Challenges in Predicting Estimated Glomerular Filtration Rate Slope Among Adult Patients With IgA Nephropathy: A Population-Based Cohort Study
Bryce Barr, Mark Canney, Navdeep Tangri, Lisa M. LixBackground
Traditional epidemiologic studies and clinical trials in IgA nephropathy (IgAN) have focused on kidney failure or large declines in estimated glomerular filtration rate (eGFR) as primary outcomes. Two-year eGFR slope has emerged as a validated surrogate endpoint and a means of earlier risk stratification, yet no tools exist to predict eGFR slope in patients with IgAN.
Objective
To develop and internally validate a prediction model for 2-year eGFR slope using routinely available clinical and pathological data in adults with IgAN.
Design
Adults with biopsy-proven IgAN (2002–2021) were identified from the Manitoba Glomerular Diseases Registry and linked to population-based laboratory and administrative health data. Mixed-effects linear regression models containing demographic, clinical, and comorbidity covariates with random intercepts and slopes predicted eGFR slope. Four pre-specified models of increasing complexity were tested. Model performance was assessed using marginal and conditional R 2 , root mean squared error, calibration, and Bland–Altman analyses. Internal validation was performed using 10-fold cross-validation.
Results
The cohort included 181 patients (median age 41 years [interquartile range (IQR) 31–54]); 137 (75.7%) cases with complete data were used in the primary analysis. Patients were high risk, with median eGFR 58 mL/min/1.73 m 2 (IQR 35–86) and proteinuria 1.73 g/day (IQR 1.06–2.95) at biopsy. Higher index proteinuria was independently associated with faster eGFR decline (–1.47 mL/min/1.73 m 2 /year per 1 g/day, SE 0.66; p=0.009). A parsimonious model including time, age, sex, index eGFR, index proteinuria, and two-way interactions of time with eGFR and proteinuria had excellent fit (conditional R 2 0.95). Conditional (i.e., patient-specific) slope estimates, which incorporated patient-level random effects estimated from longitudinal eGFR measurements, closely approximated observed trajectories, with 88% of predicted slopes within 1 mL/min/1.73 m 2 /year of observed values. In contrast, marginal (i.e., population-average) predictions were poorly calibrated (11% within 1 mL/min/1.73 m 2 /year; calibration slope 1.58, 95% CI 0.71–2.44). Cross-validation demonstrated marked deterioration in conditional performance, while marginal predictions were stable but inaccurate.
Limitations
Model performance was strongest when incorporating follow-up eGFR data obtained after the index date, which limits applicability for prediction in newly diagnosed patients. Sample size was constrained by missing data, and Oxford M and E scores were not available for all patients.
Conclusions
While mixed-effects linear regression models accurately characterized observed eGFR trajectories, predictive performance relied on patient-specific random effects and may not generalize to new patients. Baseline clinical and pathologic variables could not accurately predict eGFR slope. These findings highlight both the promise and challenges of slope-based prediction in IgAN and underscore the need for larger cohorts with multiple eGFR measurements, and models that leverage prior patient data to improve out-of-sample performance.