O38 A SOCIAL DETERMINANT AND SIMPLE CLINICAL VARIABLES AS THE BEST PREDICTORS OF ESRD BY AN XGBOOST MACHINE LEARNING MODEL IN CKD

Background and Objective: Chronic Kidney Disease (CKD) is a complex disease that affects approximately 13% of the world's population and is growing rapidly due to the increase in aging and the prevalence of type 2 diabetes mellitus, obesity and hypertension. It is difficult to predict who will develop End-Stage Renal Disease (ESRD) and, so far, no prediction model is applied in clinical practice. Our aim was to develop and externally validate an ESRD prediction model with potential utility in clinical decision making. Methods: A longitudinal study was conducted with an 8-year follow-up of sociodemographic, biological, and clinical variables. Individuals with a diagnosis of CKD belonging to the cardiovascular risk program were recruited in two cohorts for derivation and external validation of the model. Classical and machine learning models on an imputed data matrix and sub-databases were explored by bootstrapping and cross-validation, as well as the change in GFR between measurements that predicted ESRD. Measures of calibration, discrimination, and overall predictive performance were presented. Models were trained and tested in sub cohorts and the best performing model was validated in an independent cohort and its uncertainty established. Results: A total of 2514 patients were included, 1650 in the derivation cohort, mean age of 74 years, 40.4% were women. A machine learning xgboost model identified previous GFR, creatinine, blood glucose, HbA1c, HDL cholesterol, BUN and income level as the main predictors of ESRD. Likewise, an inter-measurement GFR delta of 3.09 ml/kg/min was associated with a significant increase of ESRD. In the external validation cohort, with 864 patients, the model maintained an accuracy of 0.98, an AUC of 0.776, a recall of 1.00 and an F1-score of 0.99. The uncertainty of the AUC given by the 95%CI was 0.68-0.826. Conclusions: A machine learning xgboost model identified the most important predictors of ESRD, maintaining adequate predictive performance when cross-validated in an independent cohort of patients. The model could be applied in real clinical practice.