Derivation and Validation of a Machine Learning Model for the Prevention of Unplanned Dialysis.

Abstract

Approximately half of all patients with advanced chronic kidney disease (CKD) who progress to kidney failure initiate dialysis in an unplanned fashion which is associated with high morbidity, mortality, and healthcare costs. A novel prediction model designed to identify advanced CKD patients who are at high risk for developing kidney failure over short time frames (6-12 months) may help reduce the rates of unplanned dialysis and improve the quality of transitions from CKD to kidney failure. We performed a retrospective study employing machine learning random forest algorithms incorporating routinely collected age and sex data along with time-varying trends in laboratory measurements to derive and validate six- and 12-month kidney failure risk prediction models in the advanced CKD population. The models were comprehensively characterized in three independent cohorts in Ontario, Canada - derived in a cohort of 1,849 consecutive advanced CKD patients (mean [standard deviation] age 66 [15] years, eGFR 19 [7] mL/min/1.73m2), and validated in two external advanced CKD cohorts (n=1,356; age 69 [14] years, eGFR 22 [7] mL/min/1.73m2). Across all cohorts, 55% of patients experienced kidney failure, of which 35% involved unplanned dialysis. The six- and 12-month models demonstrated excellent discrimination with area under the receiver operating characteristic curve of 0.88 (95%CI: 0.87-0.89) and 0.87 (95%CI: 0.86-0.87) along with high probabilistic accuracy with Brier scores of 0.10 (95%CI 0.09-0.10) and 0.14 (95%CI 0.13-0.14), respectively. The models were also well-calibrated and delivered timely alerts on a significant number of patients who ultimately initiated dialysis in an unplanned fashion. Similar results were found upon external validation testing. These machine-learning models using routinely collected patient data accurately predict near-future kidney failure risk among the advanced CKD population, and retrospectively deliver advanced warning on a substantial proportion of unplanned dialysis events. Optimal implementation strategies still need to be elucidated.