Cardiovascular Risk Estimation in Colombia Using Artificial Intelligence Techniques

Introduction: There is no information on the potential of machine learning (ML)-based techniques to improve cardiovascular risk estimation in the Colombian population. This article presents innovative models using five artificial intelligence techniques: neural networks, decision trees, support vector machines, random forests, and Gaussian Bayesian networks. Methods: The research is based on a cohort of 847 patients free of cardiovascular disease at baseline and followed for cardiovascular disease events over 10 years at the Central Military Hospital in Bogotá, Colombia. To enhance the robustness and reduce the risk of overfitting, model evaluation was conducted using a 5-fold cross-validation on the entire dataset. Discriminatory ability was evaluated with the area under a ROC curve (AUC-ROC) for each ML-based model and the Framingham model. Results: Experimental results showed that the neural network technique had the best discriminative ability to predict cardiovascular events, with an AUC-ROC of 0.69 (CI 95% 0.622-0.759) for unbalanced data and 0.67 (CI 95% 0.601-0.754) for balanced data. Other ML techniques also showed good discriminatory ability with AUC-ROC values between 0.56 and 0.65, superior to that observed for the Framingham model (0.53; CI 95% 0.468-0.607). Conclusion: Our study supports the flexible ML approaches to cardiovascular risk prediction as a way forward for cardiovascular risk assessment in Colombia. Our data even suggest that risk prediction using these techniques could be even more discriminative than widely used risk-stimulation models such as Framingham, adapted to the Colombian population. However, new prospective studies need to validate our data before general implementation.