Prediction of depressive disorder using machine learning approaches: findings from the NHANES

Abstract

Background

Depressive disorder, particularly major depressive disorder (MDD), significantly impact individuals and society. Traditional analysis methods often suffer from subjectivity and may not capture complex, non-linear relationships between risk factors. Machine learning (ML) offers a data-driven approach to predict and diagnose depression more accurately by analyzing large and complex datasets.

Methods

This study utilized data from the National Health and Nutrition Examination Survey (NHANES) 2013–2014 to predict depression using six supervised ML models: Logistic Regression, Random Forest, Naive Bayes, Support Vector Machine (SVM), Extreme Gradient Boost (XGBoost), and Light Gradient Boosting Machine (LightGBM). Depression was assessed using the Patient Health Questionnaire (PHQ-9), with a score of 10 or higher indicating moderate to severe depression. The dataset was split into training and testing sets (80% and 20%, respectively), and model performance was evaluated using accuracy, sensitivity, specificity, precision, AUC, and F1 score. SHAP (SHapley Additive exPlanations) values were used to identify the critical risk factors and interpret the contributions of each feature to the prediction.

Results

XGBoost was identified as the best-performing model, achieving the highest accuracy, sensitivity, specificity, precision, AUC, and F1 score. SHAP analysis highlighted the most significant predictors of depression: the ratio family income to poverty (PIR), sex, hypertension, serum cotinine and hydroxycotine, BMI, education level, glucose levels, age, marital status, and renal function (eGFR).

Conclusion

We developed ML models to predict depression and utilized SHAP for interpretation. This approach identifies key factors associated with depression, encompassing socioeconomic, demographic, and health-related aspects.

Introduction

Depressive disorder, specifically major depressive disorder (MDD), place a significant burden on individuals and society. These psychiatric disorders are distinguishable by pervasive feelings of sadness, emptiness, or hopelessness that significantly interfere with a person’s daily activities. These disorders are highly prevalent worldwide and have a significant impact on individuals’ ability to do daily tasks, their overall well-being, their cognitive function, and employment status [1]. Depression can cause significant limitations in personal, social, and occupational domains, making it an important public health concern that requires effective methods for prediction and intervention [1, 2]. Depression is the primary cause of impairment, as determined by Years Lived with Disability (YLDs), and the fourth most significant contributor in the worldwide burden of disease, according to the World Health Organization (WHO) [3]. The Global Burden of Disease (GBD) study shows that depressive disorders account for a substantial portion of total Disability Adjusted Life Years (DALYs) and YLDs, with a trend of increasing burden over time [4]. An increase in the global economic burden of mental health diseases is anticipated, with a particular emphasis on the burden of depressive disorders. This highlights the urgent need for effective strategies to address the rising burden of depressive disorders and improve access to mental health care services.

Traditional statistical methods have long been employed in the prediction and analysis of depression. These methods typically involve hypothesis-driven approaches that use predefined models to understand the relationships between various risk factors and depression outcomes. Despite being helpful, they may not capture the complex, non-linear relationships between different risk factors and depression, limiting their effectiveness in fully understanding and predicting the disorder.

Recent studies have shown that machine learning (ML) has become a revolutionary tool for predicting and diagnosing diseases [5,6,7]. It offers several advantages over traditional statistical methods [6]. Unlike conventional approaches that test hypotheses derived from theories, ML focuses on discovering hidden patterns and interactions within large datasets [6]. This capability enables ML to analyze complex, non-linear relationships among variables, leading to more accurate and nuanced predictions of depression risk [8].

One of the major challenges in adequately addressing MDD is identifying affected individuals and ensuring appropriate and timely treatment. MDD symptoms are internally experienced, and often go undetected. The application of ML in epidemiological studies and public health has revolutionized the approach to depression prediction and early intervention [9, 10]. ML algorithms can process vast amounts of data from electronic health records (EHRs), biometric markers, and patient characteristics to identify individuals at risk of developing depression [9, 11]. This ability to analyze complex, multidimensional data sets with greater precision than traditional methods make ML particularly valuable in the context of population-level health studies.

When considering the broader context of an individual or population, depression is influenced by a multitude of factors. The critical question arises: which of these identified risk factors are the most significant, and how do they contribute to the formation of the predicted outcome?

Through this study, we aim to assess the role of machine learning (ML) in epidemiological research. Additionally, we strive to explore all potential risk factors for depressive disorders (DD) based on large-scale National Health and Nutrition Examination Survey (NHANES) data from the year 2013–2014 using supervised ML. Furthermore, we aim to evaluate the contribution of each risk factor to the development of DD.

Methods

Study participants

The study employed data from NHANES 2013–2014, which is a dependable and extensive random sample designed to evaluate the health and nutritional condition of the US population (www.cdc.gov/nchs/nhanes). Participants were questioned in their residences, and had physical and laboratory examinations at a mobile examination center (MEC). The National Centre for Health Statistics Research Ethics Review Board granted ethics approval (Protocol # 2013-14) for all the study procedures, and all subjects provided signed informed consent. The studies conducted adhered to NHANES guidelines and regulations. The study specifically eliminated those who were under the age of 18 and those who provided incomplete or insufficient responses.

Data collection

Data for the study were collected from various sources within NHANES 2013–2014, including: Demographic Data, Examination Data, Laboratory Data, and Questionnaire Data. The following variables were employed:

• Age was calculated from the date of the interview to the date of birth.

• Gender was coded as male or female by NHANES personnel.

• Race was assessed by two questions: “Do you identify as Hispanic, Latino, or Hispanic?” and “What race do you consider yourself?” Based on the responses, race was categorized into six groups: Mexican American, Other Hispanic, Non-Hispanic White, Non-Hispanic Black, Non-Hispanic Asian, and Other Race.

• Education level was grouped by the question: “What is the highest grade or level of school you have completed or the highest degree you have received?” Responses were divided into five groups: less than grade 9th, 9–11th grade, high school graduate, some college or associate degree, and college graduate or above.

• Marital status was categorized as married, widowed, divorced, separated, never married, and living with a partner.

• PIR refers to the ratio of family income to the family’s appropriate poverty threshold. Annual family income was categorized into four levels: low (income < $20,000), low-medium ($20,000 ≤ income < $75,000), medium-high ($75,000 ≤ income < $99,000), and high income (income > $100,000).

• Body mass index (BMI) was calculated as weight in kilograms divided by height in meters squared (kg/m²).

• Blood pressure was measured using a mercury sphygmomanometer, with two consecutive readings of systolic blood pressure (SBP) and diastolic blood pressure (DBP) taken at 5-minute intervals. The mean of the two readings was calculated for analysis.

• The estimated glomerular filtration rate (eGFR) was calculated using the Modification of Diet in Renal Disease formula: eGFR = 186 * creatinine^(-1.154) * age^(-0.203) * (0.742 if female).

• Both smoking and drinking habits were separated into three categories: current, past, and never.

• Physical activity was categorized as mild, moderate, or vigorous.

• Hypertension was defined as a resting SBP ≥ 140 mmHg, a DBP ≥ 90 mmHg, or the use of hypertensive medication.

• Diabetes was defined as a two-hour glucose tolerance test result ≥ 200 mg/dL, plasma fasting glucose ≥ 126 mg/dL, glycohemoglobin ≥ 6.5%, or the use of diabetic medication.

• Dyslipidemia was defined as total cholesterol ≥ 200 mg/dL, triglycerides ≥ 150 mg/dL, low-density lipoprotein cholesterol ≥ 140 mg/dL, high-density lipoprotein cholesterol < 40 mg/dL, or the use of lipid-lowering medication.

Depressive disorder assessment

Depression was assessed using the Patient Health Questionnaire (PHQ-9), a nine-item depression scale that is widely used to screen, diagnose, monitor, and measure the severity of depression. The PHQ-9 screening instrument consists of nine questions about depression symptoms experienced during the past 2 weeks followed by a single question that assesses associated impairment. Each question is scored on a scale from 0 to 3, with 0 indicating the absence of the symptom, 1 indicating its presence for several days, 2 indicating its presence for more than half of the days and 3 indicating its presence nearly every day. The result of the PHQ-9 questionnaire consists of scores ranging from 0 to 27. The PHQ-9 is widely used in various areas, such as psychiatric hospitals, primary care, and the general population, and has demonstrated good reliability and validity in assessing depression severity [12, 13]. A score equal to or exceeding 10 is frequently employed as a benchmark for indicating moderate to severe depression. Studies have discovered that a PHQ-9 score of 10 or higher is both extremely sensitive and specific in identifying severe depression [14]. In this study, we employed a PHQ-9 score of 10 as the threshold for diagnosing individuals with depressive disorder.

Data pre-processing

We conducted a comprehensive analysis of the NHANES dataset to predict depression status using various statistical and machine learning techniques. Initially, we filtered and cleaned the data, creating a cumulative depression score and categorizing individuals based on their depression levels. We also derived variables for smoking, drinking, physical activity, family income, blood pressure, hypertension, dyslipidemia, diabetes, and renal function.

The data was described using the mean and standard deviation for symmetric numerical variables, and the median and interquartile range for asymmetric numerical variables. Categorical variables were described using frequency and percentage. To analyse differences in participants characteristics between those with and without depression, we utilized a range of statistical tests, including Student’s t-tests, Mann-Whitney U tests, and Chi-squared tests, depending on the suitability for each variable.

Features selection

The selection of features (variables) was conducted in a systematic manner to retain only the most relevant predictors for the model development. Initially, variables with over 50% missing data were excluded to minimize potential bias caused by imputation. We used Random Forest-based imputation for handling missing data [15]. Following this, a correlation matrix was employed to identify and remove variables exhibiting high multicollinearity, defined as having a correlation coefficient exceeding 0.8. Subsequently, Least Absolute Shrinkage and Selection Operator (LASSO) regression was used to shrink the coefficients of less influential predictors towards zero, effectively eliminating them from the model. The final set of features used for model development is presented in Supplementary Table S2.

Model development

The significant imbalance could lead to the model being biased toward the majority class (without depression), which would reduce its effectiveness in predicting depression cases. To mitigate this issue, we reduced the number of samples from the majority class (without depression) to match the number of samples in the minority class (depression). By using Random Undersampling, we ensured that both classes were equally represented during model training. This approach was chosen for its simplicity and effectiveness in improving model performance when dealing with imbalanced data. Next, we applied one-hot encoding to convert categorical variables into a binary format and normalization to scale the numerical features to prepare the data for the machine learning models. In this study, the Z-score normalization method was used for data normalization. Specifically, this technique centers each numerical feature by subtracting the mean and then scales it by dividing by the standard deviation [5, 6]. The dataset was split into training and testing sets with an 80% and 20% ratio, respectively. We trained multiple models, including LR, RF, NB, SVM, XGBoost, and LightGBM. The selection of LR, RF, NB, SVM, XGBoost, and LightGBM is based on several key considerations, including diversity, proven effectiveness, handling of different data types, and computational efficiency. We choose a diverse set of algorithms to capture both linear and non-linear relationships in the data. Each model brings unique strengths: LR is a staple for risk prediction due to its simplicity and interpretability [16], Naive Bayes efficiently handles categorical features [17], SVM is suitable for high-dimensional data [18, 19], RF and gradient boosting models capture non-linear relationships and complex feature interactions [20, 21]. XGBoost, a gradient boosting method, is known for its high efficiency and accuracy, making it a powerful tool for both classification and regression tasks in the context of medical research [22]. Additionally, LightGBM’s computational efficiency makes it ideal for large-scale datasets [23]. This diverse set of models was chosen to leverage their complementary strengths, providing a comprehensive evaluation across different algorithm types. During training, we used 5-fold cross-validation to ensure robustness. After training, we tested the models on the test set and calculated performance metrics such as accuracy, sensitivity, specificity, precision, area under the receiver operating characteristic curve (AUC) and F1-score.

The hyperparameter optimization process involved the systematic tuning of key parameters for each machine learning model. The study utilized grid search-based hyperparameter tuning in combination with 5-fold cross-validation to evaluate different combinations of parameters and ensure stability and minimizing the impact of random variations. The objective was to identify the optimal configuration that maximized model performance, specifically focusing on metrics such as accuracy and AUC. The best model for predicting depression was identified by comparing performance metrics on the test set.

Model explanation

SHAP scores are used to explain the output of our depression prediction model. SHAP values provide a common measure for interpreting how each feature in a model contributes to prediction.

We analysed feature importance using SHAP values. This thorough approach allowed us to pinpoint key factors influencing depression and evaluate how these factors contribute for predicted outcome, ultimately enhancing our understanding of the dataset and the underlying relationships within the data. The analysis was conducted using the statistical software R (version 4.4.0).

Results

Participants’ characteristics

The study evaluated the characteristics of 5,372 participants, divided into those without depression (4,861, 90.5%) and those with depression (511, 9.5%). The results show significant differences between the two groups on a range of demographic, socio-economic and health-related factors.

Participants with depression were older with high proportion of females and low level of education. However, non-Hispanic Asians were the more prevalent among those without depression.

There were fewer married individuals and more divorced individuals in the depression group. Similarly, current smokers were more prevalent among those with depression. Physical activity levels were lower, and very-low family income was more common in the depression group.

Health measures showed that the depression group had higher systolic blood pressure and BMI. Biochemical differences included lower eGFR, higher apolipoprotein B, lower HDL-cholesterol, and higher triglyceride, fasting glucose, and two-hour glucose levels.

Moreover, serum and urine cotinine and hydroxycotinine levels were significantly higher among participants with depression. Furthermore, participants with depression tended to have higher prevalence of hypertension, diabetes and dyslipidemia.

These findings indicate significant differences in demographics, socioeconomic status, lifestyle factors, biochemical parameters, and health conditions between participants with and without depression, highlighting the multifaceted nature of depression and its association with various risk factors, illustrated in Table 1.

Table 1 Characteristics of study participants with and without depression

The performance metrics of the six machine learning models applied in prediction of depression give the overview of their effectiveness, as shown in Table 2. XGBoost generally had the highest metrics across all categories, indicating strong overall performance. It showed the best accuracy (0.69), good sensitivity (0.68), and high specificity (0.71), meaning it balanced well between true positive rate and true negative rate. Additionally, it had the highest AUC (0.69), indicating good discriminatory power, and its F1-score was also the highest (0.69), suggesting it maintained a good balance between precision and recall. Naive Bayes also performed well, with similar accuracy (0.68), sensitivity (0.70), and AUC (0.68). However, it slightly lagged behind XGBoost in specificity (0.67) and F1-score (0.69). SVM and Logistic Regression were close contenders but generally fell short in some areas compared to XGBoost and Naive Bayes. Random Forest and LightGBM did not perform as well as the other models in most metrics. XGBoost appeared to be the best-performing model based on the provided metrics. It offered the highest accuracy, strong sensitivity, and specificity, along with the highest AUC and F1-score. Therefore, this indicated that XGBoost was likely the most reliable model for this classification task, offering a good balance between all evaluated metrics. See Supplementary Table S1 for more information about model performance metrics and Supplementary Figure S1 for ROC curves of all machine learning models.

Table 2 Performance metrics of different machine learning approaches

Figures 1 and 2 both highlight the importance and impact of different features on the model’s predictions using SHAP values. These visualizations, including a SHAP feature importance plot and a SHAP summary heat plot, offer critical insights into the model’s behavior and the factors influencing its prediction. See Supplementary Table S2 for more information about variables’ names.

Fig. 1

The contribution levels of all variables to depression based on SHAP values The global significance of each feature in the model is illustrated in the SHAP (blue) bar plot. It provides an overview of the features’ impact on the model’s output by displaying the mean absolute SHAP value for each feature. A feature (variable) is represented by each bar in the plot, and the length of the bar indicates the extent of the feature’s contribution to Depression

Fig. 2

The heat plot on SHAP values The relationships between the feature (variable) and Depression are revealed by the heat plot of SHAP values. The relationship between the value of a specific feature and its impact on prediction can be fundamentally understood through this. Each data point is associated with a specific participant and their corresponding Shapley value for a specific feature. The Shapley value, which is represented on the x-axis, and the feature’s prominence, which is represented on the y-axis, determine the position of a data point on this plot

The SHAP feature importance plot ranks the features based on their average absolute SHAP values, highlighting those with the most significant impact on the model’s prediction. The top features identified are PIR, which is the most important, indicating its highest impact on the model’s predictions. This is followed by sex.2 (female) and hypertension, highlighting their significant roles. sHCOT (Serum Hydroxycotinine), sCOT (Serum Cotinine) and BMI also show substantial importance, indicating their strong influence on the model. Additionally, educ.5 (Education Level: College graduate or above) significantly impacts the model’s prediction. Features such as glucose levels, age, marital status (divorced status) have moderate importance, meaning they still contribute meaningfully but to a lesser extent than the top features. Other features have the smaller impact on the model’s predictions.

The SHAP summary plot provides insights into how various features influence the model’s predictions. It uses colors to indicate feature values, with yellow representing low values and blue representing high values. The SHAP values on the horizontal axis indicate the impact of each feature on the model’s output. Negative SHAP values suggest a protective effect against depression, while positive SHAP values indicate a higher likelihood of depression.

One of the most important features is the PIR. Lower PIR values (yellow dots) generally push the model’s predictions higher, indicating that lower family income increases the likelihood of depression. Conversely, higher PIR values (blue dots) decrease this likelihood, acting as a protective factor, with the wide spread of SHAP values showing PIR’s strong and consistent impact on the model. Gender is another significant categorical variable. Females have a higher chance of depression, as the positive SHAP values suggest an increased likelihood of depression compared to the male group.

Similarly, hypertension push the model’s predictions higher, indicating a higher likelihood of depression, while not having hypertension is more protective.

Marital status (masts.3), particularly being divorced, shows that divorced individuals have a higher likelihood of depression, as indicated by the positive SHAP values, while not being in this group shows a lower likelihood.

Education level (educ.5), categorized as college graduate or above, acts as a protective factor. Higher education levels have negative SHAP values, indicating a lower likelihood of depression, while lower education levels increase the risk.

However, BMI shows varied impacts. For serum cotinine levels (sHCOT), serum cotinine (sCOT), and glucose levels (Glu), higher values increase the likelihood of depression, as shown by positive SHAP values, while lower values are protective.

Age is another crucial numerical variable that had a varied impact. Lower eGFR (Estimated Glomerular Filtration Rate) values were harmful, increasing the likelihood of depression, while higher eGFR values were protective. Figure 2 showed heat plot of all variables to depression, however, it doesn’t clearly focus on each variable. Figures 3, and 4 provide more information of important variables for both categorical and numerical data.

Fig. 3

The impact of categorical variables on depression

Fig. 4

The impact of numerical variables on depression

These insights help in understanding the model’s behavior and identifying the most important factors influencing the prediction.

Discussion

In this study, we applied machine learning (ML) approaches to predict depression using big data from the NHANES cycle 2013–2014. We used the PHQ-9 score with a cut-off point of 10 to dichotomize depressive disorder. The input factors were derived from demographic data, examination data, laboratory data, and questionnaire data of the study participants. We employed six supervised models: Logistic Regression, Random Forest, Naïve Bayes, XGBoost, and LightGBM to predict depression. AUC and F1-score were used as critical indicators for evaluating model performance, with XGBoost emerging as the best-performing model. It offered the highest accuracy, sensitivity, specificity, along with the highest AUC and F1- score, indicating its reliability for this classification task.

To explain the model, we applied SHAP values to identify the most important variables contributing to the risk of depression. The top variables included socioeconomic factors like PIR, education level, marital status, demographic factors such as age, sex, and health-related factors such as hypertension, BMI, blood glucose, eGFR and consumption of nicotine products (serum cotinine & serum hydroxycotinine).

Our findings highlight the superiority of ML models in leveraging all input data to build predictive models. Traditional analysis typically confirms one or a few predictors with the predicted outcome, but our approach demonstrates the advantage of using data-driven ML techniques to gain a comprehensive view of risk factors and their contributions to depression.

The most significant risk factor identified was PIR. Lower PIRs, indicating higher poverty, were associated with significantly higher rates of depressive symptoms. This is consistent with previous studies examining the nonlinear associations between PIR and various health outcomes, which found that lower PIRs were linked to higher vulnerability to adverse health outcomes, including mental health issues like depression [24, 25]. These findings emphasize the importance of considering income levels in public health strategies [24, 25].

We found that women were more likely to have depression compare to men. This aligns with numerous studies showing that women are more likely to suffer from depression than men [26, 27]. Hormonal changes related to the menstrual cycle, pregnancy, and menopause, along with chronic stressors and social discrimination, contribute to this higher prevalence [28].

Both hypertension and depression are linked to increased sympathetic nervous system activity and decreased parasympathetic activity, leading to elevated blood pressure and increased risk of cardiac arrhythmias [29]. Depression is associated with unhealthy behaviors such as smoking, physical inactivity, increased alcohol consumption, poor nutrition, and poor sleep, which are known risk factors for hypertension. The heightened sympathetic tone in depressed individuals may contribute to poor blood pressure control and exacerbate hypertension [30, 31]. Elevated cortisol levels due to depression can also promote vascular changes leading to sustained high blood pressure [32].

BMI and depression exhibit intricate relationships, as indicated by our findings in Fig. 4B. Individuals with a BMI of less than 20 exhibit a high SHAP value. Nevertheless, participants with a BMI of 20–30 have negative SHAP values, which indicates that they are less likely to experience depression. Conversely, individuals with a BMI greater than 30 exhibit a significantly elevated SHAP value, which is associated with a favorable correlation with depression. Malnutrition and health issues associated with low body weight can impact mood and mental health [33]. Obesity is often associated with lifestyle factors such as poor diet, physical inactivity, and sleep disturbances, all of which are risk factors for depression [33, 34]. Conversely, depression can lead to changes in appetite and physical activity, contributing to weight gain and obesity. Therefore, personalized intervention and treatment strategies tailored to specific BMI levels are necessary for optimal outcomes [35,36,37].

Individuals with lower education levels tend to have higher rates of depression compared to those with higher educational attainment. Education plays a protective role against depression through various socioeconomic pathways. Higher education levels are associated with better economic and social resources, which help individuals manage and mitigate depressive episodes. Education also influences socioeconomic position, leading to better employment opportunities, higher income, and greater social status, all contributing to lower depression rates [38, 39].

The relationship between glucose levels and depression involves multiple biological, psychological, and lifestyle factors. Insulin resistance and high blood sugar levels can stimulate the release of stress hormones like cortisol, linked to depression. Both high blood sugar levels and depression are associated with chronic inflammation, with inflammatory markers such as C-reactive protein (CRP) and interleukin-6 (IL-6) often elevated in individuals with both conditions [40, 41].

The complex relationship between depression and age was discovered, as illustrated in Fig. 4E. Depression was low among participants under the age of 50, as evidenced by their negative SHAP value. The SHAP value was positive for participants aged 50–75, with the highest value observed in the group of participants aged 60 (55–65 years) who were experiencing elevated levels of depression. The SHAP value becomes negative when the age exceeds 75 years.

Marital status is significantly associated with the prevalence of major depression. Married individuals generally report lower rates of depression compared to those who are single, divorced, or widowed. Married individuals often report higher levels of subjective well-being, which is associated with lower depression rates [42,43,44].

Poor kidney function, indicated by lower eGFR, is associated with higher rates of depression. Individuals with chronic kidney disease (CKD) are more likely to experience depressive symptoms compared to those with normal kidney function. The prevalence of depression increases as kidney function declines, with the highest rates observed in patients with end-stage renal disease (ESRD) undergoing dialysis [45, 46].

Cotinine and hydroxycotinine, biomarkers for nicotine exposure, are linked to depression through neuroinflammation, nicotine dependence, and oxidative stress. Understanding these relationships can help in developing targeted interventions for individuals with depression who are also exposed to nicotine [47,48,49,50].

Strength and limitation

In summary, the study utilized data from NHANES 2013–2014, which is a robust and comprehensive dataset including a wide range of demographic, health, and laboratory information. This broad dataset enabled a detailed and multifaceted analysis of depression predictors. Additionally, the use of SHAP values for model interpretation was a significant strength, providing clear and understandable explanations of how each feature contributes to the model’s predictions. This helped identify the most important variables influencing depression risk. Furthermore, the study successfully identified key risk factors for depression, such as PIR, education level, marital status, age, sex, hypertension, BMI, blood glucose, eGFR and nicotine products. This comprehensive identification can provide the basis for targeted public health interventions.

However, this study has several limitations that should be acknowledged. First, as it is based on cross-sectional data, the findings cannot establish causal relationships between the identified risk factors and depression, which limits the ability to draw conclusions about the temporal dynamics of these associations. Longitudinal studies are necessary to confirm these relationships and understand their progression over time. Second, some variables, particularly those related to socioeconomic status, physical activity, smoking, and drinking habits, were self-reported, introducing the potential for recall bias and inaccuracies. Third, the use of imputation to handle missing data, while effective, may not fully capture the true values, potentially leading to residual bias. Additionally, the exclusion of variables with over 50% missing data might have resulted in the loss of important information. Finally, addressing class imbalance through random undersampling, while effective in balancing the dataset, may limit the generalizability of the findings to broader populations where the prevalence of depression is lower. Furthermore, undersampling may introduce bias by underrepresenting specific subgroups or feature combinations within the control group, which could affect the robustness and interpretability of the model.

Conclusion

Our study demonstrates the advantages of leveraging ML models to predict depression by using comprehensive datasets. By identifying key risk factors, these models provide valuable insights into the multifaceted nature of depression and emphasize the significance of considering socioeconomic, demographic, and health-related factors in understanding and addressing this complex condition.