OBJECTIVES: Estimated glomerular filtration rate (eGFR) is a measure of kidney function commonly used as an outcome in clinical trials and real-world studies. Estimating eGFR slope can be challenging due to high measurement variability. Although linear models are commonly used, non-linear models have recently gained attention by providing more accurate estimates for such complex outcomes. We therefore aimed to compare statistical performance of linear and non-parametric models in estimating eGFR slope in real-world data.

METHODS: Using Clinical Practice Research Datalink AURUM data (2010-2019), adult CKD patients (≥18 years) with ≥2 serum creatinine measurements <60 ml/min/1.73m² were analyzed. The index date was the second serum creatinine measurement used to calculate eGFR. Patients with ≥4 eGFR results within the10 year study period with at least one year between the first and last tests, were included. eGFR slope was estimated using linear regression, Theil–Sen/median-based linear regression (MBLM), quantile regression with Tau = 0.5, generalized additive models (GAM), and Locally Weighted Scatter-plot Smoother (LOESS) using nonparametric kernel regression with linear regression.

RESULTS: Using a random sample of 100,000 patients with CKD (~20% of the cohort), with 1.3 million eGFR measures in the analysis, mean and standard deviation (SD) of number of eGFR records per patient was 21.60 (19.05). The LOESS model (span=0.75) provided the highest accuracy: median Efron R² (25th-75th quantiles) was 0.78 (0.54-0.99), compared to the linear model: 0.26 (0.07-0.53), while GAM is 0.43 (0.18-0.69), MBLM is 0.07 (0-0.39), quantile regression is 0.15 (0-0.46). The mean (SD) of root mean square error was 2.74 (2.38) for LOESS, versus 4.96 (2.71) for linear, GAM is 4.36 (2.54), MBLM is 6.00 (7.43), quantile regression is 5.34 (2.95).

CONCLUSIONS: In CPRD Aurum, the LOESS model demonstrated high accuracy in eGFR slope estimation in CKD patients with ≥4 eGFR measurements, outperforming linear and extended linear methods.