Prediction of Cardiovascular Events in Hypertensive Subjects With Artificial Intelligence Tools

Objective: Cardiovascular diseases contribute substantially to morbi-mortality at the global level. To effectively address the increasing incidence of cardiovascular events, advanced preventive strategies are needed. Predictive modeling is a crucial tool to identify individuals at risk and facilitate proactive interventions to improve clinical outcomes. This study develops and evaluates a model for predicting cardiovascular events (ischemic heart disease and stroke in previously event-free hypertensive subjects) by integrating clinical data, biomarkers and established risk factors. Most importantly, it predicts the likelihood of cardiovascular events after the initial patient visit. Design and method: The first step we have performed, enrich the dataset by creating variables using statistical measures, Principal Component Analysis (PCA) and Latent Class Analysis (LCA). Three different prediction algorithms are applied: XGBoost, Random Forest and Logistic Regression. The models were trained with 70% of the dataset and 30% for validation. Each algorithm maintains a balance between predictive accuracy and interpretability. The models predict cardiovascular events in patients with no history of cardiovascular disease after their baseline visit. Results: A total of 4,145 consecutive patients were referred to a hospital’s hypertension unit. The population’s mean age is 55.9±14, with 51% women. The average follow-up time is 3,028 days, yielding an incidence density of 1.93 (95% CI: 1.78-2.09) per 100 patient-years. The XGBoost model achieves 85% accuracy, 87% sensitivity, and 71% specificity. When reducing the model to the 30 most influential variables, there is a modest decrease in predictive accuracy (80%). Logistic Regression provides an interpretable model with an accuracy of 80%, sensitivity of 56%, and specificity of 84%. Meanwhile, Random Forest exhibits a high accuracy of 86%, with notably lower sensitivity at 13%. Conclusions: Our findings underscore the resilience and stability of the XGBoost model, effectively balancing predictive accuracy and interpretability. The inclusion of the Logistic Regression model, with its interpretative capacity, remains crucial, especially when transparency in model decisions is paramount. This study emphasizes the importance of tailoring predictive models to the specific requirements of applications. Future research could explore possible techniques for the interpretability of the XGBoost model.