Abstract
OBJECTIVE: To develop new risk prediction equations for kidney failure in patients with chronic kidney disease who do not require data for the urine albumin to creatinine ratio. DESIGN: Retrospective cohort study. SETTING: Stockholm Creatinine Measurements (SCREAM) database of routinely collected electronic healthcare records in primary and outpatient care from the region of Stockholm, Sweden. PARTICIPANTS: 116 158 adults with chronic kidney disease stages 3-4, defined by two estimated glomerular filtration rate (eGFR) results of <60 to ≥15 mL/min/1.73 m², at least 90 days apart, with no intermediate eGFR value ≥60 mL/min/1.73 m², between 1 January 2010 to 31 December 2018. MAIN OUTCOME MEASURE: Kidney failure, defined as starting kidney replacement therapy, recorded within five years of the index date. RESULTS: Based on temporal split sample validation, development and validation cohorts included 85 012 patients (736 kidney replacement therapy events) and 28 338 patients (114 kidney replacement therapy events), respectively. After Cox regression with automated backwards selection, the final model included 10 predictors (in order of significance): eGFR, age, diabetes, sex, atrial fibrillation, antihypertensive drugs, peripheral artery disease, reduction in eGFR slope, acute kidney injury, and hypertension. Model discrimination was excellent in both the development cohort (C statistic 0.941, 95% confidence interval (CI) 0.932 to 0.951) and validation cohort (C statistic 0.944, 0.923 to 0.965). In 26 229 patients with data for the urine albumin to creatinine ratio, the four variable kidney failure risk equation (KFRE) showed marginal improvement in discrimination over our new equation (C statistic 0.950, 95% CI 0.942 to 0.958 for KFRE v 0.926, 0.915 to 0.936, for the new equation). KFRE under-estimated the risk in the study cohort, however, with an observed-expected event probability ratio of 2.11, suggesting that recalibration is required. CONCLUSIONS: The findings of the study indicate that predicting the risk of kidney failure with high accuracy in a general population of patients with chronic kidney disease is possible based on data that are routinely available, without requiring data for the urine albumin to creatinine ratio.