Analytical data originated from the 2007–2012 Korea National Health and Nutrition Examination Survey (KNHANES) database administered by the Korea Centers for Disease Control and Prevention (CDC) [19, 20]. KNHANES constitutes a nationally representative continuous cross-sectional survey employing multistage stratified cluster sampling methodology. The survey comprises three distinct modules: health examinations, nutrition assessments, and health interviews. The institutional review board of the KCDC granted ethical approval for the study protocol, with all participants providing written informed consent.

The initial analytical cohort included 50,405 KNHANES participants (2007–2012). Screening procedures first excluded individuals with incomplete physical activity documentation, yielding 8,263 eligible subjects. Subsequent exclusions comprised participants meeting adequate physical activity thresholds (defined in Section Definition of physically inactive population) and those aged <19 years, leaving 3,750 individuals. Following exclusion of 393 subjects with missing depression-related metrics, the final analytical sample encompassed 3,357 adults. To assess temporal robustness within the same survey framework, the 2011 (n = 477) and 2012 (n = 502) KNHANES subsets were reserved as temporally separated internal validation cohorts. The remaining participants from 2007 to 2010 (n = 2,378) were randomly partitioned into training (n = 1,665) and testing (n = 713) subsets at a 7:3 ratio. Figure 1 presents the comprehensive participant selection workflow.

Consistent with World Health Organization (WHO) physical activity guidelines, this study evaluated individual activity levels using metabolic equivalent minutes per week (MET-min/week). Exclusion criteria incorporated missing essential data elements including weekly frequency and average daily duration of walking, moderate-intensity exercise, and vigorous-intensity exercise. Referencing WHO recommendations, minimum effective session durations were established: ≥15 min for vigorous exercise and ≥10 min for moderate exercise or walking. To minimize extreme value influence on cumulative activity calculations, session durations underwent Winsorization at the 99th percentile. Following KNHANES methodology, total weekly physical activity (PA) was computed as PA (MET-min/week) = MET coefficient × session duration × weekly frequency [21]. Based on WHO standards, participants achieving PA < 600 MET-min/week were classified as “physically inactive,” whereas those attaining PA ≥ 600 MET-min/week were designated “physically active” [21–23].

This investigation employed Asian-specific diagnostic criteria for obesity classification and the Patient Health Questionnaire-9 (PHQ-9) for depression assessment to define obesity-depression comorbidity (ODC). Obesity categorization integrated body mass index (BMI) and waist circumference (WC) measurements: generalized obesity was defined as BMI ≥25 kg/m²; abdominal obesity was defined as WC ≥ 90 cm for males or ≥85 cm for females. Participants were consequently stratified into four mutually exclusive categories: 1) Non-obese (below threshold values for both BMI and WC); 2) Isolated abdominal obesity (WC exceeding threshold with subthreshold BMI); 3) Isolated generalized obesity (BMI exceeding threshold with subthreshold WC); and 4) Compound obesity (exceeding thresholds for both indices). For analytical purposes, categories 2–4 were collectively classified as obese. Depression status was assessed using the Patient Health Questionnaire-9 (PHQ-9; total score range: 0–27) [24]. Consistent with prior validation studies, a PHQ-9 score of ≥10 was used to indicate clinically significant depressive symptoms [25]. Thus, in this study, depression was operationally defined on the basis of symptom screening rather than physician-diagnosed depression.

Based on existing literature and clinical expertise, this study incorporated multiple classes of potential predictor variables relevant to depression-obesity comorbidity, comprising: Demographic and sociological characteristics (sex, age, household income, educational attainment, marital status); health status indicators and disease history (hypertension, dyslipidemia, stroke, myocardial infarction, arthritis, diabetes, smoking status, alcohol consumption); clinical signs and laboratory parameters (systolic blood pressure, diastolic blood pressure, height, fasting glucose, insulin, total cholesterol, high-density lipoprotein cholesterol (HDL-C), triglycerides, hematocrit, ferritin, serum creatinine, vitamin D, white blood cell count, red blood cell count, platelet count); dietary intake metrics derived from 24-h recall (total food mass, total energy intake, water consumption, protein, fat, carbohydrates, calcium, phosphorus, iron, sodium, potassium, vitamin A, β-carotene, retinol, thiamine, riboflavin, niacin, vitamin C).

The initial dataset contained 46 predictor variables (12 categorical, 34 continuous). To develop robust, generalizable prediction models, systematic data preprocessing and modeling procedures were implemented. Samples with missing values were excluded in a complete-case analysis, rather than being imputed, to avoid introducing additional model-based uncertainty across heterogeneous demographic, clinical, and nutritional variables during preprocessing. Subsequently, three feature selection methods—logistic regression, LASSO regression, and random forest—were applied to identify optimal predictive feature subsets. The LASSO method was selected for final feature subset construction based on five-fold cross-validated area under the receiver operating characteristic curve (AUC-ROC) performance in the training set. SHAP values were used only after final model development to interpret variable contributions within the XGBoost model; therefore, SHAP-based importance rankings were not intended to replicate the feature selection results obtained in the preprocessing stage. Pearson correlation coefficients were computed to address multicollinearity, retaining variables with stronger outcome associations when pairwise correlations exceeded 0.8. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) was applied exclusively to the training set to improve recognition of the minority ODC-positive class [26]. No oversampling was performed in the internal test or external validation datasets, thereby reducing the risk of information leakage and overly optimistic performance estimates.

Multiple candidate models were trained and compared using SMOTE-processed data, including logistic regression, random forest, XGBoost, decision tree, naïve Bayes, K-nearest neighbors, and radial basis function (RBF) kernel support vector machines (SVM). Hyperparameters were systematically optimized via grid search with five-fold cross-validation on the internal test set. Following comprehensive performance comparisons, the optimally parameterized XGBoost model was selected for final evaluation on temporally independent external validation sets (2011 and 2012 data) to assess clinical generalizability and stability. Crucially, SMOTE application was restricted to training data construction, while test and external validation sets retained original class distributions to prevent information leakage and ensure objective evaluation.

To objectively evaluate model generalizability, the 2005–2020 U.S. National Health and Nutrition Examination Survey (NHANES) cohort served as an independent external validation dataset. Applying identical inclusion/exclusion criteria as KNHANES produced a validation cohort of 2,070 participants (demographic characteristics in Supplementary Table S1). NHANES was selected as an accessible and well-characterized independent population-based dataset for external testing; however, it was not intended to represent the most culturally or clinically comparable population to South Korea. We did not perform U.S.-specific recalibration. Instead, the NHANES analysis was intended to assess the external discrimination and transportability of the KNHANES-derived model in an independent population setting. In the external validation cohort, obesity was defined according to U.S.-appropriate criteria to preserve the clinical relevance of outcome ascertainment in that population. Within this cohort, XGBoost performance was benchmarked against conventional algorithms including logistic regression, SVM, and random forest, primarily using area under the receiver operating characteristic curve (AUC) for discriminative ability assessment. To interpret the optimal model’s (XGBoost) prediction patterns and identify influential predictors, mean absolute SHAP values were computed across the validation cohort, ensuring interpretative consistency and local explanation accuracy.

Binary prediction performance for ODC was comprehensively evaluated through systematic comparison of multiple machine learning algorithms: XGBoost, decision tree, logistic regression, naïve Bayes, K-nearest neighbors, random forest, and RBF-kernel SVM. Performance was assessed multidimensionally: Fundamental metrics included error rate and accuracy; class imbalance was addressed via Fβ-score (integrating precision and recall); discriminatory capacity was measured by AUC, sensitivity, and specificity; precision-recall balance was evaluated via precision-recall AUC (PR AUC). Calibration curves assessed agreement between predicted probabilities and observed event rates. Decision curve analysis (DCA) quantified clinical utility by comparing net benefits across decision thresholds. Following comparative evaluation using these metrics, the best-performing XGBoost model underwent final validation and interpretability analysis. SHAP (SHapley Additive exPlanations) values enabled quantification of individualized feature contributions to ODC predictions, enhancing model interpretability. Finally, an interactive online risk prediction tool was developed using the R Shiny framework based on the validated XGBoost architecture. All analyses were conducted in R (version 4.4.1) employing critical packages: DMwR, ggcor, mlr3, mlr3benchmark, mlr3extralearners, kernelshap, and shapviz. Two-sided statistical tests were applied with significance defined as p < 0.05.

The final analytical cohort comprised 3,357 physically inactive adult participants, stratified into training (n = 1,665), testing (n = 713), and two temporally independent internal validation cohorts (2011: n = 477; 2012: n = 502). Baseline characteristic analysis revealed no statistically significant differences in core demographic and clinical variables—including gender distribution, household income, educational attainment, marital status, and chronic disease history—across datasets (p > 0.05), with all standardized mean differences (SMD) below the 0.1 threshold. Though statistically significant variations existed for select metabolic and nutritional indicators (p < 0.05), their SMD values remained below the 0.3 benchmark for clinical relevance. Compared to the training set, validation cohorts exhibited elevated high-density lipoprotein cholesterol (HDL-C) and hematocrit concentrations, while demonstrating significantly reduced vitamin D levels and diminished dietary intakes of carbohydrates, phosphorus, sodium, and potassium (all p < 0.01; SMDs < 0.227). Statistically significant but clinically marginal differences were also observed for diastolic blood pressure, height, insulin concentrations, and alcohol consumption patterns (p < 0.05; SMDs<0.106). With all variables exhibiting SMDs below 0.3, the datasets demonstrated satisfactory clinical comparability for subsequent machine learning modeling and validation procedures (complete data in Supplementary Table S2).

To elucidate geographical heterogeneity in obesity-depression comorbidity (ODC) burden, we generated a spatial heatmap depicting ODC prevalence probabilities across Korean administrative regions (Figure 2). Results revealed marked epidemiological disparities: South Gyeongsang Province exhibited the highest prevalence (10.7%), followed sequentially by South Chungcheong, North Jeolla, Jeju Island, and North Gyeongsang (all >6.0%). Remaining regions demonstrated relatively uniform distribution patterns. This geographical stratification establishes critical context for subsequent subgroup analyses and informs assessments of model generalizability across diverse populations.

To construct an optimal predictive feature subset, we systematically evaluated three feature selection methodologies—logistic regression, LASSO regression, and random forest—on the training cohort (Figure 3A). Comparative assessment using five-fold cross-validated area under the receiver operating characteristic curve (AUC-ROC) demonstrated the random forest approach achieving superior discriminatory capacity (AUC = 0.721), significantly outperforming LASSO regression (AUC = 0.698) and logistic regression (AUC = 0.674) (Figure 3B). Consequently, random forest was selected as the definitive feature selection technique. Variable importance ranking yielded the top 30 predictors (Figure 3D), encompassing metabolic biomarkers, hematological indices, nutritional parameters, and demographic characteristics. To mitigate multicollinearity effects, Pearson correlation matrices were computed (Supplementary Figure S1). For feature pairs exhibiting correlation coefficients >0.8, we retained variables demonstrating stronger associations with the outcome (ODC), excluding six redundant parameters: serum creatinine, vitamin A, water intake, carbohydrate intake, total energy intake, and hematocrit. This refinement process yielded a final feature set comprising 24 core variables for model construction (Figure 3C). These results reflect the performance of alternative feature selection strategies when all candidate variables were initially entered for screening, with the aim of identifying the most suitable method for constructing an optimal predictor subset. Thus, the logistic regression result shown in Figure 3 represents a preliminary feature selection performance rather than the final performance of a logistic regression classifier.

After the optimal feature selection method was determined and the retained variables were used to construct the final dataset, seven machine learning classifiers were compared to identify the best-performing predictive model. Comprehensive algorithm comparison (Supplementary Figure S2; Table 1) revealed that XGBoost and random forest models substantially outperformed alternatives. Both attained accuracy exceeding 98%, ROC AUC surpassing 0.98, sensitivity approaching 100%, specificity exceeding 92%, and precision-recall AUC (PR AUC) above 0.97 when evaluated on the SMOTE-processed training data (Figure 4). Their respective Brier scores—0.0240 (ranked first) and 0.0244 (ranked second)—indicated optimal probability calibration. Moderate performance was observed for K-nearest neighbors and support vector machines, while decision trees, naive Bayes, and logistic regression showed markedly inferior metrics (notably Brier scores >0.14, specificity <75%, and PR AUC <0.5), suggesting inadequate predictive stability (detailed metrics in Supplementary Table S3). On the independent test set preserving original class distribution, the XGBoost model maintained strong generalization capability (Supplementary Table S4), achieving an AUC of 0.750—confirming robust discriminatory power. Following clinical sensitivity optimization using a 0.1 decision threshold, the model attained a recall rate of 65.38% (95% CI: 54.2%–75.4%), though precision remained constrained at 8.46%, reflecting inherent sensitivity-precision trade-offs in severely imbalanced data (F1 score = 0.150; accuracy = 72.93%). However, this sensitivity-oriented threshold was associated with a low positive predictive value, indicating a substantial false-positive burden and limiting the model’s suitability as a standalone clinical screening tool. The test set confusion matrix further delineated classification performance of the optimized XGBoost model (Supplementary Table S5).

Systematic evaluation of seven machine learning algorithms (Supplementary Figure S2) employed SMOTE-oversampled training data (ODC positive:negative ratio = 1:5). Through grid search coupled with five-fold cross-validation for hyperparameter optimization, the optimal configuration was determined (learning rate eta = 0.1, max_depth = 6, subsample = 0.8, lambda = 1.0), with early stopping regularization controlling overfitting. As documented in Supplementary Table S6, peak performance occurred at the 44th iteration (test set AUC = 0.750), establishing XGBoost as the definitive predictive framework.

To rigorously assess generalizability and clinical utility, the XGBoost model underwent comprehensive validation using two temporally distinct internal cohorts (2011: n = 477; 2012: n = 502). As presented in Supplementary Table S7, the model demonstrated excellent temporal discriminative capability (Supplementary Figure S3): ROC curve analysis yielded AUC values of 0.783 (95% CI: 0.702–0.864) for the 2011 cohort and 0.744 (95% CI: 0.652–0.835) for the 2012 cohort. Implementation of a low decision threshold (0.1) optimized for screening sensitivity achieved detection rates of 84.2% (95% CI: 73.1%–91.4%) and 70.0% (95% CI: 55.9%–81.2%) in the respective cohorts, successfully capturing over two-thirds of true positive cases—meeting fundamental requirements for early screening instruments.

In external validation using the U.S. NHANES cohort, the XGBoost model achieved an AUC of 0.886, allowing direct comparison with the internal validation performance reported above. The random forest classifier ranked second (AUC = 0.858), followed by radial basis function (RBF) kernel support vector machine (AUC = 0.831). Remaining models demonstrated comparatively limited predictive capacity: K-nearest neighbors (AUC = 0.795), logistic regression (AUC = 0.778), naive Bayes (AUC = 0.759), and decision tree (AUC = 0.667), See Supplementary Table S8 for details. Comparative ROC curves are depicted in Supplementary Figure S4.

SHAP (SHapley Additive exPlanations) was used to interpret the XGBoost model’s prediction patterns. Feature importance analysis (Figure 5) identified insulin concentration as the predominant predictor of obesity-depression comorbidity (mean |SHAP| = 0.052), exerting substantially greater influence than secondary contributors: white blood cell count (0.036), height (0.027), ferritin (0.023), HDL-C (0.021), and age (0.012). SHAP summary plots indicated that elevated insulin, advanced age, increased systolic blood pressure, and higher white blood cell counts were associated with higher predicted ODC probability, whereas greater height and elevated HDL-C concentrations were associated with lower predicted ODC probability. Supplementary Figure S5 provides additional SHAP dependence visualizations.

External validation SHAP analysis using the NHANES cohort further delineated the optimal model’s decision architecture and feature contribution patterns (Supplementary Figure S6). Insulin was reconfirmed as the most influential predictive variable, with its mean absolute SHAP value substantially exceeding those of other features—underscoring its centrality in model discrimination. Key predictors including age, retinol, height, and fasting glucose (fglu) followed in descending order of importance, exhibiting remarkable concordance with KNHANES-derived SHAP results. This cross-cohort reproducibility enhances model credibility and suggests that integrated metabolic, nutritional, and developmental features may provide a robust predictive foundation across populations.

Based on the rigorously validated XGBoost framework, we developed a clinically oriented online prediction instrument (accessible at https://zhlapp.shinyapps.io/Korea_ODC-shap-model/). Implemented via the R Shiny platform, this tool provides interactive risk assessment functionality, enabling healthcare practitioners to input 24 core indicators—including insulin concentration, white blood cell count, height, and age—for real-time ODC risk stratification, as visually demonstrated in Supplementary Figure S7.

Several methodological constraints warrant acknowledgment: First, the development dataset was based on KNHANES 2007–2012, and temporal changes in lifestyle patterns, obesity prevalence, mental health awareness, and public health policies over the past decade may limit the model’s direct applicability to contemporary populations. Future studies should therefore assess the temporal transportability of the model using more recent datasets and update or recalibrate it as needed. Second, depression was defined using PHQ-9 screening rather than a structured clinical diagnosis. Third, missing data were handled using complete-case analysis without imputation. Although this approach avoided additional assumptions introduced by imputation models, it reduced the effective sample size and may have introduced selection bias if the missingness mechanism was not completely random. Future studies should evaluate the robustness of the findings using multiple imputation or other sensitivity analyses. Fourth, though demonstrating robust performance in the NHANES validation cohort, population-specific genetic backgrounds, cultural contexts, and social structures in South Korea may limit global generalizability, necessitating further validation across diverse populations, Future research should prioritize validation and recalibration in East Asian populations that are more comparable (such as Chinese or Japanese populations). Fifth, feature engineering excluded highly correlated variables (r > 0.8); while statistically justified, this process may have omitted biologically relevant indicators (e.g., vitamin A, hematocrit). Sixth, although the model achieved acceptable sensitivity at the optimized decision threshold of 0.1, its precision remained relatively low, indicating a substantial false-positive burden. In practical clinical settings, this may reduce efficiency, contribute to alert fatigue among clinicians, and lead to unnecessary follow-up assessments or anxiety in individuals incorrectly classified as high risk. Therefore, the model should be regarded as a preliminary risk stratification tool rather than a standalone screening or diagnostic instrument. Finally, although obesity in the NHANES cohort was defined using U.S.-appropriate criteria, the absence of U.S.-specific recalibration means that the external validation results should still be interpreted primarily in terms of discrimination and transportability rather than calibration equivalence.

Utilizing a large-scale population-based dataset, this study developed and validated an interpretable XGBoost model for ODC risk prediction in physically inactive adults. SHAP analysis identified insulin as the most influential predictor within the model, indicating that metabolic variables may play an important role in model-based risk stratification. However, these feature-attribution results reflect predictive relevance within the algorithm rather than confirmed biological mechanisms. Accordingly, the findings should be interpreted as hypothesis-generating and supportive of further prospective, experimental, and interventional studies.