Data driven approach to characterize rapid decline in autosomal dominant polycystic kidney disease

Abstract

Autosomal dominant polycystic kidney disease (ADPKD) is a genetic kidney disease with high phenotypic variability. Furthering insights into patients' ADPKD progression could lead to earlier detection, management, and alter the course to end stage kidney disease (ESKD). We sought to identify patients with rapid decline (RD) in kidney function and to determine clinical factors associated with RD using a data-driven approach. A retrospective cohort study was performed among patients with incident ADPKD (1/1/2002-12/31/2018). Latent class mixed models were used to identify RD patients using differences in eGFR trajectories over time. Predictors of RD were selected based on agreements among feature selection methods, including logistic, regularized, and random forest modeling. The final model was built on the selected predictors and clinically relevant covariates. Among 1,744 patients with incident ADPKD, 125 (7%) were identified as RD. Feature selection included 42 clinical measurements for adaptation with multiple imputations; mean (SD) eGFR was 85.2 (47.3) and 72.9 (34.4) in the RD and non-RD groups, respectively. Multiple imputed datasets identified variables as important features to distinguish RD and non-RD groups with the final prediction model determined as a balance between area under the curve (AUC) and clinical relevance which included 6 predictors: age, sex, hypertension, cerebrovascular disease, hemoglobin, and proteinuria. Results showed 72%-sensitivity, 70%-specificity, 70%-accuracy, and 0.77-AUC in identifying RD. 5-year ESKD rates were 38% and 7% among RD and non-RD groups, respectively. Using real-world routine clinical data among patients with incident ADPKD, we observed that six variables highly predicted RD in kidney function.