Bootcongres

Introduction. Our recently published prediction model on renal patient survival (Hemke et al, BMC Nephrol 2013) was based on 4 variables: age of the patient at start of renal replacement therapy (RRT), primary renal disease (PRD), treatment modality at 90 days after start of RRT, and sex. Although this prediction model showed good calibration and reasonable discrimination, we aimed to improve this model by including clinical data. Methods. We analyzed data from 1837 patients registered in the NECOSAD-study. For these patients we selected 3 additional sets of potential predictors based on availability and objectivity of the data. The first set were easily available medical history and clinical data like BMI, Charlson co-morbidity score, smoking, and malignancies. The second set consisted of laboratory values like cholesterol, phosphate, and albumin. The last set contained the variables that are least readily available: GFR and Kt/V. Multiple imputation was used to deal with missing values. For model development we used multivariate Cox regression analyses with backward selection in one part of the patient group (N=1225). We first recalibrated the original model, and subsequently added the 3 sets to the prior variables. The resulting models were validated on both calibration and discrimination in the validation group (N=612). The C-index is a measure for discrimination; a C-index of 0.5 (the reference point) indicates no discrimination, while a model with a C-index of 1.0 has perfect discriminative power. Results. The performance of the original model in the NECOSAD cohort, measured by calibration and discrimination, was adequate. Addition of the easily available variables resulted in a prediction model for 10 year patient survival, based on 10 predictors. The calibration was sufficient and the discrimination measured by the C-index improved from 0.724 to 0.778, a 27% improvement based on the reference C-index. Addition of the laboratory values resulted in an alternative prediction model based on 14 predictors. The calibration was sufficient but the C-index was only slightly higher at 0.781. Adding GFR and Kt/V didn’t result in a different model. Conclusion. We conclude that the performance of the original model using only 4 prediction parameters (age, PRD, therapy and sex) is sufficient also for the NECOSAD cohort. The discriminative ability of the prediction model can be improved by adding easily available medical history and clinical parameters to the model.