SO044
PREDICTION MODELS FOR THE MORTALITY RISK IN CHRONIC DIALYSIS PATIENTS: A SYSTEMATIC REVIEW AND INDEPENDENT EXTERNAL VALIDATION STUDY

Introduction and Aims: A risk stratification model that accurately identifies high-risk patients could help in clinical decision making, informing patients on their prognosis and risk stratification in research context. Various models to predict mortality in dialysis patients have been published. However these have rarely been validated and none are used in clinical practice. A direct comparison of the predictive performance of existing models could help bridge the gap between the development of risk prediction models and their clinical application. The aim of the current study was to identify existing models for predicting the risk of death in dialysis patients through a systematic review, in order to subsequently validate these models externally in the same large independent cohort.

Methods: A systematic review was performed following the PRISMA guidelines. Titles, abstracts and articles were screened by two independent researchers following pre-defined inclusion criteria. The original prediction models with the regression coefficients per predictor and intercept or baseline hazard were extracted from the selected studies and used to evaluate the predictive performance within NECOSAD, a prospective cohort of incident dialysis patients. To account for missing data on predictors, multiple imputation was performed using the fully conditional specification. The probability of death was calculated for each individual and the predictive performance of the models was assessed based on their discrimination and calibration.

Results: In total 15 articles were included in the systematic review. Many studies used specific predictors not routinely collected and only 60% of the studies provided the full prediction formula. External validation was performed in 1943 dialysis patients from NECOSAD for a total of 7 models. Table 1 shows the discrimination in which the C-statistics are presented as the median and interquartile range from the pooled imputation results. Figure 1 shows the calibration plots per model, for Wagner's model calibration could not be determined due to a missing baseline hazard value.

Conclusions: Overall, the performance of the models was poorer in the external validation than in the original ones. Of the validated models, the ones proposed by Floege may be the best suited for predicting mortality in dialysis patients over a time-frame of 1 or 2 years. This study is a step forward in the use of a model with which personalized information on prognosis can be given to dialysis patients, aiding patient-centred decision making.