A machine learning model for predicting obesity risk in patients with diabetes mellitus: analysis of NHANES 2007–2018

This study utilized cross-sectional data from the NHANES survey, collected in the United States between 2007 and 2018. The survey, conducted by the National Center for Health Statistics (NCHS) under the Centers for Disease Control and Prevention (CDC), aimed to provide a nationally representative assessment of the health and nutritional status of non-institutionalized civilians in the United States. Since NHANES is a publicly available database and has been approved by the Institutional Review Board (IRB) of NCHS, our institution confirmed that no additional ethical approval was required. Furthermore, the IRB acknowledges that NCHS adheres to strict ethical standards in its data collection and processing procedures, including obtaining informed consent from all participants and ensuring data anonymization. These measures guarantee full compliance with ethical guidelines for secondary data analysis.

Individuals diagnosed with diabetes were included in this study if they self-reported a diabetes diagnosis, had a fasting blood glucose level ≥126 mg/dL, had a hemoglobin A1c level ≥6.5%, or reported using anti-diabetic medications. Obesity was assessed based on participants' body mass index (BMI), with a BMI ≥ 30.0 kg/m² defining obesity, and those with a lower BMI classified as non-obese. To maintain data quality and consistency, records with more than 10% missing data were excluded, while multiple imputation was applied to records with minor missing values. Following these stringent inclusion and exclusion criteria, a total of 3,794 diabetes participants were included in the final analysis, consisting of 2,163 participants in the obesity group and 1,631 participants in the non-obesity group (Figure 1).

This study carefully selected a set of variables to investigate the factors associated with obesity in DM patients. These variables include demographic, socioeconomic, lifestyle, clinical, and biochemical factors. The demographic variables consist of gender, age, race, education level, and marital status. Education level is categorized as below high school, high school, and above high school. Marital status is classified into the following categories: married, widowed, divorced, separated, never married, or living with a partner. Race is divided into categories such as Mexican American, Other Hispanic, Non-Hispanic White, Non-Hispanic Black, and Other Race—including Multi-Racial. The Poverty-to-Income Ratio (PIR) is calculated by dividing household income by the corresponding poverty line for the survey year and location (12).

Lifestyle variables include smoking status (smoker or non-smoker) and alcohol consumption behavior (drinker or non-drinker). Physical activity levels are categorized as light, moderate, and vigorous exercise based on MET values.

Anthropometric measurements collected during the NHANES survey cycle include BMI. Biochemical markers include Alb, ALP, ALT, AST, TC, Scr, TG, UA, Hb, and HbA1c. Data on essential hypertension (EH), chronic heart disease (CHD), and depression were obtained through the questionnaire modules. Essential hypertension (EH) and coronary heart disease (CHD) were defined based on self-reported information from the "Blood Pressure and Cholesterol Questionnaire (BPQ)" and the Medical Conditions Questionnaire (MCQ) modules. Depressive symptoms were assessed using the PHQ-9 score, where a score of ≥10 indicates depression, and a score of <10 indicates no depression (13). Intake of total energy, protein, carbohydrate, total sugar, dietary fiber, total fat, vitamin B12, vitamin C, vitamin D, and vitamin K was obtained from the Dietary Data module.

NHANES employs a complex multi-stage survey design to generate nationally representative data. Descriptive analysis involves calculating weighted averages with 95% confidence intervals and weighted percentages to account for the complex sampling design of the survey. Following NHANES analysis guidelines, chi-square tests were used for categorical variables, and weighted linear regression was applied to continuous variables to compare weighted groups. Unweighted data were used for model development and statistical analysis. Continuous variables are reported as means with standard deviations (SD), and categorical variables are presented as frequencies and percentages. Chi-square tests were used for categorical variables, and independent t-tests were applied to continuous variables to evaluate intergroup differences in clinical characteristics. A p-value of less than 0.05 was considered statistically significant, and all tests were two-tailed.

All statistical analyses in this study were performed using R software (version 4.2.2). Initially, LASSO regression with 10-fold cross-validation was conducted using the cv.glmnet() function from the glmnet package for variable selection. To enhance the robustness and comparability of model performance, we subsequently constructed and evaluated nine machine learning models: logistic regression (glm() with family = binomial), random forest (randomForest), radial support vector machine RSVM (e1071), k-nearest neighbors KNN (class), extreme gradient boosting XGBoost (xgboost), light gradient boosting machine LightGBM (lightgbm), decision tree DT (rpart), elastic net ENet (glmnet, α = 0.5), and multilayer perceptron MLP (nnet). Model performance was comprehensively assessed across four dimensions: (1) discrimination, measured by the area under the curve (AUC) using the pROC package; (2) calibration, evaluated via calibration plots using the rms and caret packages; (3) overall predictive accuracy, quantified using Brier scores calculated by the DescTools and ModelMetrics packages; and (4) clinical utility, determined by decision curve analysis (DCA) using the rmda package to estimate net clinical benefit at various threshold probabilities. All visualizations were generated using the ggplot2 package to ensure a clear and comprehensive presentation of the results. These methodological details have been incorporated into the revised Methods section to enhance clarity and reproducibility of the study.

LASSO regression and cross-validation were used for feature selection in the training set to identify significant predictive factors. Features with non-zero coefficients were retained for subsequent analysis. We randomly allocated 70% of the patient population to the training set and 30% to the testing set. Stratified sampling was applied to ensure a balanced distribution of the target variable between the two groups. To address the issue of class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was used to generate a balanced dataset in the training set (14).

In this study, we employed nine machine learning algorithms, including logistic regression, to model and predict the risk of obesity among patients with diabetes. These algorithms comprised logistic regression, random forest (RF), radial support vector machine (RSVM), k-nearest neighbors (KNN), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), decision tree (DT), elastic net regression (ENet), and multilayer perceptron (MLP). Model construction was performed on the training set, and only after repeated training yielded stable results were the models evaluated on the independent testing set. Model performance was assessed across four dimensions: discrimination (measured by ROC curves and AUC), calibration (evaluated using calibration plots), overall predictive accuracy (assessed using the Brier score), and clinical utility (measured by DCA to estimate net clinical benefit across a range of threshold probabilities). The model with the best overall performance was ultimately selected for clinical visualization and interpretation.

The model's performance was evaluated based on classification accuracy and robustness. After training, the model was validated using the test set. The model's classification performance was assessed using the ROC curve and the AUC metric (15). ROC curve analysis in the training set provided insights into the model's predictive capability, while the AUC value was used to evaluate its discriminative ability.

Furthermore, a nomogram was developed based on the logistic regression model, providing an intuitive and user-friendly tool for clinical use (16). The nomogram visually illustrates the relative contribution of each predictor to the obesity risk score. Each predictor is assigned a score proportional to its regression coefficient, and the total score corresponds to the predicted probability of obesity. A calibration curve was created to assess the predictive performance of the nomogram by comparing the predicted and observed probabilities of obesity. The discriminatory ability of the nomogram was evaluated using the AUC of the ROC curve. DCA was performed to evaluate the clinical utility of the nomogram by quantifying the net benefit at different threshold probabilities.

A total of 3,794 participants were included in this study, with an obesity prevalence rate of 57.01% (2,163/3,794). Among the participants, 1,990 were male and 1,804 were female, with average ages of 60.49 ± 12.66 years and 60.22 ± 12.96 years, respectively.

Compared to non-obese diabetic patients, those in the obese group were younger (57.85 ± 12.58 years vs. 63.69 ± 12.32 years), and the proportion of females was higher in the obese group (1,179/2,163 vs. 625/1,630). Additionally, significant differences were observed in education level, marital status, and race between the two groups. Participants in the obese group slept significantly less than those in the non-obese group (6.98 ± 1.67 vs. 7.26 ± 1.66, p < 0.05). Contrary to expectations, there were no significant differences in PIR, smoking, or drinking behaviors between the two groups (p > 0.05). Regarding past diseases, the obesity group had a higher prevalence of essential hypertension and depression (p < 0.05), while there were no significant differences in the prevalence of CHD and stroke (p > 0.05).

Regarding blood markers, significant differences were observed in ALB, ALP, ALT, Scr, TG, UA, and HbA1c between the two groups (p < 0.05). No differences were found in the levels of AST, TC, and Hb between the two groups (p > 0.05). In terms of energy intake, participants in the obesity group consumed more energy (1,964.70 ± 956.79 vs. 1,863.92 ± 854.06, p < 0.05). Significant differences were observed in the intake of carbohydrates, total sugar, dietary fiber, and total fat between the two groups (p < 0.05), while no differences were found in the intake of vitamin B12, C, and D (p > 0.05). Table 1 presents all the significant variables.

In the variables of DM participants, the occurrence of obesity is associated with several factors (Figure 2A). Age appears to be the most significant variable, followed by gender and UA, suggesting that age, gender, and UA may play crucial roles in the obesity risk of diabetic patients. Furthermore, variables such as ALP, essential hypertension, TG, and depression showed significant predictive value. Other influencing factors include total fat, HbA1c, and education level, among others. In comparison, variables like smoking, alcohol consumption, vitamin intake, and sleep duration demonstrated relatively low predictive performance, indicating a weaker correlation with obesity risk (Figure 2B).

To enhance the model's robustness and minimize potential overfitting, LASSO regression was applied for the initial selection of predictor variables in the training set, with the optimal regularization parameter determined using 10-fold cross-validation (17). The LASSO model initially identified 19 variables with non-zero coefficients, including sex, age, race, marital status, PIR, ALB, ALP, AST, Scr, TG, UA, Hb, and sleep duration (Figure 3 and Supplementary Table S1).

Based on the selected variables, nine machine learning models were developed, including logistic regression, RF, RSVM, KNN, XGBoost, LightGBM, DT, ENet, and MLP. All models were stably trained on the training set and subsequently evaluated on the testing set. Among these, the logistic regression model demonstrated the best overall performance across multiple evaluation dimensions. It exhibited strong discriminatory power, with an AUC of 0.751 in the training set and 0.781 in the testing set, indicating reliable differentiation between obese and non-obese patients (Figures 4A,B). In terms of clinical applicability, DCA showed that the logistic model provided the greatest net benefit across a range of threshold probabilities (Figures 4C,D). The model also achieved favorable calibration, with predicted probabilities closely aligning with observed outcomes, and yielded the lowest Brier score (0.189), reflecting high overall predictive accuracy (Figures 4E,F). These findings support the selection of logistic regression as the optimal model for obesity risk prediction in patients with diabetes. Accordingly, logistic regression was selected as the optimal predictive model in this study.

Following the identification of logistic regression as the optimal predictive model among the nine machine learning algorithms, we further visualized the model coefficients and constructed a clinically applicable prediction tool. As shown in Figure 5, among the variables included in the logistic model, depression and EH emerged as the strongest positive predictors, suggesting that patients with these conditions are at increased risk of obesity. Additionally, metabolic indicators such as ALP and total fat intake also contributed positively. In contrast, variables such as male sex, Alb, and age were negatively associated with obesity risk, indicating a potential protective effect.

Based on the estimated regression coefficients, a nomogram was constructed to facilitate individualized risk assessment (Figure 6). This nomogram incorporates all significant predictors from the model. By locating the value of each patient's characteristic on the corresponding axis, assigning points, and summing the total score, clinicians can estimate the predicted probability of obesity. The tool is intuitive, easy to interpret, and clinically applicable, offering a practical method to support early identification and personalized management of obesity risk in patients with diabetes.

To further enhance the clinical interpretability of the nomogram, we determined the optimal cutoff point based on the maximum Youden index. As shown in Figure 7, the Youden index peaked when the total nomogram points reached 131, indicating the most balanced trade-off between sensitivity and specificity at this threshold. This cutoff value can serve as a reference point in clinical decision-making, enabling physicians to identify high-risk patients who may benefit from early lifestyle intervention or closer metabolic monitoring (Table 2).

This study has several limitations that should be acknowledged. First, the data were derived from the NHANES database between 2007 and 2018, and therefore may not fully capture more recent trends or behavioral and physiological changes in the post-COVID-19 era. Second, due to the cross-sectional nature of the NHANES data, causal inferences cannot be drawn from the observed associations. Third, although multiple machine learning algorithms were used and their performance compared, external validation using independent datasets was not conducted, which may limit the generalizability of the models. Fourth, some relevant variables such as genetic, environmental, or medication-related factors were not available in the dataset, potentially affecting the model's completeness. Finally, although the study applied nomogram visualization, the clinical applicability of the prediction model should be further tested in prospective studies and real-world settings.