Hypertension is a major modifiable risk factor for cardiovascular disease and is highly prevalent in low- and middle-income countries. The authors used Bangladesh Demographic and Health Survey (BDHS) 2022 data to estimate hypertension prevalence among adults, to examine associations with socio-demographic and clinical factors, and to compare the predictive performance of multiple machine learning (ML) and deep learning (DL) algorithms for identifying individuals with hypertension.
The study aimed to identify which predictive approach best balances accuracy and sensitivity and to determine the most influential risk factors in the BDHS dataset.
The analysis used cross-sectional data from BDHS 2022, collected between June 27 and December 12, 2022. The survey employed two-stage stratified sampling and included biomarker measurements taken by trained enumerators. The analytic sample comprised 14,283 adults aged 18 years and older. Data access is through the DHS Program repository, which requires registration and approval.
Hypertension was defined according to standard thresholds used in the study: systolic blood pressure ≥ 140 mmHg and/or diastolic blood pressure ≥ 90 mmHg. Predictor variables included socio-demographic, anthropometric, and clinical factors captured in BDHS 2022. The manuscript reports associations assessed via chi-square testing and variable importance in supervised predictive models. Specific variable lists used for modeling were those available in the BDHS dataset; the source reports age, BMI, sex, diabetes status, wealth index, education, household size, and region among the examined variables.
Six supervised models were trained and tested: four machine learning methods—weighted logistic regression (WLR), random forest (RF), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM)—and two deep learning methods—TabNet and a multi-layer perceptron (MLP). The weighted logistic regression accounted for survey/sample weights. The authors applied standard classification workflows and evaluated models on held-out test data.
Model performance was reported with multiple metrics to capture different aspects of performance: accuracy, precision, recall (sensitivity), specificity, F1-score, area under the receiver operating characteristics curve (AUC-ROC), and area under the precision-recall curve (AUC-PR). The combination of discrimination (AUC-ROC), precision-recall summaries, and sensitivity/specificity was used to compare models in the context of hypertension detection.
Overall hypertension prevalence in the study sample was 18.04% (95% CI: 17.2%–18.9%). Prevalence was higher in women (18.87%) than in men (16.97%). Chi-square testing identified significant associations between hypertension and age, BMI, diabetes, wealth index, education, household size, and region (p < 0.05).
On predictive performance, weighted logistic regression achieved the highest accuracy (0.817), highest precision (0.444), highest specificity (0.981), highest AUC-ROC (0.751), and highest AUC-PR (0.357) among evaluated models. However, WLR had notably low recall (0.070), indicating it missed a large proportion of true hypertension cases.
In contrast, the random forest model produced the highest recall (0.687) and the highest F1-score (0.460) on the test data, reflecting greater sensitivity and a more balanced trade-off between precision and recall. These results indicate that while WLR performed well on overall accuracy and discrimination metrics, RF was more effective at identifying individuals with hypertension.
Across models the variables identified as most important for predicting hypertension were age, BMI, sex, family size, and educational level. Diabetes status, wealth index, and geographic region were also associated with hypertension in bivariate testing and considered among predictors in the models. The manuscript reports these predictors based on variable importance outputs from the applied algorithms.
The authors note that hypertension is common in Bangladesh and unevenly distributed across socio-demographic groups. Although weighted logistic regression delivered the best performance on several global metrics, its low sensitivity limits usefulness when the goal is to detect hypertensive individuals in a population. Random forest’s higher recall and F1-score suggest it may be more suitable for public-health screening applications where sensitivity is prioritized.
The authors recommended external validation and further assessment of clinical utility before deploying the models in practice. As this work used cross-sectional BDHS data, causal inference is not possible; the study design and reliance on survey-collected variables also impose limits on model generalizability and prospective predictive value.
Using BDHS 2022 data, the study found an 18.04% prevalence of hypertension among adults, with higher prevalence in women and significant associations with age, BMI, diabetes, wealth, education, household size, and region. While WLR achieved the highest accuracy and discrimination metrics, its low recall reduces its suitability for screening. Random forest may be preferred when sensitivity is desired, but additional external validation and clinical utility assessments are required prior to implementation.
The BDHS 2022 data used in this analysis are available from the Demographic and Health Surveys (DHS) Program subject to registration and approval; the authors stated they could not share the data directly due to data use restrictions. The study reported no specific funding and declared no competing interests.