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