Development and Validation of a Canadian Prediction Equation for Incident CKD Using Population-Based, Administrative Data

Background: Identifying individuals at risk for incident chronic kidney disease (CKD; estimated glomerular filtration rate [eGFR] <60 mL/min/1.73 m 2 ) could aid in prevention and disease surveillance. Objective: Develop and validate prediction equations to identify individuals at risk of incident CKD using routinely collected administrative data with and without urine albumin-to-creatinine ratio (ACR). Design: This is a retrospective cohort study using administrative data. Setting: This study was conducted in Manitoba and Ontario, Canada. Patients: This study included 413 948 adults (18 or older) with an eGFR > 70 mL/min/1.73 m 2 from Manitoba (derivation cohort; 2006-2016) with external validation in 7 747 513 adults from Ontario, Canada. Measurements: Routinely available variables (demographics, comorbidities, laboratory values) in administrative data sets were used to predict the outcome of incident CKD (stage G3+) defined by a single outpatient eGFR measure <60 mL/min/1.73 m 2 during and up to 10 years of follow-up. In an additional analysis, we defined incident CKD using repeat eGFR measures. Methods: Time-to-event models, accounting for the competing risk of death, were used to predict new-onset CKD from one to nine years with a data-driven model reduction. Prediction equations stratifying individuals with and without ACR measurements were derived internally and externally validated. Results: Among individuals from Manitoba [53% women, mean (SD) age 51 (17), mean (SD) baseline eGFR 95 (14) mL/min/1.73 m 2 , median (interquartile range) ACR 0.7 mg/mmol (1-3)], incident CKD occurred in 11.4% during a median follow-up time of 4.5 (Q1 = 2.3, Q3 = 7.6) years of follow-up. The final model included six variables (age, sex, baseline eGFR, hemoglobin, hypertension, and diabetes) and yielded a five-year area under the curve of 86.0 (no ACR) and 80.2 (with ACR). Model performance was excellent in external validation. Limitations: Only individuals with measures of all model predictors (complete case analysis) were included. Conclusion: Equations using routinely collected population-level, administrative data variables can accurately predict the onset of CKD with or without ACR.