Predictive Models for Kidney Recovery and Death in Patients Continuing Dialysis as Outpatients after Starting in Hospital

BACKGROUND: For patients who initiate dialysis during a hospital admission and continue to require dialysis after discharge, outpatient dialysis management could be improved by better understanding the future likelihood of recovery to dialysis independence and the competing risk of death. METHODS: We derived and validated linked models to predict the subsequent recovery to dialysis independence and death within 1 year of hospital discharge using a population-based cohort of 7657 patients in Ontario, Canada. Predictive variables included age, comorbidities, length of hospital admission, intensive care status, discharge disposition, and prehospital admission eGFR and random urine albumin-to-creatinine ratio. Models were externally validated in 1503 contemporaneous patients from Alberta, Canada. Both models were created using proportional hazards survival analysis, with the "Recovery Model" using Fine-Gray methods. Probabilities generated from both models were used to develop 16 distinct "Recovery and Death in Outpatients" (ReDO) risk groups. RESULTS: ReDO risk groups in the derivation group had significantly distinct 1-year probabilities for recovery to dialysis independence (first quartile: 10% [95% confidence interval (CI), 9% to 11%]; fourth quartile: 73% [70% to 77%]) and for death (first quartile: 12% [11% to 13%]; fourth quartile: 46% [43% to 50%]). In the validation group, model discrimination was modest (c-statistics [95% CI] for recovery and for death quartiles were 0.70 [0.67 to 0.73] and 0.66 [0.62 to 0.69], respectively), but calibration was excellent (integrated calibration index [95% CI] was 7% [5% to 9%] and 4% [2% to 6%] for recovery and death, respectively). CONCLUSIONS: The ReDO models generated accurate expected probabilities of recovery to dialysis independence and death in patients who continued outpatient dialysis after initiating dialysis in hospital. An online tool on the basis of the models is available at https://qxmd.com/calculate/calculator_874 .