How much better can we predict dialysis patient survival using clinical data?

Objective. To use three approaches to compare dialysis survival prediction based on variables included in the Standardized Mortality Ratio (SMR) with prediction based on a clinically enriched set of variables. Data Source. The United States Renal Data System Case Mix Severity data set containing demographic, clinical, functional, nutritional, and treatment d etails about a random sample of 4,797 adult dialysis patients from 291 trea tment units, incident to dialysis in 1986 and 1987. Study Design. This observational study uses baseline patient characteristic s in two proportional hazards survival models: the BASE model incorporates age, race, sex, and cause of end-stage renal disease (ESRD); the FULL model includes these and additional clinical information. We compare each model' s performance using (1) the c-index, (2) observed median survival in strata of predicted risk, and (3) predicted survival for patients with different characteristics. Principal Findings. The FULL model's c-index (0.709, 0.708-0.711) is signif icantly higher than that of the BASE model (0.675, 0.675-0.676), indicating better discrimination. Second, the sickest patients identified by the FULL model were in fact sicker than those identified as sickest by the BASE mod el, with observed median survival of 451 days versus 524. Third, survival p redictions for sickest patients using the FULL model are one-third shorter than those based on the BASE model. Conclusions, The model. with more detailed clinical information predicted s urvival better than the BASE model. Clinical characteristics enable more ac curate predictions, particularly for the sickest patients. Thus, clinical c haracteristics should be considered when making quality assessments for dia lysis patients.