Evaluation of Risk Factors Associated With Antihypertensive Treatment Success Employing Data Mining Techniques

Objective: This study aimed to evaluate the effects of potential risk factors on antihypertensive treatment success. Methods: Patients with hypertension who were treated with antihypertensive medications were included in this study. Data from the last visit were analyzed retrospectively for each patient. To evaluate the predictive models for antihypertensive treatment success, data mining algorithms (logistic regression, decision tree, random forest, and artificial neural network) using 5-fold cross-validation were applied. Additionally, study parameters between patients with controlled and uncontrolled hypertension were statistically compared and multiple regression analyses were conducted for secondary endpoints. Results: The data of 592 patients were included in the analysis. The overall blood pressure control rate was 44%. The performance of random forest algorithm (accuracy = 97.46%, precision = 97.08%, F1 score = 97.04%) was slightly higher than other data mining algorithms including logistic regression (accuracy = 87.31%, precision = 86.21%, F1 score = 85.74%), decision tree (accuracy = 76.94%, precision = 70.64%, F1 score = 76.54%), and artificial neural network (accuracy = 86.47%, precision = 83.85%, F1 score = 84.86%). The top 5 important categorical variables (predictive correlation value) contributed the most to the prediction of antihypertensive treatment success were use of calcium channel blocker (−0.18), number of antihypertensive medications (0.18), female gender (0.10), alcohol use (−0.09) and attendance at regular follow up visits (0.09), respectively. The top 5 numerical variables contributed the most to the prediction of antihypertensive treatment success were blood urea nitrogen (−0.12), glucose (−0.12), hemoglobin A1c (−0.12), uric acid (−0.09) and creatinine (−0.07), respectively. According to the decision tree model; age, gender, regular attendance at follow-up visits, and diabetes status were identified as the most critical patterns for stratifying the patients. Conclusion: Data mining algorithms have the potential to produce predictive models for screening the antihypertensive treatment success. Further research on larger populations and longitudinal datasets are required to improve the models.