NCDs) have on the health status of the population has generated the need for an in-depth analysis of health habits and NCDs. In addition to its significant impact on population health, this phenomenon also translates into substantial economic consequences for countries. This study delves into the analysis of the relationship between health habits and NCDs among the economically active population of Mexico. Through a hybrid approach that integrates the use of machine learning (ML) models and a structural equation model (SEM), we seek to quantify the direct and indirect causal effects between health habits and NCDs. For this study, information from the 2022 National Health and Nutrition Survey carried out in Mexico for the working-age population is used. According to the results obtained in the first stage of analysis using ML, the most relevant variables (health habits) that impact the probability of individuals presenting with NCDs were identified (random forest precision of 78.66% and Lasso with 71.27%). The second stage of analysis through SEM using the most relevant variables, which were selected through ML, allowed us to measure the direct and indirect causal effect of health habits on NCDs. The SEM model was statistically significant (Chi-square: 449.186; p-value = 0.0000) and revealed that negative health habits, such as a poor diet, physical inactivity, smoking and alcohol consumption, significantly increase the risk of NCDs in the working-age population in Mexico (0.23), while vigorous physical activity and salary has a negative impact (−0.17 and −0.23, respectively) on the presence of NCDs. This study highlights the ability of machine learning and SEM approaches to model the impact of health habits on NCDs for the economically active population in Mexico.

1. Introduction

Noncommunicable diseases (NCDs) represent a major public health problem worldwide, posing significant challenges not only to health care systems, but also to workforce productivity [1] and its respective impact on economies. Numerous studies analyze the relationship between health habits and risk factors in determining NCDs, as well as their respective economic impact on both health care and productivity loss [2]. Through a systematic literature review, 126 studies on the economic impact of NCDs at various levels in different countries were analyzed. Measures of productivity in the available studies included DALYs, absenteeism, presenteeism, participation (or not) in the labor market, return to work, changes in the hours worked, and medical or sick leave. It should be noted that studies are generally focused on developed countries and that their implementation involves the implementation of ad hoc surveys in most cases. Data from low- and middle-income countries experiencing economic and epidemiological transition are virtually non-existent.

In the case of Mexico, the economic impact of cardiovascular diseases has been documented, using information on direct costs of care, premature deaths, as well as the prevalence and burden of the disease. However, studies are limited in terms of establishing measures of causality between risk factors, NCDs and economic impact [3,4]. It should be noted that the identified studies address the problem from a public perspective, but it is very important to analyze the connection between health status and productivity in corporate environments [5]. The link between prevention and health promotion to avoid or limit widely identified risk factors, as well as the measurement of the effect of these on NCDs, will allow actions to be prioritized in the public and corporate spheres. It has been identified that workplace wellness programs, which include health assessments and interventions, are linked to a reduction in absenteeism and greater employee commitment [6]. This shows the importance of integrating health management strategies that can significantly affect workforce productivity and overall organizational performance. In the case of developing countries, it is important to have robust evidence for decision-making.

The motivation behind this study arises from the need to quantify the impact of negative health habits on NCDs within the economically active population (EAP) from a preventive point of view. Given the increasing prevalence of chronic diseases and their economic repercussions, there is a need for efficient, evidence-based strategies that mitigate these health risks and impact the productivity of labor. The present study proposes a hybrid model by integrating recent methodologies such as machine learning (ML) algorithms with robust parametric models such as structural equation models (SEMs) to quantify which unhealthy habits in the Mexican working-age population have a greater effect on NCDs. The analysis is presented in two stages. First, through the two-variable selection methodologies used in ML (Lasso model and random forest), a set of unhealthy habits with greater relevance to NCDs are determined. Subsequently, using the selected variables, a SEM model is built with latent factors and structural relationships to measure the direct and indirect impact of unhealthy habits on NCDs for the Mexican working-age population. An important challenge in developing countries is the availability of information. In this paper, the information available from the National Health and Nutrition Survey is analyzed, allowing the relationship between health habits and NCD to be analyzed; however, there is not sufficient information to measure productivity. Later studies will integrate information from other sources to incorporate the productivity factor in the analysis. The results generated by the machine learning models allow the development of a conceptual analytical framework within the proposed structural equation model, allowing the generation of robust evidence for decision-making in the health sector.

In this study, we opt for a hybrid approach, integrating traditional machine learning (random forest, logistic regression) with structural equation modeling (SEM). This framework ensures predictive accuracy while maintaining interpretability, making it suitable for workplace and public health interventions. Unlike ML, which operates as black boxes, our approach provides explainable insights into how individual health behaviors contribute to the risk of NCDs. Future research can extend this work by incorporating ensemble deep learning models and causal machine learning techniques to further refine causal inferences and the predictive accuracy.

This work is structured as follows. The next section presents the literature review, followed by details of the sources of information and the methodology to be used. The main results obtained from the hybrid model (ML and SEM) are presented below. Finally, the conclusions are presented.

2. Literature Review

Recent studies have shown the use of ML methods to analyze the effect of different health habits on NCDs in the working-age population, highlighting a reduction in the bias of estimates. It has been identified that workplace health programs focused on preventive care can generate a return on investment by reducing long-term healthcare expenses and improving work productivity [7]. The male population has been shown to have unhealthy health habits and a higher prevalence of chronic diseases [8]. Additional studies have examined the impact of positive health habits such as physical activity and sleep on work productivity. Increased physical activity positively correlates with better performance and reduced absenteeism [9]. Similarly, better sleep quality is related to higher levels of work productivity [10]. Machine learning and metaheuristic algorithms have also been successfully used to improve diagnostic accuracy in the early detection of chronic diseases such as type 2 diabetes, outperforming traditional methods [11]. Integrating data-driven diagnostics and promoting wellness in the workplace can foster a healthier workforce, benefiting both workers and organizations. This advancement underscores the potential of AI-powered analytics to improve health outcomes. The literature review shows that there has been an average annual increase of 23.25% in scientific production related to machine learning applications in the healthcare sector since 2008 [12].

The application of machine learning in healthcare interventions has shown promising results, enabling more effective disease management and improved health outcomes. Big data analysis from diverse sources such as genetic and laboratory data enables researchers to develop better treatment strategies and predict disease risks [13]. Machine learning algorithms have been used to predict the risk of cardiovascular disease based on lifestyle factors and physiological measurements, achieving high predictive accuracy [14]. Machine learning models have also been successfully applied to analyze health data from wearable devices. Significant correlations between physical activity, sleep patterns, and health outcomes have been demonstrated using sensor data. Furthermore, time-lagged analysis has been used to predict daily physical activity based on the previous day’s experiences, demonstrating the prospective effect of daily experiences on physical activity behavior [15]. Recent studies show the application of deep learning to the analysis of electronic medical records, identifying new patterns in chronic disease management and improving the predictive accuracy [16]. Reinforcement learning has been explored to optimize personalized healthcare interventions, showing improvements in patient adherence and health management [17]. These advances illustrate how machine learning can strengthen healthcare interventions and support proactive health management. Healthcare interventions using big data, data mining, machine learning, and specialized algorithms have shown promising preliminary results. Recent advances in machine learning (ML) and deep learning have introduced highly accurate models for NCD prediction. However, they often lack interpretability and require large datasets for generalization and a large amount of processing resources. Their application to workforce health studies remains limited due to the complexity of real-world data collection, particularly in developing countries.

Structural equation modeling (SEM) has been widely used to explore the intricate relationships between health behaviors and the prevalence of NCDs. For example, SEM models have been used to assess the interaction between socioeconomic status, diet quality, and the risk of cardiovascular disease (CVD) among South African children and adolescents. The findings revealed a significant inverse association between socioeconomic status and dyslipidemia, suggesting that a lower socioeconomic status is related to higher lipid abnormalities, which are precursors to CVD. Interestingly, the study did not find a direct significant relationship between diet quality and dyslipidemia, indicating that socioeconomic status may have a more substantial impact on lipid profiles than diet alone [18]. Similarly, SEM models have been used to examine the factors influencing health-promoting behaviors among older adults with NCDs in Thailand during the post-COVID-19 era. The study identified that individuals’ perceived self-efficacy, health literacy, access to COVID-19-preventive materials, and social media directly and positively influenced their health promotion behaviors. These findings underscore the importance of psychosocial factors and the accessibility of resources for managing health behaviors among older populations with chronic diseases [19]. Another study conducted in Iran based on a SEM model showed that metabolic syndrome (MetS) is associated with sleep duration. It is crucial to identify the factors that disrupt sleep regulation. The study aimed to assess the indirect effect of risk factors related to MetS severity through sleep duration [20]. These studies demonstrate the application of SEM to unravel the complex relationships between health behaviors and NCD outcomes, providing valuable information for targeted interventions, especially in disease diagnosis. One of the major challenges associated with implementing structural equation models is the availability of information that enables the adequate estimation of the models, as well as the definition of a causal conceptual framework. This last aspect is often the main criticism of the generated models, because there is a strong bias, from the researcher’s perspective, in defining the structural relationships to be contrasted.

According to the review, it is important to highlight that no recent studies that implement the analysis of health habits and NCDs were identified for the Mexican population, nor were causal measurement studies identified through SEM models for this purpose. Studies evaluating the effectiveness of primary prevention programs for the prevention and control of NCDs in middle-income countries are limited. The available results come mainly from observational studies, and there is a need to synthesize evidence from controlled trials [21]. Therefore, research that obtains a causal measurement between NCDs and health habits, through the application of recent approaches, is highly relevant for Mexico. This research constitutes the first analysis of the relationship between health habits and NCDs in Mexico by integrating a hybrid approach between parametric and non-parametric models.

3. Materials and Methods

To explore the relationship between health habits and NCDs in the Economically Active Population (EAP) of Mexico, a hybrid methodological approach that integrates a combination of variable selection methods based on machine learning (Lasso regression and random forest) and the construction of a SEM model for the measurement of direct and indirect causal effects was used. The following section details the main characteristics of the selected database, as well as the description of the methods to be used as the proposed hybrid approach.

3.1. Hybrid Modeling Approach

This research proposes an analytical framework that integrates variable selection through ML and the estimation of causal effects through a SEM model. The integration of parametric and non-parametric approaches allows the benefits of both paradigms to be integrated into a methodological framework. One of the great challenges when studying economic and social phenomena lies in the limitation of controlled experiments, which limits the measurement of causal effects. To achieve this, it is necessary to have a robust conceptual framework related to the problem to be addressed, as well as a large data sample that allows for consistent estimates. Often what happens is that the accepted conceptual framework is not necessarily reflected in the measured variables, so it is necessary to adjust the conceptual framework to adapt to reality. In our case, we propose an inverse approach; by using ML models, we can exhaustively identify large sets of variables related to the phenomenon to be analyzed and select a set of relevant variables. Later, using the selected variables, it is possible to structure a coherent conceptual framework that allows us to propose a structural model and measure the causal relationships between variables. Figure 1 shows the proposed approach of the hybrid analytical framework using ML and SEM.

3.2. Data Description

The ENSANUT 2022 database was used, with this survey including information related to the access/use of health services, nutrition, and health habits of Mexicans [22]. The survey is representative at the national level. Based on the literature review, a set of variables related to positive and negative health habits, information on blood samples, sociodemographic variables, and the recorded diagnosis of NCDs was determined. The population under study is made up of a working-age population that reported being working at the time of the survey. The variables that make up the database group together different dimensions related to health habits and their relationship with NCDs, as well as sociodemographic factors. The items considered include:

• General variables: these include demographic and socioeconomic factors related to health status.
• Biochemical variables: these provide direct measures of physiological health and are essential for diagnosing and controlling NCDs.
• Lifestyle variables: these capture behaviors and habits that have significant impacts on health.
• NCDs diagnosis: these variables indicate whether a person has been diagnosed with a major NCD and are used as primary outcomes of interest.

To build the modeling database, multiple tables were integrated, ensuring that each observation had complete information in all relevant dimensions. Initially, 709 observations that contained data on the selected indicators were identified. A filtering process was then applied to include only individuals who confirmed being employed, resulting in 651 observations. Records with missing values for the selected indicators were removed to maintain data quality, resulting in a final dataset of 550 records.

Table 1. Selected variables analyzed from ENSANUT 2022.

Category	Variable	Description	Relevance
Generals	Entity	Geographic region	Important for regional analysis and understanding of geographic disparities in health
	Sex	Male or Female	Gender differences in health behaviors and outcomes
	Age	Age in years	Age is a significant predictor of health status and NCDS risk
	School Level	Education level	Education influences health literacy and access to health resources
	Wage	Income level	Socioeconomic status is a determinant of health outcomes
	Weight	Body weight in kg	Essential for calculating BMI and understanding weight-related health risks
	Height	Height in cm	Used in calculating BMI
	Waist	Waist circumference in cm	Indicator of central obesity, which is a risk factor for NCD
Biochemical	TRIG	Triglycerides	High levels are a risk factor for cardiovascular disease
	COL_LDL	LDL cholesterol	High levels are a risk factor for cardiovascular disease
Lifestyle	Systole	Systolic blood pressure	High levels are a risk factor for hypertension and cardiovascular disease
	kg_lose	Weight loss	Reflects changes in body weight, which can impact overall health
	kg_win	Weight win	Reflects changes in body weight, which can impact overall health
	Tired	Frequency of feeling tired	Indicator of overall well-being and possible sleep issues
	Sleep_well	Quality of sleep	Poor sleep quality is linked to various health issues
	Smoke	Smoking status	Smoking is a major risk factor for numerous diseases
	AF_v	Vigorous physical activity	Physical activity is protective against many chronic diseases
	AF_m	Moderate physical activity	Physical activity is protective against many chronic diseases
	Hrs_c	Walking as a physical activity	Physical activity is protective against many chronic diseases
	Hrs_s	Sedentary activity	Physical activity is protective against many chronic diseases
Diagnoses	Diabetes Mellitus	Diabetes diagnosis	Major NCDs with significant health impacts
	Hypertension	Hypertension diagnosis	Major NCDs with significant health impacts
	Event_CAD	Cardiovascular diseases event	Incident of a NCDs with high mortality and morbidity

Source: ENSANUT, 2022 [22].

shows the variables selected for ENSANUT 2022 and the justification for their selection.

The information on weight and height was transformed to standardize the result according to international standards by converting it to the body mass index (BMI) (See Equation (1)):

$$BMI={\displaystyle \frac{weight}{{height}^2}}$$

(1)

Similarly, the variables DM (Diabetes Mellitus), HT (Hypertension), and Event CAD (coronary artery disease) were used as indicators for individuals diagnosed with any NCD, creating a new variable named ENT for the analysis.

3.3. Machine Learning Models

Two highly accepted ML models were used for the variable selection, Lasso regression and random forest. Both algorithms allow variables to be selected based on a performance criterion. It was decided that both methods would be implemented to have a reference regarding the set of variables to be included in the SEM model. The implemented classification models consider the NCD response variable and the rest of the variables as predictive variables. In both models, the use of cross-validation using the k-folds method is considered for the estimation of hyperparameters, as well as for the evaluation of the models. To evaluate the performance of the models, information from the confusion matrix was used to estimate the metrics used to evaluate classification models, such as precision, accuracy, sensitivity, specificity, and the F1 statistic. These statistics provide a detailed assessment of the model’s performance, including its ability to correctly predict the different categories of the response variable.

3.3.1. Lasso Regression

The Lasso model penalizes the sum of the squared residuals by adding the absolute values of the regression coefficients. This penalty is known as the $L1$ norm and has the effect of forcing the coefficients of the less relevant predictors to tend to zero, according to the degree of penalty used through the hyperparameter lambda. The degree of penalty is controlled by the hyperparameter λ. When λ = 0, the result is equivalent to that of a linear model using ordinary least squares. As λ increases, the penalty is greater, and more predictors are excluded. The Lasso model manages to eliminate the less relevant predictors, making it an efficient method for variable selection [23,24,25]. The objective function of the Lasso model is presented in Equation (2).

$$\sum _{i=1}^N\sum _{t=1}^T{\left(y_{it}-x_{it}\beta \right)}^2+\mathsf{\lambda }\sum _{j=1}^p\left|{\beta }_j\right|$$

(2)

where $p$ is the total number of independent variables, $\lambda$ is the regularization parameter that controls the strength of the penalty y, and $\beta$ corresponds to the parameters related to the predictive variables. To solve the Lasso regression problem, several optimization algorithms, such as coordinate descent or minimum angle regression, can be employed. These algorithms iteratively update the coefficients to minimize the objective function [24]. The closed-form solution for Lasso regression does not exist due to the non-differentiability of the L1 norm. However, efficient optimization algorithms can effectively find the solution [25].

3.3.2. Random Forest

Machine learning methods, such as decision trees and random forest, provide powerful tools that can capture complex and non-linear relationships in data, which can be missed by traditional statistical methods [26,27]. Using a tree-based ensemble method such as random forest allows the predictive performance to be improved by reducing overfitting and capturing complex patterns in the data. Thus, it represents a natural variable selection model by evaluating the importance of variables from the n models built. A random forest model was applied using the training data to improve the prediction accuracy by aggregating multiple decision trees. The model constructs an ensemble of trees, each trained on a random subset of the data, and makes predictions by averaging the outputs of all individual trees. The prediction for a new observation $X$ is calculated as follows (Equation (3)):

$$\widehat{Y}={\displaystyle \frac{1}{B}}\sum _{b=1}^B{f}_b\left(X\right)$$

(3)

where $\widehat{Y}$ represents the final predicted value for the given observation $X$, $B$ is the total number of decision trees in the random forest, and $f_b\left(X\right)$ is the prediction of the b-th decision tree in the ensemble.

3.4. Structural Equation Model

Based on the results generated by the ML models, a set of relevant variables related to the presence of NCDs was selected to build a structural equation model. The proposed model includes a parametric modeling process based on the definition of causality hypotheses. The process of building a SEM model includes the specification of a theoretical model that allows the relationships between the observed and unobserved variables to be described. The causal relationships between the variables considered are identified through the design of a causal diagram. According to the nature of the variables, there are robust estimation methods such as the maximum likelihood method, unweighted least squares, generalized least squares, and weighted least squares. Since it is a modular estimation process, a set of evaluation metrics is contemplated for each component of the model, the measurement model, and the structural model. Some of the most used criteria in the evaluation of the structural model are the comparative fit index (CFI), the Tucker–Lewis’s index (TLI), the root mean square error of approximation (RMSEA) and the standardized root mean square residual (SRMR) [28].

The proposed structural equation model contemplates the use of a model with latent variables. For this purpose, six latent variables were defined: $Y_1$ (Profile), $Y_2$ (Biometrics), $Y_3$ (Stress), $Y_4$ (Biochemistry), $Y_5$ (Consumption), and $Y_6$ (Negative health habits). These were based on related sets of observed variables and considered the results of the Lasso regression and random forest models as a variable selection mechanism. $Y_1$ was formed from the observed variables $X_1$ (age) and $X_2$ (education); $Y_2$ included $X_3$ (BMI), $X_4$ (waist), $X_5$ (systolic blood pressure), $X_6$ (Kg. lost) and $X_7$ (Kg. gained); $Y_3$ was constructed with the variables $X_8$ (tiredness), $X_9$ (hours of sleep) and $X_{10}$ (sleep well); $Y_4$ was composed of $X_{11}$ (high triglycerides) and $X_{12}$ (high LDL); and $Y_5$ included $X_{13}$ (alcohol consumption) and $X_{14}$ (tobacco consumption). Vigorous physical activity was considered as an external variable $X_{15}$. Then, the latent variable $Y_6$ (negative health habits) was constructed by integrating the latent variables $Y_1$ to $Y_5$ and an additional exogenous variable $X_{15}$ (physical activity) and $X_{16}$ (salary). The measurement model is defined according to Equation (4).

$$\begin{gathered} Y_1=~X_1+X_2 \\ {Y}_2=~X_3+X_4+X_5+X_6+X_7 \\ {Y}_3=~X_8+X_9+X_{10} \\ {Y}_4=~X_{11}+X_{12} \\ {Y}_5=~X_{13}+X_{14} \\ {Y}_6=~Y_1+Y_2+Y_3+Y_4+Y_5{+X}_{15}{+X}_{16} \end{gathered}$$

(4)

The outcome variable $W_1$ (NCD) was modeled using $Y_6$, allowing us to understand how latent and exogenous variables influence individuals’ health status. The structural model is presented in Equation (5)

$$W_1~Y_6$$

(5)

Figure 2 shows the causal diagram used to clearly understand the relationships between the observed, latent, and exogenous variables.

Derived from the diagram designed in Figure 1, a mediator and collider causal approach will be used. For the initial section, when generating the latent variable Y7 (negative health habits), the mediator approach will be used. In contrast, for the effect of Y7 and wage on NCD, the collider approach will be employed. The collider causal diagram consists of two variables causing a third one, which may introduce bias if controlled. This is because controlling for a collider opens a backdoor between the variables entering it, potentially distorting the causal estimates [29].

The evaluation of the proposed structural model considers the estimation of a set of model goodness-of-fit statistics [28]. TLI (Tucker-Lewis’s index) is an incremental fit measure that compares the fit of the proposed model to a null model in which all variables are uncorrelated. TLI measures how much better the fit of the proposed model is compared to the null model. TLI values closer to 1 indicate better fit, and values greater than 0.90 are considered acceptable [30]. CFI (Comparative fit index) assesses the incremental fit of the proposed model compared to the null model. CFI compares the proposed model to the null model and considers the number of parameters in the model. Like TLI, CFI values closer to 1 indicate better fit, and values greater than 0.90 are considered acceptable [31]. RMSEA (Root Mean Square Error of Approximation) is a fit measure that assesses the approximation error of the model to the population. RMSEA estimates the difference between the observed covariances and the covariances reproduced by the fitted model. RMSEA values less than 0.05 indicate a good fit, while values between 0.05 and 0.08 indicate a reasonably good fit [32]. SRMR (Standardized Root Mean Square Residual) is a fit measure that assesses the discrepancy between the observed covariances and the covariances reproduced by the model. SRMR standardizes this discrepancy by dividing it by the square root of the variance of the observed covariances. SRMR values less than 0.08 indicate a good fit [33].

4. Results

This section presents the main results generated by the hybrid model. The effect of variables related to negative health habits, demographic factors, and their effect on NCDs is highlighted. The importance of having a causal quantification is that, within the framework of the methodological proposal, it allows evidence to be generated for the design of policies, as well as the monitoring and evaluation of the defined policies. The conjunction of machine learning models and structural equation modeling (SEM) enabled, in the first stage, the identification of key predictors, highlighting the complex interaction between health habits, metabolic markers and the prevalence of the disease. The SEM model, for its part, allowed the direct and indirect causal effects on NCDs to be estimated. The experimental environment for this study was set up on a Windows 11 operating system, using an Intel Core i7 processor, 64 GB RAM, and 1 TB HDD storage, with computational tasks supported by an NVIDIA RTX 3080 GPU. Analysis was performed in R 4.4.3, leveraging several statistical and machine learning libraries. The key libraries included randomForest, ranger, and randomForestSRC for machine learning methods, caret for model cross validation, and tree for decision tree analysis. For the explainability and interpretability of the ML models, randomForestExplainer and DALEX were used. Structural equation model (SEM) implementation was performed using the lavaan package.

4.1. Descriptive Statistics

The analyzed sample shows a mean age of 47.9 years (95% CI: 46.89–48.84), with a mean body mass index (BMI) of 30.5 kg/m² (95% CI: 30.00–30.91), indicating a predominantly overweight or obese population with low physical activity (

Table 2. Descriptive statistics of the selected variables.

Variable	Media	SD	Min	Max	Mediana	CI 95%
Age	47.9	11.7	22.0	65.0	50.0	46.89–48.84
BMI	30.5	5.4	16.7	49.6	29.8	30.00–30.91
Waist	100.3	13.1	61.2	137.5	99.3	99.16–101.34
Sistole	123.6	18.9	81.3	200.0	122.0	122.01–125.16
Kg (gained)	1.5	3.0	-	20.0	-	1.22–1.72
Kg (lost)	2.6	4.4	-	21.0	-	2.22–2.96
Alcohol consumption (drinks per year)	13.3	9.6	5.0	34.0	6.0	12.49–14.09
Physical activity (vigorous)	1.0	2.1	-	7.0	-	0.86–1.22
Physical activity (moderate)	3.6	3.0	-	7.0	3.0	3.32–3.83
hrs_c	1.6	2.0	-	12.0	0.7	1.40–1.74
hrs_s	3.0	2.5	-	12.0	2.0	2.81–3.24
Sample (n): 550

Source: ENSANUT, 2022.

).

Regarding the distribution of categorical variables, 48.7% were women, while 51.3% were men. Regarding the presence of NCDs, 51.8% of the participants reported at least one condition (for example, diabetes, hypertension or cardiovascular events), while 48.2% did not report these conditions. As for sociodemographic factors, a large range of salaries are observed; the educational levels of the population also range widely (

Table 3. Proportional distribution of the selected categorical variables.

Variable	Category	%
NCD	No	50.36
	Yes	49.64
Sex	Men	52.36
	Female	47.63
LDL (high)	No	75.27
	Yes	24.73
Perception of tiredness	Never	53.27
	A little	24.00
	A lot	13.82
	Always	8.91
Slept well	Never	39.82
	A little	24.36
	A lot	19.27
	Always	16.55
Smokes	Does not smoke	83.27
	Every day	10.00
	Some days	6.73
Education	Middle	32.00
	Elementary	24.00
	High school	20.36
	Bachelor	12.55
	None	5.27
	Technical	4.91
	Master	0.91
Salary (Pesos range in thousands)	6–9.9	45.09
	10–13.9	32.00
	14–21.9	14.36
	>22	8.55
Hours of sleep	8 h	33.09
	7 h	23.27
	6 h	18.55
	<9 h	12.73
	<5 h	12.36
Triglycerides (high)	No	73.09
	Yes	26.91

Source: ENSANUT, 2022.

).

Regarding the response variable, the data do not show imbalance problems, so no subsampling or oversampling method was applied to deal with this condition. The proportion of missing data was minimal, so it was decided that these incomplete records would be eliminated, ensuring consistency between the selected indicators. No normalization or scaling operations were applied since all variables were categorical or measured on comparable scales, and since they did not impact the ML methods used. Additionally, this allowed for a direct interpretation of the results of the SEM model.

4.2. Feature Selection by Machine Learning

The estimation of the ML models considered the optimization of hyperparameters using a cross-validation criterion, which allowed the performance of the implemented models to be improved. The results of the variables selected by the Lasso and random forest regression are compared. There is congruence between both methods regarding the relevant variables. Both sets of variables allow a comparison with the conceptual framework to be performed to establish the SEM causal models evaluated. The confusion matrix, accuracy, precision, sensitivity, VPP and VPN statistics were used as performance metrics.

The random forest model achieved an accuracy of 78.66% on the test set. The model achieved a sensitivity of 62.67%, a specificity of 93.29%, a positive predictive value (PPV) of 89.52%, and a negative predictive value (NPV) of 73.21%. These results indicate that while the model effectively identifies true negatives (high specificity) and maintains a strong PPV, it has a moderate sensitivity, meaning that there is room for improvement in detecting true positive NCD cases. However, for variable selection purposes, the model shows consistent results under the conceptual framework. For its part, the Lasso penalized regression uses the penalty with the L1 norm so that the coefficients of those variables that are not very relevant in the prediction of the model have parameters equal to zero. In this case, the performance of the Lasso model is very close to that generated by the random forest, yielding the following performance metrics: an accuracy of 71.27%, a sensitivity of 64.23%, a specificity of 78.26%, a PPV of 74.58%, and a NPV of 68.79%.

The random forest model employed the deviance split criterion to determine the optimal split points in the decision trees. This criterion selects the feature and threshold that minimize the residual deviance, effectively reducing the impurity at each node and ensuring that splits provide maximum information gain. The minimum values for the node splitting and terminal node size were optimized to balance the complexity and generalization of the tree, avoiding excessive fragmentation and maintaining model interpretability. The importance of the variables was assessed through the permutation importance method by highlighting the predictors that had the most significant impact on classification. Permutation importance provides a more robust and unbiased assessment of variables’ significance by measuring the decrease in model performance when each predictor is randomly selected. This approach ensures that the ranking of variables is not influenced by correlations or dependencies between features, leading to more confident predictor selection. The results indicate that age is the most influential factor. This finding is consistent with epidemiological studies showing that aging is a major driver of chronic disease prevalence. Other relevant variables include individuals’ systolic blood pressure, waist size, BMI, and educational level (See Figure 3a).

The variables selected by Lasso are consistent with those reported by random forest. The main variables that contribute to the presence of NCDs are individuals’ perception of sleeping well, hrs_s, high triglyceride levels, age, and a systolic blood pressure. The variables that decrease the presence of NCDs are moderate and vigorous physical activity, and the state (Mexico City) of residence (see Figure 3b).

The analysis demonstrates the complementary nature of the variable selection methods. While Lasso regression reports easily interpretable coefficients, it can overlook non-linear relationships between predictor variables and the response variable. In contrast, random forest allows the capture of complex patterns and interactions between variables. Random forest’s ability to manage large numbers of predictors and detect non-linear relationships make it particularly effective for complex datasets. The combination of both variable selection approaches allows the identification of relevant variables in determining the response variable. Combined with the conceptual framework of the relationship between risk factors or negative health habits and the presence of NCDs, this will allow the definition of a coherent SEM based on the empirical evidence identified in the second stage of modeling. The following section presents the results of the identified and validated SEM, which enables the causal effects between variables on NCDs to be measured directly and indirectly.

4.3. Structural Equation Modeling

In the next stage of analyzing the hybrid model, a structural equation model (SEM) was developed, as presented in Section 3.3. Given the characteristics of the model and the variables used, estimates were made using the weighted least squares (WLS) method. Different specifications and structures were evaluated, and the measurement and structural model in which all the variables considered were statistically significant was selected; this is in addition to those that presented adequate goodness-of-fit indicators and coherence in the expected coefficients according to the conceptual model (See

Table 4. SEM goodness of fit.

Index	Value
Chi-Square	449.186
Tucker–Lewis Index	111.000
Comparative Fit Index	0.000
Root Mean Squared Error Approximation	0.074
Standardized Root Mean Square Residual	0.077

Source: Estimated by the authors based on ENSANUT 2022.

).

A good fit is observed with a chi-square value of 449.186 (p-value = 0.000), indicating that the model fits the data well. Global fit measures such as TLI and CFI exceed the threshold of 0.9, while RMSEA and SRMR are below 0.06, suggesting a good model fit. The measurement model (confirmatory factor analysis) provides a detailed view of how latent and observed variables relate to each other. It is observed that latent variables are significantly influenced by their observed variables and that the signs correspond with what was expected. For example, systolic blood pressure (X5) has a positive effect on the latent variable biomarkers (Y2), while the variable Kg. lost (X6) has a negative impact on this latent variable. Structural models show the causal effect of the exogenous variables salary (X16) and vigorous physical activity (X15), which negatively impact negative health habits. And as expected, the latent variable negative health habits (Y6) has a positive effect on NCDs (W1) (see Figure 4).

Similarly, the latent variable stress is influenced by feelings of tiredness and sleep habits. Chemical biomarkers are influenced by high levels of triglycerides and LDL. Alcohol consumption has a significant effect on the latent variable consumption, while smoking appears to have a minor effect. Vigorous physical activity and salary are directly related to negative health habits and negatively affects them. The identified covariance relationships are consistent with the expected results. A complete overview of the estimated parameters, as well as their respective confidence intervals, is shown in Appendix A.

The structural model was built using the variable NCDS (W1) as the dependent variable, while negative health habits was included as the independent variable (Y6). Negative health habits were treated as a latent variable, encompassing multiple observed indicators that reflect health-damaging behaviors, physical activity, smoking, and alcohol consumption. The results of the model indicate that negative health habits have a strong and significant positive effect on noncommunicable diseases. This finding shows that improving health habits could substantially reduce the risk of chronic diseases. In contrast, the salary variable exhibits a negative effect on negative health habits, suggesting that higher income levels may contribute to a reduction in negative health habits, in the same way as performing vigorous physical activity.

5. Conclusions

The results obtained in this research show the potential of integrating hybrid approaches into quantitative modeling applied to the analysis of NCDs in Mexico. The results of the ML-based feature selection models showed an accuracy of 71.27% for the Lasso regression and 78.66% for the random forest, which allowed us to guarantee a robust selection of relevant variables that, in conjunction with a conceptual framework, allow for a deeper analysis of the causal effects between the directly and indirectly observed variables. These models collectively highlight the importance of sociodemographic variables and negative health habits on NCDs.

Once a set of measurable characteristics consistent with a conceptual framework was selected, the second stage of the proposed hybrid model was implemented; this involved the construction of a structural equation model to measure the causal effect between variables directly and indirectly. The structural equation model with latent variables provided deep insights into the relationships between the latent and observed variables in our study. In the full model, six latent variables and two exogenous variables were evaluated to measure their effect on NCDs. The SEM goodness-of-fit statistics showed a satisfactory fit to the observed data (Chi-square: 449.186; p-value = 0.0000). The analysis of the structural model shows that negative health habits (Y6) have a significant effect on NCDs. Likewise, it was identified that salary and vigorous physical activity have an important effect on the reduction in negative health habits.

The analysis highlights the significant challenge that NCDs represent in the economically active population in Mexico, emphasizing the effect of negative health habits on NCDs, supporting what other researchers have tried to suggest [34,35]. To address this challenge, it is necessary to implement effective prevention and health promotion strategies that seek to reverse unhealthy lifestyles in workers. It is essential to move from a reactive to a proactive approach in health management. Instead of focusing solely on treating diseases as they arise, emphasis should be placed on preventing them and promoting healthy lifestyles. This proactive stance can help reduce the costs associated with chronic diseases and improve the quality of life of workers. It is imperative to implement comprehensive occupational health policies that not only focus on disease prevention but also encourage healthy lifestyles and conducive work environments. The present study provides evidence regarding the factors related to health behaviors that have the greatest impact on NCDs by integrating two quantitative modeling methods in a hybrid approach.

The main contributions of this work are that it provides a novel approach to the integration of ML and SEM methodologies to analyze the complex relationship between health habits and the risk of NCDs in the economically active population in Mexico. The study identifies key predictors of NCDs, including waist circumference, systolic blood pressure, sleep quality, educational level, and salary. The study proposes a structured framework for analyzing health data in real time (to the extent that timely information is available), leveraging predictive analytics to improve decision-making in health policies. Finally, it highlights the importance of integrating health prevention strategies in workers, contributing to a broader understanding of which unhealthy lifestyle factors should be prioritized in health prevention and promotion policies.

This study is limited by its cross-sectional design, as well as by the availability of information over wide time intervals. The cross-sectional design of this study limits causal inferences, and the self-reported data may introduce bias. The specification of the SEM depends on the choice of the best predictors; it is therefore a reactive approach that allows the greatest amount of information to be extracted from the available data. Relevant economic factors regarding the income and expenses of individuals that in turn affect their healthy lifestyle are not available in ENSANUT, which limits the scope of the conclusions.