Using random forests to understand unrecognized progression to late-stage CKD, a case-control study

Abstract Background and objectives Patients with undiagnosed CKD are at increased risk of suboptimal dialysis initiation and therefore reduced access to home dialysis and transplantation as well as poor outcomes. Improved understanding of how patients remain undiagnosed is important to determine better intervention strategies. Design, setting, participants, and measurements A retrospective, matched, case-control analysis of 1,535,053 patients was performed to identify factors differentiating 4 patient types: unrecognized late-stage CKD, recognized late-stage CKD, early-stage CKD and a control group without CKD. The sample included patients with commercial insurance, Medicare Advantage, and Medicare fee-for service coverage. Patient demographics, comorbidities, health care utilization, and prescription use were analyzed using random forests to determine the most salient features discriminating the types. Models were built using all four types, as well as pairwise for each type versus the unrecognized late-stage type. Results Area under the curve for the three pairwise models (unrecognized late-stage vs recognized late-stage; unrecognized late-stage vs early-stage; unrecognized late-stage vs no CKD) were 82%, 68% and 82%. Conclusions The lower performance of the unrecognized late-stage vs early-stage model indicates a greater similarity of these two patient groups. The unrecognized late-stage CKD group is not simply avoiding or unable to get care in a manner distinguishable from the early-stage group. New outreach for CKD to undiagnosed or undetected, insured patients should look more closely at patient sets that are like diagnosed early-stage CKD patients.