Integrating epidemiologic modeling and explainable machine learning to predict and identify factors associated with self-reported depression among adults in Tennessee, United States Authors: Mustapha Aliyu Muhammad, Jamilu Sani, Salad Halane Published: 2026 ABSTRACT AbstractBackgroundDepression is a major public health concern, with Tennessee ranking among the U.S. states with the highest prevalence. Despite its burden, many cases remain undetected due to limited screening and access to mental health services. This study integrated epidemiologic modelling and explainable ML techniques to predict self-reported depression and identify key risk factors among Tennessee adults using the Behavioral Risk Factor Surveillance System (BRFSS) 2023 data.MethodsWe conducted a cross-sectional analysis of 5596 adults from the 2023 Tennessee BRFSS, representing 5,569,707 weighted respondents. The primary outcome was lifetime self-reported diagnosis of depression. The Oregon BRFSS 2023 was used as an external validation dataset. Eight machine learning algorithms were trained using 5-fold stratified cross-validation. Model performance was evaluated using AUROC, PR-AUC, accuracy, precision, recall, F1-score, balanced accuracy, DeLong’s test, and McNemar’s test, while model interpretability was assessed using SHapley Additive exPlanations (SHAP).ResultsThe weighted prevalence of self-reported depression among Tennessee adults was 27.3%. Among the evaluated algorithms, XGBoost, Gradient Boosting, Random Forest, and Logistic Regression demonstrated the strongest and highly comparable external validation performance. DeLong’s test for AUROC and paired bootstrap resampling for PR-AUC showed no statistically significant differences among these four leading models. McNemar’s test produced a similar pattern for paired classification errors. SHAP interpretation identified sex, ACEs category, memory decline, disability category, race/ethnicity, poor physical activity, and age group as the most influential predictors of self-reported depression.ConclusionsThis study demonstrates the utility of integrating explainable machine learning approaches to predict and identify factors associated with self-reported depression, thereby enhancing the use of public health surveillance systems in early identification of high-risk populations and informing targeted mental health interventions.Supplementary InformationThe online version contains supplementary material available at 10.1007/s44192-026-00506-4.Keywords: Depression, Machine learning, SHAP analysis, Behavioral risk factor surveillance system (BRFSS), Public health informatics, Risk prediction, Population surveillance, Mental health Background Depression is a widespread and serious public health concern. It is a common mental health condition that significantly impairs an individual’s psychological well-being and social functioning, ultimately diminishing their quality of life. It also imposes substantial emotional and financial burdens on the families of those affected. The condition is typically characterized by ongoing fatigue, a consistently low mood, loss of interest in activities, and difficulty concentrating [1]. The impact of depression extends beyond individual suffering, it is also associated with reduced quality of life, increased risk of chronic illness, and even elevated mortality rates [2]. Globally, this disorder affects more than 300 million people and has become the leading cause of disability worldwide [3, 4]. In the United States, depression is recognized as a major contributor to mortality, morbidity, disability, and economic costs [2]. Recent estimates indicate that approximately one in five U.S. adults (18.5%) has been diagnosed with a depressive disorder in their lifetime [5]. The annual estimated cost of treating depression in the US is reported to be about $210 billion [6]. According to the CDC, Tennessee ranks third among U.S. states in depression prevalence, with 24.4% of adults reporting a diagnosis of depression, surpassed only by Kentucky and West Virginia [5]. Without timely intervention, depression’s burden is projected to grow, and it is expected to become the leading cause of disease burden globally by 2030, underscoring the urgency for improved detection and prevention efforts [7]. Early detection of depression remains challenging, despite its high prevalence. Many individuals with depression go undiagnosed or untreated even in high-income countries, nearly half of people suffering from depression are not receiving the care they need [4, 8]. Contributing factors include social stigma, lack of routine mental health screening in primary care, and barriers to accessing behavioral health services [9]. Consequently, depression often remains hidden until it reaches a more severe stage, leading to prolonged suffering and higher risk of complications. This treatment gap highlights the need for proactive surveillance and community-level data to identify those at risk. Public health experts emphasize that early identification and intervention are essential to reducing depression’s long-term impact [4]. Robust surveillance data can help fill this gap by providing insights into who is affected and what risk factors are associated with depression in the general population, enabling health authorities to respond sooner with targeted outreach and resources. To monitor conditions like depression in the general adult population, public health agencies rely on surveillance systems that capture data beyond clinical settings. The Behavioral Risk Factor Surveillance System (BRFSS) is a cornerstone of such surveillance in the U.S., providing a source of recent data on health status and behaviors. BRFSS is a publicly available, ongoing state-based health survey that each year collects information from a large, representative sample of adults in every state on topics ranging from chronic conditions to health behaviors and mental health status [10]. These extensive scale and sampling design ensure that BRFSS data are broadly generalizable to the non-institutional adult population. BRFSS is the primary source of timely and accurate health data, used to identify emerging health problems and to inform the development and evaluation of interventions [10]. Notably, BRFSS includes measures related to mental health (such as lifetime self-reported depression diagnoses), making it a valuable tool for tracking depression prevalence and trends across communities. Additionally, the BRFSS collects extensive data on established risk factors for depression, including sociodemographic characteristics, general and mental health status, physical well-being, disability, and health behaviors such as tobacco and alcohol use [10]. It also captures information on comorbid conditions like cancer, hypertension, diabetes, stroke, heart attack, cognitive impairment, and adverse childhood experiences (ACEs), all of which are known to be associated with increased risk of depression [11–13]. This study capitalizes on the latest available data for the state of Tennessee and ensures that our analysis addresses the current scope of depression in the state’s adult population and thereby supporting contemporary public health planning. Traditional epidemiological analyses of surveillance data (for example, using logistic regression to identify individual risk factors) provide important insights but have limitations when dealing with complex, high-dimensional data. Depression, like many health outcomes, results from an interplay of diverse factors: demographics, socioeconomic status, health behaviors, comorbid conditions, and more, which may interact in non-linear ways. Machine learning (ML) methods offer a powerful complement to traditional analytic approaches by automatically identifying patterns in large datasets that might be missed by parametric analysis [14]. Unlike conventional statistical models that are typically linear and require pre-specified predictors, modern ML algorithms can handle non-linear relationships and high-dimensional feature spaces with minimal human guidance [15]. Prior research has demonstrated that ML-based risk models can improve the accuracy of depression predictions compared to traditional methods, effectively handling “mass data” and uncovering subtle correlations [4, 13, 16]. This study aims to integrate traditional epidemiological methods and machine learning models to predict and identify risk factors of lifetime self-reported depression using demographic, behavioral, and health-related variables from the 2023 BRFSS Tennessee dataset. We also employ SHapley Additive exPlanations (SHAP) to identify and interpret the most influential risk factors contributing to the prediction of the outcome. This approach can yield new insights, thus helping public health officials identify vulnerable subpopulations in the community. The detection of patterns and high-risk profiles can guide targeted prevention efforts, inform resource allocation, and ultimately support interventions to reduce the burden of depression in Tennessee and beyond. Methods Study design and data source This study employed a cross-sectional design, utilizing secondary data from the 2023 Behavioral Risk Factor Surveillance System (BRFSS) specific to the state of Tennessee. The BRFSS is a nationally representative, state-based surveillance system conducted annually by the Centers for Disease Control and Prevention (CDC) in collaboration with U.S. states and territories [10]. It collects information on health-related behaviors, chronic conditions, and preventive health practices among non-institutionalized adults aged 18 years and older via telephone interviews [10]. The 2023 BRFSS dataset was selected for its comprehensive coverage of variables related to mental health, physical health, sociodemographic characteristics, and health-related behaviors. For this study, we extracted the subset of respondents residing in Tennessee and included only those with complete responses to the outcome variable, lifetime self-reported depression diagnosis. After applying these inclusion criteria, the final analytical sample comprised of 5,596 adult individuals (5,569,707 weighted population). For transparency and reproducibility, all scripts used for the epidemiological and machine learning modeling are available in a public GitHub repository at: https://github.com/Mustapha0287/Explainable-ML-Self-Reported-Depression-Tennessee-Adults-Evidence-from-the-2023-BRFSS/tree/main. Study variables The primary outcome variable in this study was lifetime self-reported depression diagnosis, based on whether a respondent reported ever being told by a health professional that they had a depressive disorder. This aligns with the standard BRFSS approach for assessing lifetime prevalence of depression. A range of predictor variables were selected based on their theoretical and empirical associations with depression, as documented in previous literature [13, 17–20]. These variables span across four broad domains: sociodemographic characteristics, health status, health behaviors, and psychosocial or medical risk factors. Sociodemographic variables included age (categorized into six age groups), sex, race/ethnicity (five categories), marital status (binary indicator of married vs. not married), education level (categorized as less than high school, high school graduate, some college, and college graduate), employment status (employed vs. unemployed), and home ownership status. These variables were included to capture structural and socioeconomic factors that may influence mental health outcomes. Health status indicators included general self-reported well-being, the number of days of poor physical health, disability status based on functional limitations, and reported experiences of memory loss or cognitive decline. These variables reflect both perceived and clinically relevant aspects of health that are frequently associated with depression. Health behavior variables comprised physical activity in the past 30 days, smoking status, alcohol consumption in the past 30 days, and current health insurance coverage. These behaviors and access-related variables were included as potential protective or risk factors influencing mental health. Additional psychosocial and clinical predictors included veteran status, pregnancy status, caregiving responsibilities, cancer diagnosis, the presence of comorbid chronic conditions (grouped into categories), and exposure to ACEs. Urban-rural residence classification was also included to account for geographic and environmental differences in mental health risk and service access. (Supplementary Table S1) Data processing and analysis A comprehensive analytical approach was employed for data processing and analysis. Preliminary exploratory and statistical analyses were conducted using Stata/BE version 19, where the complex sampling design of the BRFSS was accounted for using the svyset commands. Weighted frequencies and percentages were computed for categorical variables to generate population-representative descriptive summaries. Bivariate associations between self-reported depression diagnosis and each predictor were assessed using design-adjusted chi-square tests. Weighted crude odds ratios (CORs) and corresponding 95% confidence intervals (CIs) were estimated through bivariate logistic regression. Variables demonstrating associations at a liberal threshold (p ≤ 0.20) were retained through backward elimination for inclusion in the multivariable logistic regression model, from which adjusted odds ratios (AORs) and 95% CIs were obtained [21]. Statistical significance was defined as p ≤ 0.05, with p ≤ 0.001 considered highly significant. The machine learning analyses were subsequently performed in a Google Colab Python 3.12.12 environment using Jupyter Notebook. This workflow leveraged widely used data science libraries, including Pandas 2.2.2, NumPy 2.0.2, Scikit-learn 1.6.1, SciPy 1.16.3, Matplotlib 3.10.0, Seaborn 0.13.2, and SHAP 0.50.0, to support data cleaning, transformation, exploratory analysis, model development, and visualization. These tools facilitated the implementation and evaluation of multiple machine learning algorithms, performance assessment on an independent hold-out test set, and interpretation of model predictions using SHAP to quantify feature importance and directionality of effects. (Fig. 1) Machine learning workflow for predicting self-reported depression diagnosis in Tennessee Data pre-processing and handling missingness The raw BRFSS data underwent a comprehensive preprocessing phase to ensure suitability for machine learning analysis. As all variables included in the analysis were categorical, normalization or scaling was not required. Outliers were assessed for each variable, and rare response categories were either consolidated with conceptually similar categories or excluded when deemed analytically insignificant to improve model stability. The data cleaning process further involved resolving inconsistencies, recoding values for clarity, and standardizing category definitions across variables. Several individual variables were consolidated into composite measures to enhance interpretability and analytical relevance. Specifically, variables capturing ACEs, comorbid conditions, and functional disabilities were aggregated to create categorical indicators reflecting the number of reported outcomes within each domain. Missingness across analytic variables varied between the Tennessee training dataset and the Oregon external validation dataset. In the Tennessee dataset, missingness ranged from 0.3% for self-rated health to 27.7% for ACEs, with most variables demonstrating relatively low levels of missingness; excluding ACEs, missingness across the remaining variables was below 8.9%. In the Oregon external validation dataset, missingness ranged from 0.10% for pregnancy status to 44.58% for memory decline, with higher missingness also observed for ACEs (28.20%) and caregiver status (15.62%). Most other Oregon variables had low to moderate missingness, generally below 8.2%, suggesting that missingness was concentrated in a small number of variables rather than distributed broadly across the dataset (Supplementary Table S2 & S3). Given the categorical nature of all analytic predictors, missing values were addressed using multiple imputation by chained equations with the MICE package in R (version 4.6.0) [22]. Binary variables were imputed using logistic regression, variables with more than two categories were imputed using polytomous logistic regression, and BRFSS survey design variables were excluded from the imputation process to preserve the original sampling design information [23]. This preprocessing strategy streamlined the dataset, improved interpretability, and supported robust model performance. Addressing class imbalance using synthetic minority over-sampling technique for nominal (SMOTEN) Exploratory analyses identified class imbalance in the outcome distribution, with fewer participants reporting a lifetime diagnosis of depression compared with those without depression. Because all predictors in the analytic dataset were categorical, we used the Synthetic Minority Over-sampling Technique for Nominal (SMOTEN) rather than the standard SMOTE algorithm. Unlike traditional SMOTE, which creates synthetic observations through numeric interpolation and is more appropriate for continuous predictors, SMOTEN is designed for categorical predictors and generates synthetic minority-class examples based on categorical feature patterns [24, 25]. To reduce model bias toward the majority class and data leakage, SMOTEN was applied only to the Tennessee training data during model development and cross-validation. The Oregon external validation dataset was not resampled, preserving its original class distribution and allowing an unbiased assessment of model generalizability. After SMOTEN resampling, the training data achieved balanced class proportions, improving the opportunity for algorithms to learn patterns associated with self-reported depression while maintaining the natural outcome distribution in the external validation dataset. (Fig. 2) Impact of SMOTE on minority-class representation in the training dataset Feature selection A comprehensive feature selection process was conducted to identify the most relevant predictors of self-reported depression. Exploratory data analysis (EDA), including descriptive statistics and visualizations, was used to examine the distribution of each variable and its potential relationship with the outcome. Bivariate analyses were employed to assess the strength and direction of associations between individual predictors and depression status. Thereafter, the epidemiologic model was fitted in Stata, using backward elimination (p ≤ 0.20) to retain potentially informative variables [21]. To further refine the predictor set for the machine learning modelling, Cramer’s V statistic was used to measure associations among categorical predictors, helping to detect and minimize multicollinearity. In addition, Recursive Feature Elimination (RFE) was implemented to iteratively remove less informative variables, thereby improving model efficiency. This multi-pronged approach, grounded in statistical reasoning, machine learning techniques, and existing literature, ensured the selection of a robust and interpretable set of features for model development. Feature importance Feature importance was evaluated using SHAPley Additive exPlanations (SHAP), a unified and model-agnostic interpretability framework that enables transparent interpretation of complex machine learning models [26, 27]. SHAP values quantified the marginal contribution of each predictor to the predicted probability of self-reported depression, allowing assessment of both the magnitude and direction of feature effects. This approach provided granular insight into how individual variables influenced model predictions and facilitated identification of the most influential risk and protective factors [27]. SHAP analyses were performed for the best-performing models and visualized using beeswarm and bar plots to summarize feature importance and distributional effects. To enhance interpretability, SHAP values were examined at both the encoded feature level and the aggregated variable level, with one-hot-encoded categories recombined into their original predictors. This dual representation ensured consistency in interpretation while preserving model fidelity. The SHAP waterfall plot for individual prediction index 0 was used to illustrate how specific respondent characteristics contributed to the model’s predicted probability of self-reported depression. Features with positive SHAP values pushed the prediction toward depression diagnosis, whereas features with negative SHAP values reduced the predicted probability, providing a transparent case-level explanation of the model decision. SHAP dependence plots for the selected top variables further showed how categories of key predictors contributed differently to predicted depression risk across respondents. Model development and optimization Given the complex and multifactorial nature of self-reported depression, a diverse set of machine learning algorithms was implemented to capture heterogeneous patterns in the data. The selected models represented multiple learning paradigms, including linear (Logistic Regression), probabilistic (Naïve Bayes), kernel-based (Support Vector Machine), tree-based (Decision Tree and Random Forest), distance-based (K-Nearest Neighbors), and ensemble boosting methods (Gradient Boosting and Extreme Gradient Boosting [XGBoost]). This comprehensive modeling strategy enabled systematic comparison of predictive performance across distinct algorithmic approaches. For model optimization, default hyperparameters were initially applied, and where appropriate, tuning was conducted using grid search with 5-fold stratified cross-validation to refine performance (Supplementary Table S4). All models were implemented within a consistent preprocessing framework and trained on the SMOTEN-balanced training dataset. Logistic Regression served as the baseline model, providing a transparent and interpretable linear classification approach. Naïve Bayes was included as a probabilistic classifier well-suited for high-dimensional categorical data [28]. Support Vector Machine was employed for its capacity to model non-linear decision boundaries using kernel functions. Decision Tree and Random Forest models were utilized to capture hierarchical decision structures, with Random Forest benefiting from ensemble averaging to enhance stability and reduce overfitting [28]. The K-Nearest Neighbors algorithm classified observations based on proximity within the feature space, while boosting methods, Gradient Boosting and XGBoost, were incorporated for their iterative learning processes that sequentially minimize classification errors and improve predictive accuracy [28]. Model training and evaluation Model performance was evaluated using a structured training and validation framework. The Tennessee BRFSS dataset was used for model training and internal cross-validation, while the Oregon BRFSS dataset was used as an external validation dataset to assess model generalizability. Oregon was selected as the external validation dataset because it contributed a similar proportion of the overall BRFSS sample as Tennessee (1.44% vs. 1.33%) and was among the few states with available responses for both the adverse childhood experiences (ACEs) and caregiver modules [10]. Stratified 5-fold cross-validation was applied within the Tennessee training data to preserve the distribution of self-reported depression during model development and hyperparameter tuning. To address class imbalance, SMOTEN was applied exclusively to the training data, while the test set was left unaltered to provide an unbiased assessment of model generalizability. Each machine learning algorithm was trained on the resampled training data and evaluated on the Oregon test set. Model performance was assessed using a comprehensive set of metrics, including accuracy, precision, recall, F1 score, balanced accuracy, area under the receiver operating characteristic curve (AUROC), and precision–recall area under the curve (PR-AUC). In addition, 5-fold stratified cross-validation was conducted within the training dataset to obtain robust performance estimates and reduce the risk of overfitting. Performance metrics were averaged across folds to ensure stability. Confusion matrices were generated to summarize true positives, false positives, true negatives, and false negatives, and AUROC and PR-AUC curves were plotted to visually compare classification performance across models. This evaluation strategy enabled balanced assessment of both discrimination and classification error patterns while preventing information leakage between training and test data. Model selection Model selection was based on a systematic comparison of performance metrics across all evaluated machine learning algorithms, with AUROC selected as the primary indicator of discriminative performance. AUROC was prioritized because it evaluates the model’s ability to distinguish between adults with and without self-reported depression across a range of classification thresholds, rather than relying on a single cutoff point. Additional metrics, including accuracy, precision, recall, F1-score, balanced accuracy, and PR-AUC, were used to provide complementary information on classification performance. Precision quantified the proportion of predicted depression cases that were truly positive, while recall (sensitivity) measured the model’s ability to correctly identify individuals with self-reported depression [29]. The F1 score, representing the harmonic mean of precision and recall, was emphasized for its ability to balance these measures under conditions of class imbalance [30]. In addition, the precision–recall area under the curve (PR-AUC) was evaluated to provide a more informative assessment of performance when the outcome was imbalanced [31]. Confusion matrices were examined to characterize classification error patterns by summarizing true positives, true negatives, false positives, and false negatives for each model. DeLong’s test was used to assess whether AUROC differences between algorithms on the external validation dataset were statistically significant. Paired bootstrap resampling was used to compare PR-AUC differences while preserving the paired structure of predictions from the same validation participants. McNemar’s test was also applied to compare paired classification errors at the selected probability threshold, providing additional evidence on whether competing algorithms differed significantly in their classification decisions (Fig. 1). Results Sociodemographic characteristics and prevalence of self-reported depression Among 5,596 adult respondents included in the analysis, representing an estimated 5.57 million Tennessee adults after weighting, 27.3% reported having been diagnosed with depression, while 72.7% reported no such diagnosis. The weighted sample was predominantly aged 65 years and older (24.1%), followed by adults aged 25–34 years (16.7%) and 55–64 years (16.1%). Females constituted a slight majority of the population (52.1%). The racial and ethnic composition was largely non-Hispanic White (72.4%), with Black respondents accounting for 15.1% and Hispanic respondents for 6.3%. Nearly one-third of respondents had a high school education (32.4%), while 26.8% were college graduates. Approximately half of the population was married (50.6%), and 60.4% were employed. Most respondents resided in urban areas (88.2%) and owned their homes (70.2%). The majority reported having no children (67.4%), while smaller proportions reported one or more children. Most respondents reported having health insurance coverage (90.6%), and 11.3% identified as veterans. Overall, self-reported wellbeing was favorable, with 76.9% reporting good or better health, although 23.1% reported fair or poor health. Obesity was highly prevalent, affecting 38.2% of the weighted population, and 66.3% reported at least one comorbid condition. A cancer diagnosis was reported by 12.8% of respondents, while pregnancy was rare (0.8%). Indicators of cognitive and functional health showed that 30.5% reported memory decline, and more than one-third reported either a single disability (16.3%) or multiple disabilities (23.5%). Additionally, 25.5% reported serving in a caregiver role. With respect to health behaviors, 41.4% reported current or former smoking, and 47.4% reported alcohol use in the past 30 days. Exposure to adverse childhood experiences was common, with 45.2% reporting one to three ACEs and 25.9% reporting four or more. Finally, 43.7% reported at least one day of poor physical health in the past month, and 73.9% reported engaging in physical activity during the previous 30 days. (Table 1) Table 1Weighted and unweighted distribution of variables associated with self-reported depression diagnosisSNVariableCategoryUnweighted Frequencyn = 5,596Weighted Frequencyn = 5,569,707Weighted percentage Outcome 1.Self-reported Depression diagnosisNo4,0724,050,87072.7%Yes15241,518,83727.3% Sociodemographic characteristics 2.Age Group18–24383669,55912.0%25–34665927,68416.7%35–44758880,42815.8%45–54910852,27515.3%55–641,024895,98916.1%65+1,8561,343,77324.1%3.SexMale2,6792,665,17747.9%Female2,9172,904,53052.1%4.Race/EthnicityWhite4,4614,034,74672.4%Black687840,20015.1%Other152206,3533.7%Multiracial122139,6942.5%Hispanic174348,7156.3%5.Education Level< High School407631,12211.3%HS Graduate1,6761,804,46832.4%Some college1,4821,640,75329.5%College graduate2,0311,493,36426.8%6.Marital StatusMarried2,7922,818,36250.6%Unmarried2,8042,751,34649.4%7.Employment StatusEmployed3,0953,362,94160.4%Unemployed2,8042,751,34639.6%8.Home OwnershipOwn3,8393,912,00070.2%Rent1,3861,278,14222.9%Other371379,5666.8%9.UrbanicityUrban4,9284,914,52688.2%Rural668655,18111.8%10.Number of ChildrenNo children4,0653,751,90867.4%One695764,47613.7%Two500595,11410.7%Three229284,9615.1%Four70104,4311.9%Five+3768,8161.2% Health Access & Health Conditions 11.Health Insurance CoverageInsured51505,048,32890.6%Not insured446521,3809.4%12.Veteran StatusYes710628,84511.3%No4,8864,940,86288.7%13.Self-reported WellbeingGood/Better4,2534,280,60376.9%Fair/Poor1,3431,289,10423.1%14.BMI CategoryUnderweight8583,9631.5%Normal1,5121,543,88727.7%Overweight1,8831,816,05432.6%Obese2,1162,125,80338.2%15.Comorbidity Burden0 comorbidities1,5841,871,91633.6%1–2 comorbidities2,3192,252,70040.4%3 + comorbidities1,6931,445,09125.9%16.Cancer DiagnosisYes951712,85912.8%No4,6454,856,84987.2%17.Pregnancy StatusYes2544,7850.8%No5,5715,524,92399.2% Cognitive and functional status 18.Memory DeclineYes1,6451,699,02530.5%No3,9513,870,68369.5%19.Functional Disability BurdenNo disability3,3013,350,09160.1%Single disability977908,19216.3%Multiple disabilities1,3181,311,42423.5%20.Caregiver RoleYes1,3921,422,37825.5%No4,2044,147,32974.5% Health behavior & exposure 21.Smoking StatusEvery day689637,71811.4%Somedays248288,4215.2%Former1,4991,380,44424.8%Never3,1603,263,12558.6%22.Alcohol UseYes2,5682,639,35747.4%No3,0282,930,35152.6%23.Adverse Childhood Experiences (ACEs)0 ACEs1,8101,606,37728.8%1–3 ACEs2,4222,519,15445.2%4 + ACEs1,3641,444,17625.9% Self-reported health status 24.Poor Physical Health DaysZero days3,1553,136,54056.3%1–13 days1,4931,572,54228.2%14 + days948860,62515.5%25.Physical Activity in the last 30 daysYes4,0694,115,17573.9%No1,5271,454,53226.1% Weighted and unweighted distribution of variables associated with self-reported depression diagnosis Factors associated with self-reported depression diagnosis In bivariate analyses, several sociodemographic, health, and behavioral factors were significantly associated with self-reported depression diagnosis. The prevalence of depression varied by age, with lower odds observed among older adults, particularly those aged 65 years and older (COR = 0.55, 95% CI: 0.41–0.75), compared with adults aged 18–24 years. Females had substantially higher odds of reporting depression than males (COR = 2.35, 95% CI: 1.99–2.78). Racial and ethnic differences were observed, with Black respondents exhibiting lower odds of depression relative to White respondents (COR = 0.62, 95% CI: 0.48–0.81). Lower educational attainment, unmarried status, unemployment, renting or other non-ownership housing arrangements, and poorer self-reported wellbeing were each associated with higher odds of depression. Health-related factors showed strong associations, including increasing comorbidity burden, memory decline, functional disability, ACEs, frequent poor physical health days, smoking, and physical inactivity. Notably, individuals reporting four or more ACEs had markedly elevated odds of depression (COR = 5.6, 95% CI: 4.44–7.07), as did those with multiple functional disabilities (COR = 3.83, 95% CI: 3.15–4.67). After adjustment for potential confounders in the multivariable logistic regression model, several associations remained statistically significant. Increasing age was independently protective, with substantially reduced odds of depression observed among adults aged 35–44 years (AOR = 0.54, 95% CI: 0.34–0.85), 45–54 years (AOR = 0.58, 95% CI: 0.38–0.89), 55–64 years (AOR = 0.43, 95% CI: 0.28–0.66), and 65 years or older (AOR = 0.22, 95% CI: 0.15–0.34). Female sex remained strongly associated with higher odds of depression (AOR = 2.82, 95% CI: 2.29–3.47). Black respondents continued to show lower odds compared with White respondents (AOR = 0.41, 95% CI: 0.30–0.56). Poor or fair self-reported wellbeing (AOR = 1.47, 95% CI: 1.13–1.92), increasing comorbidity burden (AOR = 2.34 for 1–2 comorbidities; AOR = 3.29 for ≥ 3 comorbidities), memory decline (AOR = 0.5, 95% CI: 0.39–0.65 for no decline), and functional disability demonstrated strong independent associations. Exposure to adverse childhood experiences remained one of the most influential predictors, with individuals reporting four or more ACEs experiencing more than threefold higher odds of depression (AOR = 3.56, 95% CI: 2.71–4.68). Additionally, frequent poor physical health days and lack of physical activity were independently associated with increased odds of depression, underscoring the multifactorial nature of depression risk in this population. (Table 2) Table 2Bivariate & Logistic Regression Analysis (COR & AOR) of Factors Associated with Self-reported Depression DiagnosisSNSelf-reported depression diagnosisVariablesSelf-reported Depression[Yes]Frequency (%)Self-reported Depression[No]Frequency (%)Crude OR[95% C.I.]Adjusted OR[95% C.I.] Sociodemographic characteristics 1. Age Group 18–24206,957 (30.9%)462,602 (69.1%)RefRef25–34303,089 (33.7%)624,595 (67.3%)1.08 (0.78–1.51)1.12 (0.75–1.68)35–44258,852 (29.4%)621,575 (70.6%)0.93 (0.66–1.31)0.54 (0.34–0.85)*45–54226,204 (26.5%)626,071 (73.5%)0.81 (0.59–1.11)0.58 (0.38–0.89) *55–64257,518 (28.7%)638,472 (71.3%)0.90 (0.65–1.25)0.43 (0.28–0.66) **65+266,217 (19.8%)1,077,556 (80.2%)0.55 (0.41–0.75) **0.22 (0.15–0.34) **2. Sex Male498,837 (18.7%)2,166,340 (81.3%)RefRefFemale1,020,000 (35.1%)1,884,530 (64.9%)2.35 (1.99–2.78) **2.82 (2.29–3.47) **3. Race/Ethnicity White1,158,846 (28.7%)2,875,900 (71.3%)RefRefBlack168,867 (20.1%)671,332 (79.9%)0.62 (0.48–0.81) **0.41 (0.3–0.56) **Other55,623 (27.0%)150,730 (73.0%)0.92 (0.52–1.62)0.51 (0.24–1.08)Multiracial43,714 (31.3%)95,980 (68.7%)1.13 (0.70–1.82)0.57 (0.34–0.95)*Hispanic91,788 (26.3%)256,928 (73.7%)0.89 (0.58–1.37)0.83 (0.51–1.37)4. Education level < High School216,716 (34.3%)414,405 (65.7%)RefHS Graduate496,638 (27.5%)1,307,831 (72.5%)0.73 (0.58–1.00) *Some college481,129 (29.3%)1,159,624 (70.7%)0.79 (0.58–1.09)College graduate324,354 (21.7%)1,169,010 (78.3%)0.53 (0.39–0.73) **5. Marital status Married619,707 (22.0%)2,198,655 (78.0%)RefRefUnmarried899,130 (32.7%)1,852,216 (67.3%)1.72 (1.46–2.03) **1.31 (1.07–1.6) *6. Employment status Employed805,808 (24.0%)2,557,133 (76.0%)RefRefUnemployed713,029 (32.3%)1,493,737 (67.7%)1.51 (1.28–1.79) **1.19 (0.94–1.52)7. Home ownership Own928,575 (23.7%)2,983,425 (76.3%)RefRent443,240 (34.7%)834,902 (65.3%)1.71 (1.43–2.04) **Other147,023 (38.7%)232,543 (61.3%)2.03 (1.50–2.75) **8. Urbanicity Urban1,341,998 (27.3%)3,572,528 (72.7%)RefRefRural176,839 (27.0%)478,342 (73.0%)0.98 (0.77–1.25)0.78 (0.60–1.01)9. Number of children No children1,024,459 (27.3%)2,727,449 (72.7%)RefRefOne217,819 (28.5%)546,657 (71.5%)1.06 (0.84–1.33)1.0 (0.75–1.33)Two158,828 (26.7%)436,286 (73.3%)0.97 (0.72–1.3)0.88 (0.61–1.28)Three82,943 (29.1%)202,019 (70.9%)1.09 (0.75–1.6)1.08 (0.69–1.69)Four30,322 (29.0%)74,109 (71.0%)1.09 (0.61–1.96)0.68 (0.32–1.41)Five+4,466 (6.5%)64,351 (93.5%)0.18 (0.07–0.52) **0.14 (0.05–0.44) ** Health access & health conditions 10. Health insurance coverage Insured1,352,075 (26.8%)3,696,253 (73.2%)RefNot insured166,762 (32.0%)354,617 (68.0%)1.29 (0.97–1.7)11. Veteran status Yes146,026 (23.2%)482,820 (76.8%)RefRefNo1,372,811 (27.8%)3,568,050 (72.2%)1.27 (0.99–1.63)0.74 (0.56–0.97) *12. Self-reported Wellbeing Good/Better (Ref)946,671 (22.1%)3,333,932 (77.9%)RefRefFair/Poor572,166 (44.4%)716,938 (55.6%)2.81 (2.34–3.38) **1.47 (1.13–1.92) *13. BMI category Underweight36,483 (43.5%)47,480 (56.5%)RefNormal397,155 (25.7%)1,146,732 (74.3%)0.45 (0.24–0.85) *Overweight410,721 (22.6%)1,405,333 (77.4%)0.38 (0.2–0.71) *Obese674,479 (31.7%)1,451,325 (68.3%)0.60 (0.32–1.13)14. Comorbidity Burden 0 comorbidities349,245 (18.7%)1,522,671 (81.3%)RefRef1–2 comorbidities636,242 (28.2%)1,616,458 (71.8%)1.72 (1.39–2.12) **2.34 (1.8–3.04) **3 + comorbidities533,350 (36.9%)911,742 (63.1%)2.55 (2.04–3.18) **3.29 (2.4–4.5) **15. Cancer diagnosis Yes208,647 (29.3%)504,212 (70.7%)RefNo1,310,190 (27.0%)3,549,659 (73.0%)0.89 (0.71–1.12)16. Pregnancy status Yes15,633 (34.9%)29,152 (65.1%)RefNo1,503,205 (27.2%)4,021,718 (72.8%)0.70 (0.21–2.28) Cognitive and functional status 17. Memory decline Yes720,591 (42.4%)978,433 (57.6%)RefRefNo798,246 (20.6%)3,072,437 (79.4%)0.35 (0.3–0.42) **0.5 (0.39–0.65) **18. Functional disability burden No disability594,181 (17.7%)2,755,911 (82.3%)RefRefSingle disability331,136 (36.5%)577,056 (63.5%)2.66 (2.13–3.33) **1.79 (1.39–2.32) **Multiple disabilities593,521 (45.3%)717,903 (54.7%)3.83 (3.15–4.67) **1.93 (1.48–2.51) **19. Caregiver role Yes448,630 (31.5%)973,749 (68.5%)RefNo1,070,208 (25.8%)3,077,121 (74.2%)0.75 (0.63–0.91) * Health behavior & exposure 20. Smoking status Every day259,811 (40.7%)377,907 (59.3%)RefRefSomedays130,848 (45.4%)157,573 (54.6%)1.21 (0.8–1.83)1.5 (0.9–2.52)Former365,771 (26.5%)1,014,673 (73.5%)0.52 (0.4–0.68) **0.84 (0.63–1.13)Never762,407 (23.4%)2,500,718 (76.6%)0.44 (0.35–0.57) **0.76 (0.57–1.02)21. Alcohol use Yes676,720 (25.6%)1,962,637 (74.4%)RefNo842,118 (28.7%)2,088,233 (71.3%)1.17 (0.99–1.38)22. Adverse childhood experiences (ACEs) 0 ACEs220,257 (13.7%)1,386,120 (86.3%)RefRef1–3 ACEs618,306 (24.5%)1,900,848 (75.5%)2.05 (1.63–2.57) **1.74 (1.35–2.24) **4 + ACEs680,275 (47.1%)763,901 (52.9%)5.6 (4.44–7.07) **3.56 (2.71–4.68) ** Self-reported health status 23. Poor physical health days Zero days (Ref)543,116 (17.3%)2,593,424 (82.7%)RefRef1–13 days551,288 (35.1%)1,021,254 (64.9%)2.58 (2.12–3.13) **1.61 (1.29–2.02) **14 + days424,433 (49.3%)436,192 (50.7%)4.65 (3.72–5.8) **1.82 (1.35–2.46) **24. Physical activity in the last 30 days Yes946,671 (22.1%)3,333,932 (77.9%)RefNo572,166 (44.4%)716,938 (55.6%)1.68 (1.41–2.0) **Abbreviations: C.I.: Confidence Interval, OR: Odds Ratio, Ref: Reference Group, ** p-value ≤ 0.001, * p-value ≤ 0.05 Bivariate & Logistic Regression Analysis (COR & AOR) of Factors Associated with Self-reported Depression Diagnosis Abbreviations: C.I.: Confidence Interval, OR: Odds Ratio, Ref: Reference Group, ** p-value ≤ 0.001, * p-value ≤ 0.05 Class imbalance adjustment and assessment of predictor associations Figure 2 illustrates the effect of applying the Synthetic Minority Over-sampling Technique for Nominal (SMOTEN) to the Tennessee training dataset to address class imbalance in self-reported depression status. Prior to resampling, respondents reporting depression constituted a smaller proportion of the training data compared with those without depression (1524 vs. 4072). Following SMOTEN, the minority depression class was up-sampled to achieve balance with the majority class, resulting in 4072 respondents in each outcome category. Because all predictors were categorical, SMOTEN was used instead of standard SMOTE to generate synthetic minority-class observations based on categorical feature patterns rather than numeric interpolation. Importantly, resampling was applied only to the Tennessee training data during model development, while the Oregon external validation dataset was left unchanged, preserving its original distribution of 4569 respondents without depression and 1615 respondents with depression for unbiased external model evaluation. Figure 3 presents the Cramér’s V heatmap assessing the strength of association among categorical predictors included in the analysis. Overall, associations between variables were predominantly weak to moderate, indicating limited redundancy among predictors. Stronger correlations were observed between employment status and age group. However, none of the observed associations approached levels indicative of problematic multicollinearity (≥ 0.7), supporting the inclusion of these variables in both traditional and machine learning models [32, 33]. Strength of association between categorical predictors using Cramér’s V heatmap Model performance evaluation Across all evaluated machine learning algorithms, predictive performance varied, reflecting trade-offs between discrimination, sensitivity, precision, and classification error patterns in the Oregon external validation dataset. XGBoost, Logistic Regression, and Gradient Boosting demonstrated the strongest and highly comparable external validation performance, with AUROC values of 0.751, 0.750, and 0.749, respectively. These differences were minimal, indicating that the three leading models performed similarly in discriminating between adults with and without self-reported depression. XGBoost achieved accuracy of 0.711, precision of 0.460, recall of 0.619, F1-score of 0.528, balanced accuracy of 0.681, and PR-AUC of 0.528. Logistic Regression showed very similar performance, with accuracy of 0.716, precision of 0.466, recall of 0.607, F1-score of 0.528, balanced accuracy of 0.681, and the highest PR-AUC among the models (0.535). Gradient Boosting also performed comparably, with accuracy of 0.711, precision of 0.459, recall of 0.598, F1-score of 0.519, balanced accuracy of 0.674, and PR-AUC of 0.525. Other models showed different performance trade-offs. Random Forest achieved the highest accuracy (0.721) but lower recall (0.538), suggesting stronger identification of non-depression cases but fewer correctly identified depression cases. Naïve Bayes achieved the highest recall (0.758) but had the lowest precision (0.355) and accuracy (0.578), indicating a tendency to classify more respondents as depression cases at the cost of substantially more false positives. KNN and Decision Tree showed lower overall discrimination, with AUROC values of 0.728 and 0.689, respectively. Confusion matrix analyses further illustrated these differences in model behavior. XGBoost correctly identified 1000 depression cases and 3397 non-cases, with 1172 false positives and 615 false negatives. Logistic Regression correctly identified 981 depression cases and 3447 non-cases, with 1122 false positives and 634 false negatives. Gradient Boosting correctly identified 966 depression cases and 3429 non-cases, with 1140 false positives and 649 false negatives. In contrast, Naïve Bayes identified more depression cases (TP = 1224) but produced the largest number of false positives (FP = 2220). Overall, these results indicate that XGBoost, Logistic Regression, and Gradient Boosting formed a comparable leading model group, with no clear practical separation in external validation performance based on the observed metrics. To statistically evaluate algorithm superiority, we used DeLong’s test to compare AUROC differences, paired bootstrap resampling to compare PR-AUC differences, and McNemar’s test to compare paired classification errors between models on the Oregon external validation dataset (Supplementary Table S5, S6 & S7). DeLong’s test showed that XGBoost had significantly higher AUROC than Decision Tree, Naïve Bayes, KNN, and SVM, but its AUROC was not statistically different from Gradient Boosting, Random Forest, or Logistic Regression. Paired bootstrap resampling for PR-AUC showed that XGBoost had significantly higher PR-AUC than Naïve Bayes, KNN, Decision Tree, Random Forest, and SVM, but was not statistically different from Gradient Boosting or Logistic Regression. McNemar’s test showed significant differences in paired classification errors between XGBoost and Naïve Bayes, Decision Tree, and Random Forest, but not between XGBoost and Logistic Regression, KNN, SVM, or Gradient Boosting. Therefore, the statistical tests indicate that XGBoost, Gradient Boosting, and Logistic Regression should be interpreted as a comparable leading group rather than evidence of clear superiority of a single algorithm. Given the class imbalance in self-reported depression, PR-AUC was interpreted alongside AUROC to provide additional insight into performance for identifying depression cases while accounting for false positives (Table 3; Fig. 4). Table 3Performance evaluation metrics and confusion matrix for ML models in self-reported depression diagnosis (oregon external validation dataset)MetricGBLRXGBSVMRFNBKNNDT Algorithms performance evaluation metrics Accuracy0.7110.7160.7110.7140.7210.5780.7150.694Precision0.4590.4660.4600.4610.4690.3550.4580.426Recall0.5980.6070.6190.5700.5380.7580.5070.525F1 Score0.5190.5280.5280.5080.5010.4840.4810.456Balanced Accuracy0.6740.6810.6810.6660.6620.6360.6480.628AUROC0.7490.7500.7510.7420.7470.7010.7280.689PR-AUC0.5250.5350.5280.5140.5140.4840.4740.425 Algorithms confusion matrix TP96698110009218691224819848FP114011221172108098222209691232FN649634615694746391796767TN34293447339734893587234936003337LR: Logistic Regression, NB: Naïve Bayes, SVM: Support Vector Machine, KNN: K-Nearest Neighbors, DT: Decision Tree, RF: Random Forest, GB: Gradient Boosting, XGB: Extreme Gradient Boosting (XGBoost), TP: True Positive, FP: False Positive, FN: False Negative, TN: True Negative, AUROC: Area Under the Receiver Operating Characteristic Curve, PR-AUC: Precision–Recall Area Under the Curve Performance evaluation metrics and confusion matrix for ML models in self-reported depression diagnosis (oregon external validation dataset) LR: Logistic Regression, NB: Naïve Bayes, SVM: Support Vector Machine, KNN: K-Nearest Neighbors, DT: Decision Tree, RF: Random Forest, GB: Gradient Boosting, XGB: Extreme Gradient Boosting (XGBoost), TP: True Positive, FP: False Positive, FN: False Negative, TN: True Negative, AUROC: Area Under the Receiver Operating Characteristic Curve, PR-AUC: Precision–Recall Area Under the Curve Model evaluation metrics for all models AUROC and PR-AUC curve analysis Receiver operating characteristic and precision–recall curve analyses were used to evaluate model discrimination and performance under class imbalance in the Oregon external validation dataset. The ROC curves showed that XGBoost, Logistic Regression, and Gradient Boosting had highly comparable discriminatory performance, with AUROC values of 0.751, 0.750, and 0.749, respectively. Random Forest and SVM also showed similar but slightly lower discrimination, with AUROC values of 0.747 and 0.742, respectively, while KNN (0.728), Naïve Bayes (0.701), and Decision Tree (0.689) demonstrated lower AUROC values. Overall, the AUROC curves suggest that the top-performing models were closely clustered, with no large practical separation between XGBoost, Logistic Regression, and Gradient Boosting. Because self-reported depression was less common than non-depression in the Oregon external validation dataset, PR-AUC was also examined to better assess model performance for the positive class. PR-AUC is particularly useful in imbalanced classification settings because it emphasizes the trade-off between precision and recall among respondents classified as depression cases. Logistic Regression achieved the highest PR-AUC (0.535), followed by XGBoost (0.528) and Gradient Boosting (0.525), while Random Forest and SVM each had PR-AUC values of 0.514. Naïve Bayes (0.484), KNN (0.474), and Decision Tree (0.425) showed lower precision–recall performance. Taken together, the AUROC and PR-AUC curves indicate that XGBoost, Logistic Regression, and Gradient Boosting formed a comparable leading group, rather than showing clear dominance of a single algorithm across all evaluation criteria. ROC curve for all models PR-AUC curve for all models Features importance analysis using shapley additive explanations (SHAP) SHapley Additive exPlanations (SHAP) were used to quantify and visualize the relative importance and directional effects of predictors contributing to the XGBoost model’s classification of self-reported depression. XGBoost was selected for SHAP-based interpretation because it provided a nonlinear, tree-based complement to the weighted logistic regression used in the epidemiologic modeling component of the study, not because it demonstrated clear statistical superiority over the other leading models. Aggregated SHAP feature importance showed that sex was the most influential predictor, followed by ACEs category, memory decline, disability burden, race/ethnicity, poor physical health days, age group, and number of children. Health and psychosocial burden variables, including ACEs, disability status, memory decline, poor physical health days, and comorbidity category, ranked prominently among the top predictors, suggesting that both life-course adversity and current health limitations contributed meaningfully to predicted depression risk. Other variables, including self-rated health, urban/rural status, alcohol use, cancer diagnosis, health insurance status, smoking status, caregiver status, home ownership, and BMI category, showed comparatively smaller contributions to the model’s predictions (Fig. 7). Aggregated SHAP feature importance bar plot for predicting self-reported depression The SHAP beeswarm plot further illustrated heterogeneity in predictor effects across individuals. Higher feature values for sex, corresponding to female respondents, were associated with positive SHAP values, indicating increased predicted probability of self-reported depression, while lower values were associated with negative SHAP values. Higher ACEs category and greater disability burden also generally shifted predictions toward higher depression probability, supporting the importance of life-course adversity and functional limitation in the model. Having memory decline showed an increased predicted probability of self-reported depression too. Poor physical health days and comorbidity category also tended to contribute positively to higher burden groups, while age group showed a pattern in which older age categories contributed lower predicted depression probability relative to younger groups. Race/ethnicity and number of children demonstrated more heterogeneous and category-specific effects, suggesting that their influence on predicted depression classification varied across subgroups rather than following a simple linear pattern (Fig. 8). SHAP feature importance beeswarm plot for predicting self-reported depression Case-level prediction using SHAP model This SHAP waterfall plot illustrates how individual predictor values contributed to the XGBoost model’s prediction for one participant in the test dataset. The baseline model output was close to zero, and the participant’s final predicted model output increased to f(x) = 0.935, indicating that the combined feature contributions shifted the prediction toward self-reported depression. Positive SHAP values, shown in red, increased the predicted probability of depression, while negative SHAP values, shown in blue, reduced the prediction. Features that increased the predicted probability of self-reported depression included poor physical health days, sex-related category indicators, self-rated health, disability category, cancer diagnosis, race/ethnicity, ACEs category, comorbidity category, number of children, and caregiver status. In contrast, older age group, memory decline category, ACEs category, and marital status contributed negative SHAP values, reducing the predicted probability (Fig. 9). Local SHAP waterfall plot explaining an individual prediction for self-reported depression Category-specific SHAP contributions of the top predictors This figure presents SHAP dependence plots for the top predictors contributing to the model’s prediction of self-reported depression in the test dataset. Each panel shows how categories of a given predictor influenced the model output, with positive grouped SHAP values indicating increased predicted probability of depression and negative values indicating decreased predicted probability. Female sex, higher ACEs category, memory decline, disability burden, and more frequent poor physical health days were associated with positive grouped SHAP values, indicating higher predicted probability of self-reported depression. In contrast, lower ACEs, absence of disability, no memory decline, and no poor physical health days generally contributed negative SHAP values, lowering the predicted probability. Race/ethnicity, age group, and number of children showed more heterogeneous patterns, with non-Hispanic White respondents contributing positively relative to other race/ethnicity categories, older age groups (especially 65 years and older) showing lower predicted depression probability, and higher numbers of children generally shifting predictions downward (Fig. 10). SHAP dependence plots for key predictors of self-reported depression Discussion This study represents the application of machine learning models to predict self-reported depression diagnosis outcomes within a population-based dataset, an area where traditional epidemiological methods often face limitations due to the complex, multifactorial, and non-linear nature of contributing risk factors. By integrating machine learning techniques with traditional epidemiological methods, this research utilizes a framework to explore population health patterns, especially in scenarios involving numerous interdependent variables [34]. The results highlight the practicality and effectiveness of explainable machine learning in examining risk factors of depression, particularly in cases where standard statistical approaches may fall short in capturing intricate, non-linear relationships [35]. Unlike traditional approaches that assume linear relationships, ML models provide a more flexible and scalable framework for uncovering intricate patterns in the data [36]. For instance, this study identified significant non-linear relationships between individuals’ age-group, adverse childhood experiences, disability, and comorbidity burden and depression risks. The capacity of ML to reveal such interactions aligns with emerging evidence advocating for the integration of artificial intelligence into public health research to enhance analytical power and predictive accuracy [37]. An important contribution of this study is the identification of key predictors associated with self-reported depression among adults in Tennessee. Using traditional epidemiological analyses, the weighted multivariable logistic regression model identified several factors independently associated with self-reported depression. Older age (≥ 35 years) was consistently protective, while female sex was associated with significantly higher odds of depression. Compared with White respondents, Black adults exhibited lower adjusted odds of reporting depression. Unmarried status, veteran status, poor self-reported wellbeing, and frequent poor physical health days were each associated with elevated depression risk. Household composition also demonstrated a protective pattern, with having five or more children in the household associated with reduced odds of depression. Behavioral and cognitive factors further contributed to risk stratification: memory decline, greater comorbidity burden, functional disability burden, and exposure to ACEs were associated with substantially increased odds of depression. Several predictors demonstrated clear dose–response trends in adjusted odds ratios, particularly age, comorbidity burden, ACEs exposure, and disability burden, indicating progressively stronger associations with depression as exposure levels increased. The largest effect sizes were observed for female sex, higher comorbidity burden, greater ACEs exposure, and increasing disability burden, underscoring the cumulative impact of health burden and early-life adversity on depression risk. These results are consistent with prior research highlighting the complex interplay between psychological, social, and physical health in depression [5, 13, 17, 38–40]. Numerous studies have shown that early life trauma, cognitive decline, and lack of social support are strong predictors of depression in adult populations [41, 42]. Our findings demonstrate how machine learning can complement traditional epidemiological approaches to inform the development of mental health programs [43]. The machine learning models developed achieved relatively high evaluation metrics. XGBoost, Gradient Boosting, and Logistic Regression demonstrated comparable external validation performance, with minimal differences across key metrics and no statistically significant superiority of one model over the others. To strengthen interpretation, SHAP was used to quantify the relative contribution and direction of influence of predictors in the XGBoost model, because it provided a nonlinear, tree-based complement to the weighted logistic regression used in the epidemiologic modeling component of the study, not because it demonstrated clear statistical superiority. Aggregated SHAP results identified sex, ACEs category, memory decline, disability burden, race/ethnicity, age group, poor physical health days, and number of children as the most influential predictors of self-reported depression, ranked by mean absolute SHAP values. Female sex, higher ACEs category, greater disability burden, poor physical health, and selected health-burden patterns generally shifted predictions toward a higher probability of self-reported depression. This method addresses the limitations of “black-box” models by generating interpretable results that can support data-driven policy and intervention planning [44]. National surveillance research supports these findings, individuals who reported frequent psychological distress, functional impairments, or histories of adverse experiences were significantly more likely to report depression [13, 39, 45]. Furthermore, being female, unmarried, and experiencing memory problems were linked to elevated depression risk, highlighting the roles of cognitive aging and social isolation [5, 13, 46]. This study demonstrates the complementary value of integrating traditional epidemiological methods with explainable machine learning to advance understanding of depression risk at both the population and individual levels. It also aligns with emerging literature demonstrating the strengths of ensemble-based machine learning models in predicting depression [13, 17, 18, 26, 47]. Weighted logistic regression provided interpretable, population-averaged effect estimates like odds ratios, allowing identification of statistically significant predictors and directionality of associations [48]. Additionally, its reliance on the assumption of linear relationships can oversimplify the multifaceted nature of depression risk [48]. However, depression is inherently multifactorial, with risk factors such as age, disability status, comorbidity burden, and adverse childhood experiences interacting in complex, non-linear ways that are not fully captured by parametric models. In contrast, the ensemble-based models achieved superior predictive performance by accommodating non-linear relationships and higher-order interactions among predictors [43]. Importantly, SHAP extended this advantage by translating model predictions into interpretable, observation-level contributions, revealing how specific features increased or decreased depression risk for individual respondents rather than only estimating average effects [27, 49]. While traditional regression quantifies whether a factor is associated with depression, SHAP values elucidate how and to what extent each variable contributes to a given prediction. Collectively, the integration of machine learning models and SHAP interpretation strengthens both predictive accuracy and interpretability, supporting their use in public health surveillance and targeted mental health policy development [27]. A major strength of this study is the integration of population-representative epidemiological analysis with explainable machine learning using a large, state-level surveillance dataset. By leveraging the complex survey design and weighting structure of the 2023 BRFSS, the findings are generalizable to the adult population of Tennessee. The use of multiple machine learning algorithms allowed for rigorous comparison across modeling approaches. Addressing class imbalance through SMOTEN applied exclusively to the training data strengthened model sensitivity without compromising external validity and minimizing the risk of data leakage. Importantly, the application of SHAP enhances transparency by providing both global and case-level insights into predictor contributions, bridging the gap between predictive accuracy and interpretability. This hybrid analytical framework improves the robustness, reproducibility, and policy relevance of the findings, offering a scalable approach for integrating advanced analytics into public health surveillance and mental health risk stratification. A key strength of this study is the use of an independent external validation dataset, with models trained on the Tennessee BRFSS dataset and evaluated on the Oregon BRFSS dataset. This approach provided a more rigorous assessment of model generalizability beyond the development population and strengthened confidence in the transportability of the predictive framework. Another strength of this study is the incorporation of composite indicators, such as adverse childhood experiences, functional disability, and comorbidity burden, which were largely absent from prior studies identified in our literature review and therefore represent a valuable extension to existing depression prediction research. Future research can expand on this work by integrating longitudinal data, incorporating electronic health records or social determinants from other sources, and evaluating the predictive stability of machine learning models over time. It should also aim to test the generalizability of these findings across different states or populations and incorporate mixed-methods designs to gain a deeper understanding of individual trajectories and contextual influences on depression. Additionally, exploring more sophisticated algorithms such as deep neural networks or ensemble meta-models may further enhance performance. Limitations Although this study offers valuable insights, it has limitations that should be acknowledged. First, the use of a cross-sectional design limits the ability to draw causal inferences between the identified predictors and the self-reported depression outcome. As a result, while associations are identified, temporal relationships cannot be confirmed. Second, the study relied on self-reported data from the BRFSS, which is susceptible to reporting biases such as recall error and social desirability bias, particularly concerning sensitive topics like mental health and childhood adversity. Third, the analysis was limited to variables available in the BRFSS dataset, potentially omitting important psychosocial and environmental factors such as stigma, access to mental health services, or community-level stressors. Fourth, while SHAP facilitated model interpretability by decomposing individual predictions into feature contributions, the inherent complexity of machine learning models continues to limit the clarity and straightforward interpretability of these explanations. Conclusion This study demonstrated the integration of traditional epidemiological methods and machine learning algorithms in predicting self-reported depression among adults in Tennessee using 2023 BRFSS data, thereby enhancing mental health surveillance and support data-driven public health interventions. By identifying key sociodemographic and health-related predictors, including sex, disability burden, ACEs, marital status, memory decline, and comorbidity burden, the findings provide actionable insights to inform population-level mental health strategies. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Author contributions MAM conceptualized the study, developed the methodology, and performed the formal analysis. JS & SH assisted with data curation, implemented the software, and supported model evaluation. All authors contributed to the manuscript writing, reviewed the final draft, and approved the submitted version. Funding This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Data availability The dataset analyzed during the current study is publicly available through the Centers for Disease Control and Prevention (CDC) Behavioral Risk Factor Surveillance System (BRFSS) 2023 annual data release: [https://www.cdc.gov/brfss/annual\_data/annual\_2023.html](https:/www.cdc.gov/brfss/annual_data/annual_2023.html). Code availability All scripts used to conduct the epidemiological analyses and machine learning modeling are available in a public GitHub repository at: https://github.com/Mustapha0287/Explainable-ML-Self-Reported-Depression-Tennessee-Adults-Evidence-from-the-2023-BRFSS/tree/main. Declarations Competing interests The authors declare no competing interests. Consent to participate This study used de-identified, publicly available BRFSS data; informed consent was obtained from participants by the Centers for Disease Control and Prevention at the time of data collection. Consent to publish Not applicable. Clinical trial number Not applicable. Ethics approval This study utilized a fully anonymized, publicly available dataset obtained from the Behavioral Risk Factor Surveillance System (BRFSS) 2023. Since the data were de-identified prior to analysis and no human subjects were directly involved in this secondary data analysis, ethical approval was not required. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Data Availability Statement The dataset analyzed during the current study is publicly available through the Centers for Disease Control and Prevention (CDC) Behavioral Risk Factor Surveillance System (BRFSS) 2023 annual data release: [https://www.cdc.gov/brfss/annual\_data/annual\_2023.html](https:/www.cdc.gov/brfss/annual_data/annual_2023.html). All scripts used to conduct the epidemiological analyses and machine learning modeling are available in a public GitHub repository at: https://github.com/Mustapha0287/Explainable-ML-Self-Reported-Depression-Tennessee-Adults-Evidence-from-the-2023-BRFSS/tree/main.