Incorporating Social Determinants of Health in the Prediction of Chronic Kidney Disease Progression in a National Cohort of US Veterans
Author(s): Karandeep Singh, Jennifer Bragg-Gresham, Yun Han, Brenda Gillespie, William Weitzel, Susan Crowley, Rajiv Saran
Background: As new therapies such as Sodium-Glucose Cotransporter-2 (SGLT2) inhibitors become available that may be effective in earlier stages of chronic kidney disease (CKD), newer risk prediction tools are required earlier in the development of CKD, and also to evaluate whether social determinants of health play a role in addition to individual-level risk factors.
Methods: We examined CKD progression among US Veterans who had known CKD in the VA health system using data from 2006-2016. CKD progression was defined based on two separate outcomes: 1) rapid CKD progression based on the eGFR slope < -3.7 mL/min/1.73m2 as a binary outcome, and 2) time-to-end stage kidney disease (ESKD) as a survival outcome. Veterans whose eGFR values were declining more steeply than -3.7 per year were considered “fast progressors,” representing 9.8% of the overall cohort. ESKD was identified by linking the VA data with US Renal Data System (USRDS) data, a national ESKD registry. After randomly dividing the dataset into a training, tuning, and testing set, tree ensemble models were trained and evaluated.
Results: We identified 1,550,526 patients meeting inclusion criteria, of which 930,615 patients were assigned to a training cohort, 309,831 to a tuning cohort, and 310,044 to a testing cohort. Tree ensemble models predicted fast progression with a C-statistic of 0.79 and time to ESKD with a C-statistic of 0.90. Baseline eGFR was the most important variable in predicting both outcomes, though social determinants constituted more of the important variables in rapid progression.
Conclusions: CKD progression can be accurately predicted, though the predictors differ for fast progression and ESKD onset.