Bayesian network and artificial intelligence to predict cardiovascular events in chronic kidney disease patients = Uso de inteligencia artificial y estadística bayesiana para predecir un evento cardiovascular en pacientes con enfermedad renal crónica

Background: Current scores do not adequately predict cardiovascular risk in patients with chronic kidney disease who are at a very high CV risk in short and medium term. Aim: The aim of our analysis was to create a Bayesian network to predict the 2-year occurrence of a cardiovascular event in patients with chronic kidney disease. Methods and results: The data originated from the observational and prospective Photo-Graphe V3 cohort. Sixty-two nephrologists in 20 French regions included 1144 non-dialysed patients with chronic kidney disease. Seven hundred and thirty patients with known medical status at 2 years were analysed. An initial Bayesian model was first built using 26 variables related to the characteristics of the patients, their medical background, and treatments. A cardiovascular event (heart failure, acute coronary syndrome, transient ischemic attack, stroke or cardiovascular death) occurred in 20.0% of the patients after two years of follow-up. The model was first optimized using synthetic data (created from the original database) to increase its reliability. The number of variables was then reduced using the 13 most informative variables to increase its clinical applicability. The10-fold cross validation showed that the optimized clinical model with 13 variables had an area under the ROC curve of 0.90+0.02, a sensitivity of 82.5+5.9%, a specificity of 80.6+5.1%, a predictive positive value of 81.2+3.4% and a negative predictive value of 82.5+4.4%. The percentage of misclassified subjects was 18.4+2.6%. Conclusion: Using artificial intelligence methods, a new clinical tool to predict cardiovascular events in patients with chronic kidney disease is proposed