Interpretable Machine Learning for Predicting Early Mental Health Care-Seeking Among Reproductive-Age Women in Bangladesh Using BDHS 2022 Data

Interpretable Machine Learning for Predicting Early Mental Health Care-Seeking Among Reproductive-Age Women in Bangladesh Using BDHS 2022 Data | Research Square Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Interpretable Machine Learning for Predicting Early Mental Health Care-Seeking Among Reproductive-Age Women in Bangladesh Using BDHS 2022 Data Masudur Rahman Kanchon, Jamilu Sani, Tanvir Ahmed, Md. Rakib Akon, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8595064/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 23 Jun, 2026 Read the published version in Discover Public Health → Version 1 posted 13 You are reading this latest preprint version Abstract Machine learning (ML) holds promise for predicting complex health behaviors such as mental health care-seeking, yet its use and interpretation in low-resource settings remain underexplored. This study utilized the 2022 Bangladesh Demographic and Health Survey to predict early anxiety and depression care-seeking among reproductive-age women in Bangladesh using multiple ML algorithms. We analyzed data from 4,255 ever-married women aged 15–49 with anxiety (GAD-7 ≥ 6) or depression (PHQ-9 ≥ 10). Nine ML algorithms; Logistic Regression, Decision Tree, Random Forest, K-Nearest Neighbors (KNN), Gradient Boosting, XGBoost, LightGBM, CatBoost, and AdaBoost were evaluated using accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUROC). The Synthetic Minority Oversampling Technique (SMOTE) was applied to address class imbalance, and SHapley Additive exPlanations (SHAP) were used to interpret the best-performing model. The Random Forest model achieved the best predictive performance (accuracy = 0.66, recall = 0.69, F1-score = 0.67, AUROC = 0.70), followed by CatBoost and Gradient Boosting (F1 ≈ 0.66). SHAP analysis identified administrative division, age group, number of children, household size, and mass media exposure as the most influential features. Geographic disparities and information access were found to have the highest SHAP values, highlighting their strong role in predicting care-seeking behavior. This study demonstrates that explainable ML, combined with SHAP interpretability, provides a reliable, transparent framework for predicting mental health care-seeking behavior in low-resource settings. These insights can inform equitable, evidence-based interventions to target at-risk populations in Bangladesh and similar contexts. Machine Learning SHAP Random Forest Anxiety Depression Bangladesh Women’s Mental Health Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 1. INTRODUCTION One in seven persons globally have a mental health issue, contributing to the global disease burden [ 1 ]. Anxiety and depression are the most prevalent disorders [ 2 ], responsible for nearly 50% of this burden and disproportionately affecting women of reproductive age in low- and middle-income countries (LMICs) [ 3 , 4 ]. Many developing countries have low mental health help-seeking due to stigma, awareness, service availability, and sociocultural barriers, despite a significant prevalence [ 5 – 8 ]. Early care-seeking for mental distress is critical, as timely diagnosis and intervention improve recovery outcomes, prevent chronicity, and reduce social and economic consequences for affected women and their families. In Bangladesh, mental health is increasingly recognized as an emerging public health challenge. National surveys have reported that around 17% of adults experience some type of mental disorder, with higher prevalence among women and younger adults [ 9 ]. However, only a small proportion of affected individuals access formal treatment, reflecting a considerable treatment gap. Structural barriers such as the scarcity of trained professionals, urban concentration of services, and limited integration of mental health into primary care contribute to this gap [ 10 , 11 ]. Additionally, societal stigma, gender norms, and lack of decision-making autonomy hinder women from seeking psychological distress care [ 8 , 12 , 13 ]. Understanding the determinants of early care-seeking behavior among women experiencing anxiety and depression is therefore essential for designing effective and equitable mental health programs. Previous studies in Bangladesh and other LMICs have examined factors influencing mental health service utilization, often using traditional regression techniques [ 4 , 13 – 17 ]. While these studies have identified important associations such as the role of age, education, income, and social support, linear models may not fully capture the complex, nonlinear relationships between socioeconomic, demographic, and contextual determinants of mental health behavior. Moreover, such approaches provide limited predictive capability and lack interpretability when dealing with large, multidimensional datasets like the Demographic and Health Surveys (DHS). For more accurate health behavior prediction and risk factor identification, public health research has increasingly used machine learning (ML) methodologies [ 18 , 19 ]. Machine learning models, especially ensemble techniques such as Random Forest (RF), Gradient Boosting (GBoost), and XGBoost, have shown superior performance in health behavior prediction tasks due to their ability to manage nonlinearities and high-dimensional data. However, their “black-box” nature has raised concerns about interpretability and practical utility for policy decisions. Explainable artificial intelligence (XAI) methods like SHapley Additive exPlanations (SHAP) quantify each feature's model prediction contribution, enhancing ML-based insights' transparency and trust[ 19 ]. Incorporating SHAP analysis allows researchers to not only achieve accurate predictions but also understand which factors most strongly influence outcomes, making such approaches valuable for guiding evidence-based health interventions. Against this background, using 2022 Bangladesh Demographic and Health Survey data, this study uses numerous machine learning algorithms to predict early anxiety and depression management for reproductive-age women in Bangladesh. Furthermore, the study employs SHAP-based model interpretability to identify the most influential determinants of care-seeking behavior. By combining the predictive power of ML with explainable analytics, this research seeks to provide actionable insights into the socioeconomic, demographic, and contextual factors that influence women’s mental health care utilization. The findings are expected to inform targeted policy and programmatic strategies to promote early detection and equitable access to mental health services across Bangladesh. 2. METHODS 2.1. Data source This cross-sectional analytical analysis used secondary data from the 9th round Bangladesh Demographic and Health Survey (BDHS) of 2022 (NIPORT & ICF, 2024). The National Institute of Population Research and Training (NIPORT) and the Health Education and Family Welfare Division of the Ministry of Health and Family Welfare (MoH&FW), Government of Bangladesh, conducted this study. Mitra and Associates implemented the survey with ICF International's technical help [ 20 ]. The survey was conducted in four phases, each lasting approximately 4 weeks, from June to December 2022. 2.2. Sampling design and sample size A two-stage stratified random sample procedure was used to interview 30,078 ever-married women aged 15–49 for the nationally representative BDHS 2022 survey (99% response rate). For this study, the Individual Recode (IR) dataset was used, with a final analytical sample of 4,255 women (weighted) who had either anxiety (GAD-7 ≥ 6) or depression (PHQ-9 ≥ 10). This entire sample selection process is detailed in Fig. 1 . The BDHS-2022 final report covers sampling, data collection, quality control, and the survey questionnaire [ 20 ]. 2.3. Study variables 2.3.1 Outcome variable : The primary outcome variable in this study was early care-seeking for symptoms of anxiety or depression. Help-seeking was measured as a binary variable created from the question, “Have you ever tried to seek help for the things you experienced?” , which followed the GAD-7 and PHQ-9 questionnaires among women with anxiety or depression symptoms. Anxiety and depression were assessed using the GAD-7 (score range: 0–21; ≥6 indicating moderate to severe anxiety) and the PHQ-9 (score range: 0–27; ≥10 indicating moderate to severe depression), both based on four-point Likert scale responses[ 20 ]. 2.3.2 Explanatory variables: The explanatory variables for this study were categorized into several key domains to provide a comprehensive analysis. Sociodemographic characteristics included the woman’s age (15–19, 20–24, 25–29, 30–34, 35–39, 40–44, and 45–49 years), educational attainment (no education, primary, secondary, and higher secondary or above), and occupation (housewife, agriculture, non-agriculture, or officer/business/others). Additional variables included marital status (currently married, widowed/divorced/separated), religion (Islam or others), wealth index (poorest, poorer, middle, richer, and richest), sex of the household head (male or female), and household size (≤ 4 or 4 + members). Media and communication exposure was measured media access via newspapers/magazines, radio, and television (not at all, sometimes, or regularly). Maternal and reproductive health factors included number of children (0, 1, 2, 3, or 4+), current pregnancy status (yes or no), and current breastfeeding status (yes or no). Finally, geographic factors comprised administrative division (Barishal, Chattogram, Dhaka, Khulna, Mymensingh, Rajshahi, Rangpur, or Sylhet) and type of residence (urban or rural). 2.4. Data processing and analysis STATA 17 was used for basic exploratory data analysis and statistical analysis to provide a detailed dataset overview with cluster design effect adjustment and Rao–Scott's adjustment. After subsequently, Pandas, NumPy, Scikit-learn, Matplotlib, SHAP, SciPy, and Seaborn were used for data processing, exploration, and model construction in Python 3.10 in Jupyter Notebooks. 2.5. Data pre-processing The BDHS dataset on mental health care-seeking was further processed for machine learning applications. The preliminary phase included data cleansing, addressing inconsistencies and fixing missing values to generate an accurate dataset. A key part of this procedure involved featuring engineering, wherein specific continuous variables were transformed into categorical variables to improve model interpretability and robustness. The Tukey approach detected outliers in continuous variables, but its impact was minor, therefore no adjustments were requiredCategorical outliers were either merged or eliminated if they lacked analytical significance. Normalisation was unnecessary as most variables were already categorical. This comprehensive methodology ensured that the final dataset was properly balanced and optimised for predicting mental health care-seeking behaviour. 2.6. Addressing class imbalance using SMOTE Initial evaluation identified a class imbalance in the outcome variable, with early sought care for symptoms of anxiety or depression being under-represented. To handle this disparity and strengthen the model performance, a synthetic minority oversampling technique (SMOTE) was deployed. By interpolating between existing instances, SMOTE creates synthetic samples for the minority class, "early sought care for symptoms of anxiety or depression". This could provide a balanced dataset for productive model learning from both classes as well as improve predictive accuracy for the under-represented class. 2.7. Feature selection A thorough feature selection procedure identified the most important determinants of early mental health care. Exploratory data analysis (EDA), descriptive statistics, and visualizations can reveal variable distributions and relationships with outcomes. Bivariate study examined correlations between individual variables and early mental health care. Recursive Feature Elimination (RFE) was used to eliminate less significant predictors to refine feature selection. Cramer's V statistic was used to measure predictor correlations, reducing multicollinearity and enhancing feature set resilience. Statistical analysis, machine learning methods, and a comprehensive literature assessment produced excellent model development predictors. 2.8. Feature importance The SHAP was carried out to assess the feature importance. SHAP values offer a consistent way to explain complex ML models by revealing how each feature influences the final prediction. By assigning a significance score to each feature, this analysis provides a more profound insight into the relative importance of multiple factors in predicting outcomes associated with early sought care for mental health. 2.9. Model development This study selected nine machine learning models to identify the predictors of early sought care for mental women: Logistic Regression (LR), K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Gradient Boosting (GBoost), XGBoost, LightGBM, CatBoost, and AdaBoost approaches. These models were chosen to compare projected precision among modeling paradigms. LR was used as a baseline model to predict early sought care using a linear mix of predictors Based on proximity, K-Nearest Neighbours classified data. DT and RF used ensemble methods to increase accuracy. Lastly, GBoost, XGBoost, LightGBM, and CatBoost developed models in an iterative way to enhance performance, with LightGBM optimising efficiency for large datasets, CatBoost handling categorical variables effectively, and, in addition, AdaBoost sequentially correcting errors from weak learners to improve model performance.To assess their performance, we utilised metrics such as accuracy, precision, recall, and F1-score, as well as evaluated their capacity to identify key factors that had an effect. The best-performing model was chosen after considering all these factors. 2.10. Model training and evaluation The dataset was split into training (80%) and testing (20%) sets for model robustness. After training on the training set, the models were tested on the unseen test set using accuracy, precision, recall, F1-score, Area Under the Receive Operating Characteristic Curve (AUROC), and confusion matrices. Stratified k-fold cross-validation (k = 5) enhanced model generalisability and minimized overfitting by separating data into k-folds for iterative training and testing. 2.11. Model selection To identify the most effective model for predicting early sought care for symptoms of anxiety or depression, multiple assessment metrics were performed. The confusion matrix provides intricate analysis of the model by categorizing predictions into true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). The fundamental parameter for evaluating models was accuracy, defined as the proportion of right predictions. We evaluated model performance using precision (rate of accurately predicted positive instances to all positive predictions) and recall (percentage of correctly identified positive cases). The F1 score, balancing various metrics, offers a complete assessment of model effectiveness. The model with the best mix of indicators predicted early sought care for anxiety or depression symptoms after comparing metrics across all tested machine learning algorithms. 3. RESULTS 3.1 Characteristics of the Study Population Among 4,255 ever-married women aged 15–49 years with symptoms of anxiety or depression, approximately 20% reported ever seeking care. Care-seeking was significantly higher among women aged 30–39 years (18.0%-20.2%) than those aged 15–24 years (10.1%; p = 0.0316), among non-breastfeeding women (86.3%; p = 0.0128), and differed significantly by region (p < 0.001), with higher rates in Chattogram (21.9%), Khulna (17.5%), and Rajshahi (14.3%) and lower rates in Barishal (4.0%) and Sylhet (4.8%); women’s occupation was also strongly associated with care-seeking (p < 0.001). In contrast, education level (p = 0.2484) and urban-rural residence (p = 0.1057) was not significantly associated with care-seeking behavior (Table 1 ). Table 1 Socio-Demographic Characteristics of the Study Participants (N = 4255) Characteristics Total (weighted) N (%) Ever sought care (weighted) P-value * No n (%) Yes n (%) Woman’s age (years) 15–19 318 (7.5%) 269 (7.9%) 49 (5.7%) 20–24 536 (12.6%) 449 (13.2%) 87 (10.1%) 25–29 694 (16.3%) 563 (16.6%) 131 (15.2%) 30–34 753 (17.7%) 598 (17.6%) 155 (18.0%) 0.0316 35–39 797 (18.7%) 624 (18.4%) 173 (20.2%) 40–44 596 (14.0%) 464 (13.7%) 132 (15.4%) 45–49 562 (13.2%) 430 (12.7%) 132 (15.4%) Woman’s education No education 714 (16.8%) 585 (17.2%) 129 (15.0%) Primary 1217 (28.6%) 980 (28.9%) 236 (27.5%) 0.2484 Secondary 1865 (43.8%) 1476 (43.5%) 389 (45.3%) Higher secondary or above 459 (10.8%) 354 (10.4%) 104 (12.1%) Woman’s occupation Housewife 2630 (61.8%) 2177 (64.1%) 453 (52.8%) Non-agriculture 341 (8.0%) 271 (8.0%) 70 (8.1%) Agriculture 415 (9.8%) 328 (9.7%) 88 (10.2%) 0.0000 Officer/ Businessmen/ Others 868 (20.4%) 620 (18.3%) 248 (28.9%) Marital status Currently married 3921 (92.2%) 3131 (92.2%) 790 (92.0%) Widowed/ Divorced/ Separated 334 (7.8%) 265 (7.8%) 69 (8.0%) 0.8469 Religion Islam 3922 (92.2%) 3137 (92.4%) 785 (91.4%) 0.5056 Others 333 (7.8%) 259 (7.6%) 74 (8.6%) Wealth index Poorest 832 (19.6%) 685 (20.2%) 147 (17.2%) Poorer 896 (21.1%) 725 (21.4%) 171 (19.9%) Middle 910 (21.4%) 728 (21.5%) 182 (21.1%) 0.0971 Richer 846 (19.9%) 644 (19.0%) 202 (23.5%) Richest 771 (18.1%) 614 (18.1%) 157 (18.3%) Household head’s sex Male 3548 (83.4%) 2850 (83.9%) 698 (81.3%) Female 707 (16.6%) 546 (16.1%) 161 (18.7%) 0.1215 Household members ≤ 4 2161 (50.8%) 1720 (50.7%) 440 (51.3%) 4+ 2094 (49.2%) 1676 (49.4%) 419 (48.7%) 0.7759 Mass media exposure No 2108 (49.5%) 1701 (50.1%) 406 (47.3%) Yes 2147 (50.5%) 1695 (49.9%) 453 (52.7%) 0.2079 Number of children 0 423 (10.0%) 352 (10.4%) 71 (8.3%) 1 882 (20.7%) 713 (21.0%) 169 (19.6%) 2 1458 (34.3%) 1178 (34.7%) 280 (32.6%) 0.0676 3 943 (22.2%) 733 (21.6%) 211 (24.5%) 3+ 549 (12.9%) 420 (12.4%) 128 (14.9%) Current pregnancy status No 4005 (94.1%) 3203 (94.3%) 802 (93.4%) Yes 250 (5.9%) 193 (5.7%) 57 (6.7%) 0.3414 Current breastfeeding status No 3539 (83.2%) 2798 (82.4%) 741 (86.3%) 0.0128 Yes 716 (16.8%) 598 (17.6%) 118 (13.7%) Division Barishal 245 (5.8%) 210 (6.2%) 35 (4.0%) Chattogram 812 (19.1%) 624 (18.4%) 188 (21.9%) Dhaka 944 (22.2%) 785 (23.1%) 159 (18.5%) Khulna 535 (12.6%) 385 (11.3%) 150 (17.5%) Mymensingh 279 (6.6%) 231 (6.8%) 48 (5.6%) 0.0009 Rajshahi 563 (13.2%) 440 (13.0%) 123 (14.3%) Rangpur 640 (15.0%) 525 (15.5%) 114 (13.3%) Sylhet 237 (5.6%) 197 (5.8%) 41 (4.8%) Residence area Urban 1173 (27.6%) 961 (28.3%) 212 (24.6%) Rural 3082 (72.4%) 2435 (71.7%) 647 (75.4%) 0.1057 *Chi-Square test of Independence with Rao-Scott’s Adjustment for Survey Design. 3.2 Model Performance Evaluation Table 2 and Fig. 2 summarize the performance of the nine machine learning algorithms applied to predict early care-seeking for anxiety or depression. Model accuracy ranged from 0.55 (LR) to 0.66 (RF). RF achieved the highest overall performance with an accuracy of 0.66, recall of 0.69, and F1-score of 0.67, outperforming other models in most metrics. CatBoost and Gradient Boosting also performed competitively, with F1-scores of 0.67 and 0.66, respectively. LR recorded the lowest predictive performance (accuracy = 0.55, F1 = 0.54), indicating its limited suitability for this complex, non-linear dataset. Table 2 Model Evaluation Models accuracy precision recall f1_score auroc LR 0.55 0.57 0.52 0.54 0.55 KNN 0.61 0.65 0.52 0.58 0.61 DT 0.61 0.63 0.57 0.60 0.61 RF 0.66 0.66 0.69 0.67 0.66 GB 0.63 0.63 0.69 0.66 0.63 XGB 0.64 0.65 0.66 0.66 0.64 LightGBM 0.64 0.64 0.68 0.66 0.64 CatBoost 0.65 0.64 0.71 0.67 0.64 AdaBoost 0.60 0.61 0.64 0.62 0.60 3.3 Confusion Matrix Analysis The confusion matrix analysis further supported these findings (Table 3 and Fig. 3 ). The RF model correctly classified 368 true positive cases and 311 true negatives, yielding balanced sensitivity and specificity. Ensemble-based models such as Random Forest, CatBoost, and LightGBM demonstrated superior discrimination compared to individual classifiers like KNN and DT. Table 3 Confusion Matrix Model True Negative False Positive False Negative True Positive LR 291 211 257 275 KNN 351 151 255 277 DT 321 181 227 305 RF 311 191 164 368 Gboost 287 215 164 368 XGBoost 314 188 181 351 LightGBM 299 203 169 363 CatBoost 290 212 154 378 AdaBoost 281 221 190 342 3.4 ROC Curve Analysis The ROC curve (Fig. 4 ) and AUROC analysis indicated that Random Forest exhibited the best discriminative ability (AUROC = 0.70), followed closely by XGBoost, CatBoost, and LightGBM (AUROC = 0.69 each). In contrast, Logistic Regression showed the weakest discrimination (AUROC = 0.57). While the AUROC values across models were below 0.75 indicating moderate classification power the ensemble models consistently outperformed traditional and distance-based classifiers. 3.5 Model Explainability: SHAP Analysis The SHAP approach was used to interpret the feature importance and direction of effects in the best-performing RF model. The SHAP summary plot (Fig. 5 ) showed that variables such as administrative division, age group, number of children, household size, and media exposure exerted the strongest influence on the model’s predictions. Both positive and negative SHAP values indicated heterogeneous effects across individual predictions. The SHAP feature ranking plot (Fig. 6 ) highlighted division and age group as the most influential predictors (mean SHAP value = 0.05), followed by number of children, household size, mass media exposure, and residence. In contrast, pregnancy status, religion, current breastfeeding, and sex of household head contributed minimally to model predictions. These findings underscore the importance of demographic and regional disparities in shaping women’s likelihood of seeking mental health care, emphasizing the complex interplay between social, familial, and contextual factors. 4. DISCUSSION Using data from the 2022 Bangladesh Demographic and Health Survey, this study used ML algorithms to predict early anxiety and depression care-seeking in reproductive-age women in Bangladesh. Among the nine algorithms tested, ensemble-based models particularly Random Forest demonstrated superior performance, achieving the highest accuracy (0.66), recall (0.69), and F1-score (0.67). These results are consistent with findings from previous ML-based health behavior studies where ensemble learners, notably Random Forest (RF) and Gradient Boosting (GBoost), outperformed traditional models in predicting fertility preferences [ 21 ], optimal antenatal care utilization [ 22 ], and institutional delivery dropout [ 23 ]. In addition, RF model demonstrated the highest accuracy in predicting depression [ 24 ] (Rahman et al., 2024), and COPD admissions [ 25 ]. The ability of these models to handle nonlinear interactions and multicollinearity among features underscores their robustness for analyzing complex, multidimensional health datasets such as the BDHS. The overall moderate discriminative performance (AUROC = 0.55–0.70) reflects the inherent complexity of mental health-related behaviors, which are shaped by multifaceted psychological, social, and structural influences. Unlike biomedical outcomes, care-seeking for anxiety or depression is strongly influenced by intangible factors such as stigma, perceived need, and family support that are not typically captured in large demographic surveys [ 26 ]. Despite these limitations, the observed predictive accuracy demonstrates that ML approaches can extract meaningful behavioral insights even from limited sociodemographic and contextual data. This observation parallels findings from earlier ML applications in reproductive and maternal health, where similar moderate accuracies were achieved but still provided actionable insights for programmatic intervention [ 22 , 23 ]. The integration of SHapley Additive exPlanations (SHAP) in this study provided a transparent and interpretable view of how each feature contributed to model predictions. The SHAP analysis revealed that administrative division, age group, number of children, household size, and mass media exposure were the most influential predictors of early care-seeking for anxiety and depression. These findings are consistent with evidence from other ML studies that demonstrated the dominant role of geographic, demographic, and household characteristics in predicting health-seeking behaviors [ 21 , 24 ]. Regional disparities observed in SHAP outputs reaffirm that place of residence remains a strong determinant of access to mental health services, likely due to uneven availability of infrastructure and social stigma varying across divisions. The significance of age and household characteristics similarly reflects life-stage and caregiving constraints that may influence women’s prioritization of their own mental well-being. The interpretability provided by SHAP is a major strength of this study and a key differentiator from earlier predictive modelling approaches. One of the major challenges of ML in public health research is the “black box” nature of most algorithms, which limits their utility for policy formulation. SHAP mitigates this by decomposing predictions into additive feature contributions, thus quantifying how individual variables influence the outcome probability. This interpretability makes ML models more transparent and trustworthy, a methodological advancement that has also been emphasized in previous explainable ML studies in maternal and reproductive health domains [ 22 , 27 ]. By highlighting feature importance in interpretable terms, SHAP enables translation of computational insights into practical, evidence-based recommendations for mental health policy and service delivery. This analysis corroborates earlier ML-based studies in low- and middle-income countries (LMICs) that revealed sociodemographic inequalities as important drivers of health service utilization [ 15 , 24 , 28 ]. The consistent role of structural and regional factors across multiple health outcomes indicates a systemic influence of geographic access, socioeconomic status, and autonomy on women’s health-seeking behavior. For mental health, this underscores the need to address social and infrastructural inequities and strengthen awareness efforts, particularly in underserved regions. Nevertheless, this study provides important methodological and policy insights, demonstrating the superiority of ensemble models and the value of SHAP for interpretability in public health research. Explainable ML can help policymakers identify subgroups with low mental health care–seeking and design targeted interventions to address context-specific barriers [ 21 , 23 , 28 ], integrating data-driven predictions with interpretability frameworks strengthens the empirical foundation for health policy design in LMICs. In Bangladesh, such approaches could be instrumental for scaling up decentralized mental health services, promoting community-based counselling, and addressing regional inequalities in access to care. In summary, this study highlights the growing importance of explainable machine learning in understanding and predicting health behaviors. By combining predictive accuracy with interpretability, Random Forest modeling and SHAP analyses provide an innovative framework for identifying key determinants of mental health care-seeking among women, enabling data-driven strategies to promote equitable and effective mental health interventions in Bangladesh and similar LMIC settings. 6. Conclusion Explainable machine learning was shown to predict early anxiety and depression care-seeking in reproductive-age Bangladeshi women using nationally representative data from the 2022 BDHS. In nine algorithms studied, the Random Forest model had the best prediction performance, capturing complicated, nonlinear sociodemographic and environmental interactions. SHapley Additive exPlanations (SHAP) improved model transparency by identifying administrative division, age group, number of children, household size, and media exposure as the biggest mental health care-seeking factors. These findings emphasize the importance of predictive precision and interpretability, geographical differences, information access, and women's autonomy in mental health behavior. Explainable ML in public health research and surveillance can help policymakers identify at-risk populations and design data-driven, equitable interventions to improve early mental health care-seeking and universal mental health coverage in Bangladesh and other countries. 7. Strengths And Limitations A major strength of this study lies in its use of a large, nationally representative dataset (BDHS 2022) and its application of multiple state-of-the-art machine learning (ML) algorithms, enabling a robust comparative evaluation of predictive performance. The inclusion of SHAP interpretability represents an additional methodological advantage, as it bridges the gap between computational complexity and policy relevance by quantifying the contribution of each predictor to model outcomes. This study thus extends prior ML-based public health work by offering an interpretable and generalizable model of women’s mental health care-seeking behavior in a low- and middle-income context. Several limitations must be noted. First, the models' predictive ability was moderate because DHS surveys did not include psychosocial and attitudinal factors such perceived stigma, social support, and mental health service awareness, which are important mental health behavior predictors. Second, the cross-sectional design reduces predictor-outcome causality. Third, the Synthetic Minority Oversampling Technique (SMOTE) reduced class imbalance but may cause synthetic noise that hinders model generalization. SHAP enhances interpretability but does not show causality. Finally, since the models were trained and validated on one country dataset, their external validity to other populations should be verified before generalization. Despite these limitations, the integration of explainable ML into national health research represents a methodological advancement. The study provides a replicable framework for predicting and interpreting health-seeking behaviors using DHS data, contributing to the broader goal of data-driven, equitable, and transparent public health policymaking in low-resource settings. Declarations Acknowledgements We thank the National Institute of Population Research and Training (NIPORT), ICF, and the Monitoring and Evaluation to Assess and Use Results Demographic and Health Surveys for permission to utilize the BDHS dataset. Declaration of generative AI use The AI-assisted platforms used included SciSpace (for literature searches), Gemini (for brainstorming), and QuillBot (for paraphrasing). After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article. Author contributions M.R.K. and J.S. conceptualized the study, developed the methodology, performed data curation, formal analysis, software development, and visualization, and drafted the initial manuscript. T.A. contributed to data curation, methodology, and reviewed and edited the manuscript. M.R.A. contributed to data curation, and reviewed and edited the manuscript. F.M.A. supervised and validated the study, refined the methodology, and contributed to interpreting results, editing, and proofreading the manuscript. All authors reviewed, edited, and approved the final manuscript. Funding The authors declare that no funds, grants, or other financial support were received for this research. Data availability The datasets analyzed during the current study are available from the Demographic and Health Surveys Program repository upon request and approval: https://dhsprogram.com/data/. Clinical Trial Number Not applicable. Ethical approval The 2022 Bangladesh Demographic and Health Survey (BDHS) received ethical approval from the ICF Institutional Review Board (IRB), Maryland, USA, and the National Research Ethics Committee of the Bangladesh Medical Research Council (BMRC), Dhaka, Bangladesh. Survey procedures complied with the U.S. Department of Health and Human Services regulations for the protection of human subjects (45 CFR 46) and with national ethical standards. Written informed consent was obtained from all participants prior to enrolment. This study used anonymized, publicly available secondary data obtained with permission from the DHS Program; therefore, no additional ethical approval was required. All data were handled confidentially. Full details of DHS ethical approvals are available https://dhsprogram.com/Methodology/Protecting-the-Privacy-of-DHS-Survey-Respondents.cfm. Consent to participate All participants provided informed and voluntary consent before being interviewed. As the use of the dataset adhered to the ethical standards set by the DHS Program, this consent was obtained in accordance with ethical guidelines, ensuring that participants fully understood the purpose of the study and their right to withdraw at any time without consequence. Consent for publish Not applicable. Competing interests The authors declare that they have no known financial or personal relationships that could have influenced the work reported in this manuscript. References World Health Organization. Mental disorders [Internet]. 2025 [cited 2025 Nov 24]. Available from: https://www.who.int/news-room/fact-sheets/detail/mental-disorders World Health Organization. Over a billion people living with mental health conditions – services require urgent scale-up [Internet]. 2025 [cited 2025 Nov 24]. Available from: https://www.who.int/news/item/02-09-2025-over-a-billion-people-living-with-mental-health-conditions-services-require-urgent-scale-up Patel V, Saxena S, Lund C, Thornicroft G, Baingana F, Bolton P, et al. The Lancet Commission on global mental health and sustainable development. The Lancet. Volume 392. Lancet Publishing Group; 2018. pp. 1553–98. Shawon MSR, Hossain FB, Hasan M, Rahman MR. Gender differences in the prevalence of anxiety and depression and care seeking for mental health problems in Nepal: Analysis of nationally representative survey data. Global Mental Health. 2024;11. Islam MS, Rahman ME, Moonajilin MS, van Os J. Prevalence of depression, anxiety and associated factors among school going adolescents in Bangladesh: Findings from a cross-sectional study. PLoS ONE. 2021;16(4 April). Ahad AA, Sanchez-Gonzalez M, Junquera P. Understanding and Addressing Mental Health Stigma Across Cultures for Improving Psychiatric Care: A Narrative Review. Cureus. 2023. Komu CK, Ngigi M, Melson AJ. Barriers and Facilitators to Accessing Mental Health Services for Adults in Sub-Saharan Africa: A Systematic Review. Mental Health Sci. 2025;3(1). Koly KN, Tasnim Z, Ahmed S, Saba J, Mahmood R, Farin FT et al. Mental healthcare-seeking behavior of women in Bangladesh: content analysis of a social media platform. BMC Psychiatry. 2022;22(1). Government of the. People’s Republic of Bangladesh. National Mental Health Strategic Plan 2020–2030. Dhaka. World Health Organization. World mental health report: transforming mental health for all. Geneva; 2022. Rameez S, Nasir A. Barriers to mental health treatment in primary care practice in low- and middle-income countries in a post-covid era: A systematic review. J Family Med Prim Care. 2023;12(8):1485–504. Al Azdi Z, Saif SI, Ashraf Kushal S, Islam MT, Maaz L, Reza S et al. Gender differences in mental health help-seeking behaviour in Bangladesh: Findings from a cross-sectional online survey. BMJ Open. 2025;15(5). Rafi MA, Anika US, Hasan MT, Hossain MG. Association between women’s empowerment and mental health help-seeking behaviour in Bangladesh: findings from a nationally representative survey. BMJ Open. 2025;15(9). Raza S, Banik R, Noor STA, Sayeed A, Saha A, Jahan E, et al. Anxiety and depression among reproductive-aged women in Bangladesh: burden, determinants, and care-seeking practices based on a nationally representative demographic and health survey. Arch Womens Ment Health. 2025;28(5):1125–41. Huque R, Azad AK, Islam K, Ahmed HU, Amin MR. Mental health care seeking behavior in Bangladesh: determinants and treatment gaps. BMC Psychiatry. 2025;25(1). Ahmed T, Luthfa ST, Chaahat AH, Kanchon MR, Sharif A, Bin. Determinants of early mental health help-seeking among women in Bangladesh: A nationally representative bootstrapped regression analysis. PLOS Mental Health. 2025;2(9):e0000420. Alamirew B, Darge BD, Terefe B, Gebremeskel F. Utilization of mental health services and associated factors among residents of southern Ethiopia; a community based cross-sectional study. BMC Health Serv Res. 2025;25(1). Obermeyer Z, Emanuel EJ. Predicting the Future — Big Data, Machine Learning, and Clinical Medicine. N Engl J Med. 2016;375(13):1216–9. Lundberg S, Lee SI. A Unified Approach to Interpreting Model Predictions. 2017; Available from: http://arxiv.org/abs/1705.07874 National Institute of Population Research and Training (NIPORT), ICF. Bangladesh Demographic and Health Survey 2022 Final Report [Internet]. Dhaka, Bangladesh, and Rockville, Maryland, USA. 2024. Available from: Sani J, Halane S, Ahmed AM, Ahmed MM. Application of machine learning algorithms and SHAP explanations to predict fertility preference among reproductive women in Somalia. Sci Rep. 2025;15(1). Tadese ZB, Sani J, Kebede SD, Kitil GW, Dube GN, Nimani TD. Sociodemographic determinants and regional disparities of first-trimester antenatal care initiation among Nigerian women: a multilevel analysis of 2018 NDHS data. Front Glob Womens Health [Internet]. 2025;6. Available from: https://www.frontiersin.org/articles/ 10.3389/fgwh.2025.1502905/full Sani J, Halane S, Ahmed MM, Ahmed AM, Mohamoud JH. Exploring the application of machine learning and SHAP explanations to predict health facility deliveries in Somalia. Discover Artif Intell. 2025;5(1). Rahman MA, Kohli T. Mental health analysis of international students using machine learning techniques. PLoS ONE. 2024;19(6 June). Alhur AA, Sani J, Ahmed MM. Predicting COPD admissions using machine learning and SHAP: An exploratory multi-hospital study in Riyadh, Saudi Arabia. Intell Based Med [Internet]. 2025;12:100312. Available from: https://linkinghub.elsevier.com/retrieve/pii/S2666521225001164 Gulliver A, Griffiths KM, Christensen H. Perceived barriers and facilitators to mental health help-seeking in young people: A systematic review. BMC Psychiatry. 2010;10. Ribeiro MT, Wu T, Guestrin C, Singh S. Beyond Accuracy: Behavioral Testing of NLP Models with CheckList [Internet]. 2020. Available from: https://github.com/marcotcr/checklist Raihana Z, Kader M, Islam MZ, Bornee FA, Mondal MNI, Chowdhury MRK et al. Factors associated with the presence of anxiety and depression symptoms in rural hypertensive adults in Bangladesh: leveraging extreme gradient booster machine learning algorithm. Front Psychol. 2025;16. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 23 Jun, 2026 Read the published version in Discover Public Health → Version 1 posted Editorial decision: Revision requested 27 Mar, 2026 Reviews received at journal 26 Mar, 2026 Reviewers agreed at journal 20 Feb, 2026 Reviews received at journal 17 Feb, 2026 Reviewers agreed at journal 17 Feb, 2026 Reviewers agreed at journal 17 Feb, 2026 Reviewers agreed at journal 15 Feb, 2026 Reviewers agreed at journal 11 Feb, 2026 Reviewers invited by journal 09 Feb, 2026 Editor assigned by journal 06 Feb, 2026 Editor invited by journal 29 Jan, 2026 Submission checks completed at journal 28 Jan, 2026 First submitted to journal 28 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information

Related documents

• Predictors of Undiagnosed Diabetes in the Bangladeshi Female Population: A Propensity Score-Weighted Machine Learning Analysis of BDHS Biomarker Data.#305095
• Machine Learning Approaches to Identify Influential Factors of the Comorbid Psychiatric Symptoms and Hypertension Among Rural Adults in Bangladesh: A Cross-Sectional Study.#430762
• Machine learning-based prediction and risk factor analysis of depression among reproductive-aged women in Bangladesh: Findings from the BDHS 2022.#358524
• Short-term predictive ability of selected cardiovascular risk prediction models in a rural Bangladeshi population: A case-cohort study#222222
• Integrative Early Breastfeeding Support in NICU#842201