Application of Supervised Machine Learning Models in the Prediction of Arterial Hypertension Risk

Alhandra Alves Viégas; Renato Máximo Sátiro

DOI: 10.22167/2675-6528-202602620

Article derived from a Course Conclusion Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

Hypertension constituted a significant public health problem, associated with clinical, metabolic, and behavioral factors. This study aimed to develop and evaluate supervised XGBoost and Logistic Regression models to classify hypertension risk in adults, as well as identify the variables with the greatest contribution to the predictions. The Hypertension Risk Prediction Dataset, available on the Kaggle platform, comprising 174,982 records and 23 variables, was used. The research adopted a quantitative, descriptive, exploratory, and predictive approach. The data were subjected to exploratory analysis, categorical variable encoding, and stratified splitting into 80% for training and 20% for testing. Model performance was evaluated using classification metrics and the Precision-Recall curve, and interpretability was analyzed using the SHAP technique. The results indicated a greater contribution of variables such as systolic and diastolic blood pressure, age, and body mass index to the predictions. However, XGBoost presented an Average Precision of 0.715, a value lower than the baseline classifier of 0.719, evidencing restricted discriminative capacity. Logistic Regression also exhibited limited performance, with an AUC-ROC of 0.49. It was concluded that the models allowed for the exploration of factors associated with hypertension classification and increased the interpretability of predictions, but the class imbalance and the obtained performance limited their predictive application, indicating the need for further adjustments. The models are experimental data analysis instruments, not validated clinical tools.

Keywords: Binary classification; Clinical factors; Interpretability; Logistic Regression; XGBoost.

1. Introduction

Arterial hypertension is a significant public health problem, given its association with morbidity and mortality and the compromise of vital organs, such as the brain, kidneys, and heart. High blood pressure is recognized as one of the main mortality risk factors, being influenced by genetic and environmental components (Springer, 2024; Mroz et al., 2024).

Estimates indicate that the number of individuals with hypertension may reach 1.5 billion by 2025, with a higher prevalence among people in older age groups (Felix et al., 2024). The prevalence of the disease increases with age, potentially reaching approximately 60% among older people. This scenario is related to factors such as sedentary lifestyle, increased body fat, and less physical exercise (Adeleke et al., 2024).

Besides age, several behavioral and metabolic characteristics are associated with the development of arterial hypertension. Inadequate diet, characterized by high consumption of salt and saturated fats and low intake of fruits and vegetables, is a risk factor (Pereira and Fisberg, 2024). Other elements, such as smoking, sedentary lifestyle, alcohol consumption, and excess weight, also contribute to the increased risk of the disease (Cheraghi et al., 2025). The joint identification of these characteristics is fundamental to broadening the understanding of factors associated with hypertension and supporting monitoring strategies.

In this context, Machine Learning techniques have been employed in the analysis of health data. They allow the processing of large volumes of information and the identification of patterns and relationships between different variables, complementing traditional statistical methods. These approaches contribute to classification and prediction tasks, aiding in decision-making (Bisong et al., 2024; Martínez-García et al., 2025). Different methods, including ensemble models, neural networks, and supervised learning techniques, are used in the analysis of medical data to identify relevant patterns in clinical conditions (Guo et al., 2024).

Supervised learning, one of the Machine Learning approaches, uses previously labeled observations to develop models capable of classifying new records (Bora et al., 2025). In this study, the XGBoost and Logistic Regression algorithms were employed. XGBoost allows exploring non-linear relationships and interactions between variables, while Logistic Regression serves as a linear comparison approach. The combination of these methods enables the analysis of the behavior of different classification strategies applied to clinical, metabolic, demographic, and lifestyle data (Hua et al., 2025).

Given this scenario, the investigation sought to answer whether it is possible to classify the risk of arterial hypertension in adults through supervised Machine Learning algorithms, considering clinical and lifestyle data. The research is relevant for allowing the evaluation of predictive performance and identifying which variables contribute most to the models’ classifications. The interpretation of these results also makes it possible to identify limitations related to the characteristics of the database and the behavior of the algorithms used.

Thus, this study aimed to develop and evaluate supervised XGBoost and Logistic Regression models to classify the risk of arterial hypertension in adults based on clinical, metabolic, demographic, and lifestyle variables, as well as to identify the variables with the greatest contribution to the predictions made.

2. Material and Methods

The research was conducted in the second half of 2025 and adopted a quantitative, descriptive, exploratory, and predictive approach. The central objective was the development and evaluation of supervised Machine Learning models, intended for the classification of arterial hypertension risk in adults.

The stages of organization, processing, exploratory analysis, modeling, and data evaluation were executed in the Python language. For this, the pandas, seaborn, scikit-learn, matplotlib, XGBoost, and SHAP libraries were used, which assisted in the implementation of the methodological procedures.

The data source was the Hypertension Risk Prediction Dataset, a dataset made publicly available on the Kaggle platform (Pandey, 2025). This dataset comprised 174,982 records of adult individuals, aged between 18 and 89 years, and a total of 23 variables.

The variables included demographic, anthropometric, clinical, metabolic, and behavioral characteristics. The target variable, named Hypertension, was categorized into two classes, High and Low, for the binary classification task, while the other variables were considered predictor variables.

Initially, the dataset structure and variable types were inspected. The database was provided pre-processed and without missing records, dispensing with the application of value imputation techniques.

Categorical variables were converted to numerical representation through “one-hot” encoding, allowing their use by supervised algorithms. For the characterization of quantitative variables, the mean, standard deviation, minimum value, and maximum value were calculated.

Additionally, a correlation matrix was developed using Spearman’s coefficient, applied to the quantitative variables. This procedure aimed to evaluate the intensity and direction of monotonic associations between clinical and metabolic attributes.

After preprocessing, the dataset was stratified into 80% for model training and 20% for testing. Stratification ensured the preservation of the proportion of target variable categories in both subsets.

The test set was reserved for evaluating the models’ performance on records not used during the training phase. Imbalance was observed between the classes of the target variable; however, no resampling techniques, such as Synthetic Minority Over-sampling Technique [SMOTE], nor class weight adjustments were applied.

It was decided to preserve the original distribution of the dataset (Knox, 2026), considering that this approach would be more representative of the real data scenario. This decision was relevant for the interpretation of the performance metrics.

The first model employed was Extreme Gradient Boosting [XGBoost], a supervised algorithm based on the sequential construction of decision trees. In this procedure, each new tree was added to reduce the errors of previous estimators, incorporating a regularization term to control complexity and mitigate overfitting (Chen and Guestrin, 2016; Zhuo et al., 2025).

The XGBoost objective function was defined by the sum of the loss function, which measures the difference between the observed value and the prediction, and a regularization term, which controls the complexity of the trees added to the model. The model was configured with 300 estimators, a maximum depth of four, a learning rate of 0.05, and a sampling proportion of 0.8 for records and variables.

Binary classification was used, with the area under the curve as the evaluation metric during training and a random seed equal to 42. The hyperparameters were previously defined, without automated search or cross-validation.

Logistic Regression was used as a comparative linear model to XGBoost to estimate the probability of the positive class from a linear combination of the explanatory variables (Dey et al., 2025; Hua et al., 2025). The estimated probability of the positive class occurrence was modeled through the logistic function, which linearly relates the explanatory variables to the logarithm of the outcome’s odds, using coefficients and an intercept.

The coefficients of Logistic Regression were used to analyze the relationship between the explanatory variables and the outcome through odds ratios (OR). OR values greater than one indicated a positive association, while values less than one indicated an opposite association, with confidence intervals considered in the interpretation of the magnitude and uncertainty of the associations.

The evaluation of the models’ predictive performance was carried out on the test set, considering metrics such as accuracy, sensitivity, specificity, precision, recall, F1-score, and area under the Receiver Operating Characteristic curve [AUC-ROC]. Due to the imbalance of the target variable, the Precision-Recall curve was used to analyze the relationship between precision and sensitivity at different classification thresholds.

The overall performance in this analysis was synthesized by the Average Precision [AP] metric and compared to a baseline corresponding to the proportion of the positive class. The decision curve was also used to analyze the net benefit of classifications at different probability cutoffs, comparing the model with strategies of considering all or no records as positive.

The calibration of the forecasts was analyzed to verify the correspondence between the estimated probabilities and the observed proportions, complementing the discriminative metrics. This analysis allowed us to verify the behavior of the predicted probabilities in different risk ranges.

The interpretability of XGBoost predictions was deepened by the SHapley Additive exPlanations [SHAP] technique, which measured the individual contribution of variables to the model’s output. SHAP values allowed for the analysis of the magnitude and direction of attribute contributions in predictions.

For the SHAP analysis, an XGBoost model was used, configured with 100 estimators, a maximum depth of four, a learning rate of 0.1, a sampling proportion of 0.8, “log loss” as the evaluation metric, and a random seed of 42. SHAP values were analyzed globally, through the distribution of variable contributions and their average absolute importance, and specifically for systolic blood pressure.

The research exclusively used publicly accessible secondary data, with no primary data collection, intervention, or direct contact with participants. No elements that could enable individual identification of records were used, and the study fell under Resolution of the National Health Council No. 510, of April 7, 2016, art. 1st, sole paragraph, item II, thus waiving submission to a Research Ethics Committee. The anonymous characteristics of the database were preserved, and all sources used were duly cited and referenced.

3. Results and Discussion

The research results were analyzed to characterize the dataset used, evaluate the behavior of the developed supervised models, and identify the variables that contributed most to the classification of arterial hypertension risk. The interpretation of the findings considered descriptive analyses, correlations between attributes, variable importance, interpretability through the SHAP technique, the discriminative performance of the models, the calibration of predictions, the decision curve, and the odds ratios obtained by Logistic Regression. The inherent limitations of the dataset and the imbalance between the categories of the target variable were also considered.

Distribution of the target variable

The initial evaluation of the target variable distribution, Hypertension, revealed a significant imbalance between its categories. The “High” (high risk) category comprised 125,781 records, which corresponded to 71.88% of the database. In contrast, the “Low” (low risk) category presented 49,201 records, equivalent to 28.12% of the total. This predominance of the “High” class was a crucial factor for interpreting model performance, as classifiers can exhibit apparently favorable results by prioritizing the majority category.

The presence of this imbalance between classes required a careful analysis of performance metrics. To mitigate the influence of the majority class, measures such as the Precision-Recall curve and the Average Precision (AP) metric were employed, which provide a more robust evaluation in imbalanced data scenarios. This approach allowed for a more precise understanding of the models’ ability to distinguish between the two hypertension risk categories, avoiding biased interpretations from simple accuracy.

Correlation analysis between clinical variables

To investigate the monotonic associations between the quantitative variables, a correlation matrix was calculated using Spearman’s coefficient. The obtained coefficients, for the most part, remained close to zero, indicating a low monotonic association between the evaluated variables. Attributes such as age, body mass index (BMI), systolic and diastolic blood pressure, cholesterol, glucose, salt intake, sleep duration, stress level, and heart rate showed small magnitude correlations among themselves.

The absence of high correlations suggested that the attributes did not present strong linear relationships among themselves. However, this observation did not imply that each variable contributed independently to the classification of hypertension. Tree-based algorithms, such as XGBoost, are capable of exploring non-linear relationships and complex interactions between variables, which are not adequately represented by a linear correlation matrix. Therefore, the Spearman analysis served as an exploratory tool to characterize the relationships between attributes, and not as an exclusive criterion for selecting or excluding variables from the model.

Descriptive analysis of the Dataset

The descriptive characteristics of the quantitative variables were evaluated to understand the data distribution. The average age of the individuals was 53.49 years, with a range between 18 and 89 years. The average body mass index (BMI) was 27.49 kg m⁻², with values ranging from 15.0 to 40.0 kg m⁻². The average systolic blood pressure was 134.51 mmHg, ranging from 90 to 179 mmHg, while the average diastolic blood pressure was 89.45 mmHg, with values between 60 and 119 mmHg.

Metabolic variables also exhibited considerable amplitude. The average total cholesterol was 224.54, LDL 129.53, HDL 64.57, triglycerides 149.57, and glucose 134.49. The standard deviations associated with these variables indicated heterogeneity in the evaluated records. Regarding behavioral variables, alcohol consumption showed an average of 14.99, stress level 5.01, salt intake 8.48, and sleep duration 7.00 hours. The average heart rate was 74.50 beats per minute. This diversity of values in the variables provided a rich basis for predictive analyses, although descriptive variability alone does not establish causality with hypertension.

Importance of variables in XGBoost

The importance attributed to the characteristics by the XGBoost model was used as a preliminary analysis of the relative contribution of the attributes in the classification process. The importance values showed a relatively close distribution among the various attributes, without significant concentration in a single variable. Among the characteristics that stood out with higher importance values, the presence of categories associated with smoking, gender, alcohol consumption, heart rate, sleep duration, and BMI was observed. Several clinical and metabolic variables also showed similar importance values among themselves.

This result suggested that the classification process did not depend exclusively on a single attribute, but rather on a combination of factors. However, the importance generated by the algorithm should be interpreted with caution, as it represents the relative contribution of the variables to the model’s structure and does not, by itself, inform whether a particular value increased or decreased the predicted probability. Furthermore, this metric should not be interpreted as evidence of causality or isolated clinical relevance, as highlighted by Bisong et al. (2024), who emphasized the dependence of performance and feature importance on the properties of the population and the analyzed dataset.

Decision curve analysis

Beyond the traditional classification metrics, the utility of the XGBoost model was examined at different decision thresholds through the decision curve. At lower and intermediate thresholds, the XGBoost curve remained practically overlapped with the strategy of considering all individuals as belonging to the positive class. This behavior indicated that, in this range, the model offered little additional net benefit when compared to the strategy of classifying all records as positive.

The proximity between the curves may be related to the predominance of the “High” category, which represented 71.88% of the records. As the threshold increased and approached approximately 0.70, the model’s net benefit progressively decreased, approaching zero. At higher thresholds, the model’s practical advantage became reduced. The decision curve complements the discriminative metrics, relating classifications to decision thresholds, being useful for evaluating net benefit in risk contexts, as pointed out by Effati et al. (2024). However, the observed results did not support a clinical superiority of the model across the analyzed range, indicating limited incremental utility in a relevant portion of the evaluated thresholds.

Performance evaluation by the Precision-Recall curve

Considering the imbalance of the target variable, the discriminative performance of the XGBoost model was evaluated by the Precision-Recall curve. The model presented an Average Precision (AP) of 0.715, a value that was slightly lower than the baseline reference of 0.719, associated with the distribution of the positive class. This result indicated a limited discriminative capacity of the model, suggesting that it had difficulties in simultaneously combining high levels of precision and sensitivity.

The Precision-Recall curve remained close to the level corresponding to the prevalence of the positive class for most “recall” values. This proximity between the model’s AP and the baseline classifier can be attributed to the imbalance between categories, the limited separability of classes, and the available features in the dataset. Qi et al. (2025) highlighted the importance of analyzing performance curves and different thresholds to avoid interpretations based solely on a single metric. Therefore, although XGBoost allowed for the exploration of attributes and the production of interpretable analyses, its overall discriminative performance remained restricted, which should be considered in any evaluation of its practical utility.

Interpretability of predictions by SHAP

The interpretation of the XGBoost predictions was deepened through the SHapley Additive exPlanations (SHAP) technique, which allowed measuring the individual contribution of variables to the model’s output. The distribution of SHAP values for the analyzed variables, such as age, cholesterol, glucose, systolic blood pressure, BMI, diastolic blood pressure, and heart rate, showed overlap of positive and negative values. This demonstrated that the contribution of an attribute did not occur uniformly for all observations, indicating that the relevance of a variable can vary depending on the individual record.

This behavior reinforced one of the advantages of the SHAP technique, which allows observing how a variable can have global relevance and, at the same time, contribute differently in individual records. Huang et al. (2025) highlighted the utility of SHAP for increasing the interpretability of Machine Learning models, allowing an evaluation of how different attributes participate in predictions. It is important to emphasize that SHAP values represent an explanation of the model’s behavior and not estimates of causal effects. Therefore, it was not possible to infer that a given variable caused an increase or reduction in risk based solely on these SHAP values.

Average contribution of the variables

To synthesize the global magnitude of the contributions obtained by the SHAP analysis, the average absolute values for the presented features were calculated. Age demonstrated the highest average absolute SHAP value, followed by cholesterol and glucose. Subsequently, systolic blood pressure, BMI, diastolic blood pressure, and heart rate were observed as variables with significant average contributions. This importance ranking partially differed from the importance ranking presented directly by XGBoost, as the two procedures represent distinct concepts of importance.

While XGBoost’s internal importance is associated with the use of variables in tree construction, the absolute average of SHAP values represents the average magnitude of the contribution of attributes in the evaluated predictions. The presence of age among the variables with the highest contribution is consistent with the knowledge that the occurrence of hypertension increases in older age groups. Cholesterol and glucose are also components of the metabolic profile relevant to the classifications, reinforcing the multifactorial nature of the problem, as pointed out by Islam et al. (2025). However, the absolute average importance does not indicate the direction of the association, only the average magnitude of participation in the predictions.

Impact of systolic blood pressure on prediction

The specific contribution of systolic blood pressure was examined through its relationship with the respective SHAP values, using age as an additional representation variable. The SHAP values of systolic blood pressure showed wide dispersion across practically the entire evaluated range, evidencing a non-strictly linear behavior. Positive values indicated a contribution in the direction defined as positive by the model, while negative values represented a contribution in the opposite direction, demonstrating the complexity of the interaction between the variables.

Between approximately 90 and 110 mmHg, a relatively high dispersion of SHAP values was observed. In the intermediate region, between approximately 120 and 150 mmHg, the contributions were concentrated closer to zero. In the upper ranges, especially above approximately 160 mmHg, greater dispersion occurred again, including positive and negative contributions. This pattern demonstrated that systolic pressure did not produce a uniform contribution to all observations, and its interpretation depended on the context formed by the other characteristics of the record, making it inadequate to attribute to systolic pressure, in isolation, the classification produced by the model.

The colors used to represent age also indicated the presence of older individuals at different levels of systolic pressure, corroborating the literature that points to a higher frequency of hypertension in older age groups (Zhao et al., 2025). However, the visualization was used to interpret the algorithm’s predictions and not to establish a causal relationship between age, blood pressure, and hypertension. This result demonstrated the SHAP’s ability to observe behaviors that would not be represented by a simple linear relationship, reinforcing that the predictions depended on the combination of multiple attributes.

XGBoost Probability Calibration

Beyond the discrimination between classes, the correspondence between the probabilities produced by XGBoost and the observed proportions was evaluated through calibration. The observed points showed greater proximity to the reference line in part of the intermediate probability range, while more pronounced differences were observed at the extremes. This behavior indicated that the calibration was not uniform across the entire distribution of predictions, which may impact the confidence in the predicted probabilities in different risk scenarios.

The Brier score obtained was 0.203 and the log loss was 0.596. These metrics quantified the error between the predicted probabilities and the observed outcomes, but their isolated values did not allow us to affirm that the model presented satisfactory calibration, with its interpretation being more robust when compared to reference models or validation on independent data. The instability identified in certain ranges may be related to the number of observations available in each interval, the imbalance of the target variable, and the model’s limited ability to separate the patterns corresponding to the two categories. Thus, the calibration analysis added an important dimension to the evaluation of XGBoost, but did not eliminate the limitation already evidenced by the Precision-Recall curve, confirming that the overall discriminative performance remained restricted.

Logistic Regression and odds ratios

Logistic Regression was employed as a comparative linear model, allowing for the evaluation of the direction and magnitude of the estimated associations between some quantitative variables and the outcome. The odds ratios (OR) remained very close to one for all variables presented. Triglycerides, diastolic blood pressure, salt intake, and HDL showed estimates slightly above one, while BMI, systolic blood pressure, LDL, age, and cholesterol showed slightly lower values. Heart rate, glucose, and sleep duration remained close to the reference value.

The confidence intervals represented in the graph crossed the value one for all variables, indicating high uncertainty in the estimates and the absence of evidence of robust linear associations between the attributes and the outcome. This result demonstrated the need for caution when interpreting OR values close to one, as the estimates did not support isolated clinical conclusions nor allowed attributing a protective or risk character to the variables based solely on the presented model. Regarding predictive performance, Logistic Regression showed an accuracy of 50.6% and an AUC-ROC of 0.49. These values indicated a performance close to chance, evidencing limited discriminative capacity, which may be related both to the characteristics of the dataset and to the limitation of a linear model to represent more complex relationships between the attributes.

Synthesis of results and limitations

Jointly, the developed models were capable of producing classifications and allowed the investigation of variable participation, but presented relevant performance limitations. XGBoost provided greater flexibility to analyze non-linear relationships and allowed interpretation by SHAP, while Logistic Regression exhibited low discriminative capacity, with an AUC-ROC of 0.49. The Precision-Recall analysis constituted one of the main limitations observed for XGBoost, as its Average Precision of 0.715 was lower than the reference of 0.719, not demonstrating a global discriminative advantage in relation to the reference used. Interpretability visualizations and the identification of relevant attributes for the algorithm’s decisions were achieved, but the predictive capacity still presented restrictions for practical use.

Other limitations of the study included the use of a single dataset, meaning that the results specifically reflected the characteristics of the analyzed database and were not subjected to external validation in an independent population. The decision not to apply treatment strategies for class imbalance, such as resampling or class weighting, preserved the original distribution of the database, but may have influenced the models’ ability to adequately distinguish the minority class. Furthermore, the absence of automated hyperparameter search and cross-validation limited the investigation into alternative model configurations. Thus, the models should be understood as experimental data analysis instruments, rather than validated clinical tools for diagnosis, screening, or decision-making in patients.

4. Conclusion

This study aimed to develop and evaluate supervised XGBoost and Logistic Regression models to classify the risk of arterial hypertension in adults, as well as to identify the variables with the greatest contribution to the predictions. It was found that both models were capable of producing classifications and allowed for the investigation of variable participation, although with limited predictive performance. XGBoost presented an Average Precision of 0.715, a value slightly lower than the baseline reference of 0.719, indicating restricted discriminative capacity. Logistic Regression, in turn, exhibited performance close to chance, with an AUC-ROC of 0.49. Interpretability analysis using the SHAP technique revealed that variables such as age, cholesterol, glucose, systolic and diastolic blood pressure, and body mass index (BMI) were the ones that contributed most to the predictions, reinforcing the multifactorial nature of hypertension. However, the significant imbalance of the target variable, with a predominance of the high-risk category, directly influenced the observed performance and the interpretation of the metrics.

The study’s main contribution lies in demonstrating the technical applicability of Machine Learning methods to explore factors associated with hypertension classification and in expanding the interpretability of predictions, providing a methodological basis for future investigations. However, the identified limitations, such as the use of a single set of secondary data without external validation, the absence of class balancing strategies, and the restriction of available variables, impacted the models’ predictive capacity. The lack of systematic hyperparameter optimization also limited the exploration of alternative configurations. Therefore, the developed models are understood as experimental data analysis instruments, not as validated clinical tools. For future studies, it is suggested to incorporate databases from different populations, include new clinical and behavioral variables, evaluate balancing strategies, and validate on independent datasets, aiming to improve generalization and practical utility in healthcare.

Bibliographic References

Adeleke, O.; Adebayo, S.; Aworinde, H.; Adeniyi, Α.Ε.; Aroba, O.J. 2024. Machine learning evaluation of a hypertension screening program in a university workforce over five years. Scientific Reports 14(1): 30255.

Bisong, E.; Jibril, N.; Premnath, P.; Buligwa, E.; Oboh, G.; Chukwuma, A. 2024. Predicting high blood pressure using machine learning models in low- and middle-income countries. BMC Medical Informatics and Decision Making 24(1): 234.

Bora, K.; Kalimuthusamy, N.; Gogoi, A.J.; Garh, N.; Rabidas, M.; Chanda, G.; Borah, P.K. 2025. A machine learning approach to predict hypertension using cross-sectional and two-year follow-up data from a health and demographic cohort of Assam, North East India. The Indian Journal of Medical Research 161(4): 394.

Chen e Guestrin, 2016 [Referência completa não encontrada no documento original]

Cheraghi, Z.; Doosti-Irani, M.; Sohrabi, M.; Doosti-Irani, A. 2025. Predicting factors associated with uncontrolled hypertension using machine learning methods: a cross-sectional study in Western Iran. International Journal of Hypertension 2025: 4011397.

Dey, D.; Haque, M.S.; Islam, M.M.; Aishi, U.I.; Shammy, S.S.; Mayen, M.S.A.; Uddin, M.J. 2025. The proper application of logistic regression model in complex survey data: a systematic review. BMC Medical Research Methodology 25(1): 15.

Effati, S.; Kamarzardi-Torghabe, A.; Azizi-Froutaghe, F.; Atighi, I.; Ghiasi-Hafez, S. 2024. Web application using machine learning to predict cardiovascular disease and hypertension in mine workers. Scientific Reports 14(1): 31662.

Felix, P.V.; Pereira, J.L.; Fisberg, R.M. 2024. Associations between four diet quality indexes and high blood pressure among adults: results from the 2015 Health Survey of São Paulo. Nutrients 16(5): 629.

Guo, K.; Ni, W.; Du, L.; Zhou, Y.; Cheng, L.; Zhou, H. 2024. Environmental chemical exposures and a machine learning-based model for predicting hypertension in NHANES 2003–2016. BMC Cardiovascular Disorders 24(1): 544.

Hua, Y.; Stead, T.S.; George, A.; Ganti, L. 2025. Clinical risk prediction with logistic regression: best practices, validation techniques, and applications in medical research. Academic Medicine & Surgery.

Huang, Q.; Jiang, Z.; Shi, B.; Meng, J.; Shu, L.; Hu, F.; Mi, J. 2025. Characterisation of cardiovascular disease incidence and machine learning risk prediction in middle-aged and elderly populations: data from the China Health and Retirement Longitudinal Study. BMC Public Health 25(1): 518.

Islam, H.; Iqbal, M.S.; Hossain, M.M. 2025. Blood pressure abnormality detection and interpretation utilizing explainable artificial intelligence. Intelligent Medicine 5(1): 54-65.

Knox, S.W. 2026. Machine Learning: A Concise Introduction. John Wiley & Sons, Hoboken, NJ, USA.

Martínez-García et al., 2025 [Referência completa não encontrada no documento original]

Mroz, T.; Griffin, M.; Cartabuke, R.; Laffin, L.; Russo-Alvarez, G.; Thomas, G.; Habboub, G. 2024. Predicting hypertension control using machine learning. PLOS ONE 19(3): e0299932.

Pandey, S.; Pandey, A.; Neupane, A.; Subedi, D.; Guragain, A. 2025. Development of a hypertension risk prediction model using nationally representative survey data: a machine learning approach and web application deployment. medRxiv. Disponível em: https://www.medrxiv.org/content/10.1101/2025.08.30.25334758v1. Acesso em: 08 set. 2025.

Qi, X.; Wang, S.; Fang, C.; Jia, J.; Lin, L.; Yuan, T. 2025. Machine learning and SHAP value interpretation for predicting comorbidity of cardiovascular disease and cancer with dietary antioxidants. Redox Biology 79: 103470.

Springer, 2024 [Referência completa não encontrada no documento original]

Zhao, L.; Zheng, L.; Wang, R.; Gong, X.; Wu, Y.; Han, S.; Xue, M. 2025. Association between triglyceride glucose combined with body mass index and hypertension in the NHANES 2017 to 2020. Scientific Reports 15(1): 9092.

Zhuo, H.; Li, T.; Lu, W.; Zhang, Q.; Ji, L.; Li, J. 2025. Prediction model for spontaneous combustion temperature of coal based on PSO-XGBoost algorithm. Scientific Reports 15(1): 2752.

Article originating from the Final Course Project of the Specialization in Data Science and Analytics of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform