#4784 IDENTIFICATION OF FACTORS ASSOCIATED WITH DEATH IN DIALYSIS PATIENTS USING A MACHINE LEARNING-BASED PREDICTIVE MODEL

Abstract Background and Aims Few studies including unplanned dialysis starts have used machine learning for the prediction of death in patients. The objective study was to use R software algorithms develop predictive models identification death-related factors patients undergoing hemodialysis (HD) peritoneal (PD). Method This included adult HD PD started a planned or urgent manner center between January 2014 2019. Epidemiological, clinical laboratory data were collected. Univariate analysis followed by ML-based analyses. Then, multivariate regressions obtained using stepwise Cox regression Finally, Random Forest model generated after variables with missing values >30% removed. Results Of 581 included, 170 died (29,2%). On univariate associated age, number comorbidities, modality switching, creatinine, PTH albumin at initiation, presence diabetes (DM), hospitalization, function recovery central venous catheter (CVC) access. Patients who CVC had worse survival (p = 0.0034) than those did not CVC, AVF, received (Figure 1). Data split into 20% testing model, 80% training model. preprocessing imputing some bag impute (decision trees), creating dummy variables, removing collinear variables. Death older age < 0.001), fewer ESI-free months 0.001) lower initial creatinine 0.008) (Table C-index 0.8099. forest ranked following descending order importance: months; age; levels albumin; comorbidities; dialysis-related infection; phosphorus hemoglobin; hospitalizations; male gender; switching 2). agreement 0.8110. Conclusion months, on both