Validation of the Klinrisk Machine Learning Model for CKD Progression in a Large Representative US Population.

Abstract

BACKGROUND Early identification of high-risk chronic kidney disease (CKD) can facilitate optimal medical management and improve outcomes. We aimed to validate the Klinrisk machine learning model for prediction of CKD progression in large US commercial, Medicare, and Medicaid populations. METHODS We developed three cohorts, consisting of insured adults enrolled in a) commercial, b) Medicare, and c) Medicaid plans between 01/01/2007 and 12/31/2020 with ≥1 serum creatinine test, an eGFR between 15ml/min/1.73m 2 and 180ml/min/1.73m 2 , and ≥7 of the 19 other laboratory analytes available. Two primary sub-cohorts were evaluated within each insurer: (1) all patients with ≥7 laboratory analytes; and (2) patients in (1) with available urinalysis results. Disease progression was defined as the composite outcome of a sustained 40% decline in eGFR or kidney failure. Discrimination, accuracy, and calibration were assessed using the area under the receiver operator characteristic curve (AUC), Brier scores, and calibration plots. RESULTS In the commercial cohort, the Klinrisk model achieved AUCs ranging from 0.83 (95% confidence interval: 0.82 - 0.83) to 0.87 (0.86 - 0.87) and a maximum Brier score of 0.005 (0.0005 - 0.005) at 2 years. In Medicare patients, AUCs ranged from 0.80 (0.79 - 0.80) to 0.81 (0.80 - 0.82), with a maximum Brier score of 0.026 (0.025 - 0.027). In Medicaid patients, we found AUCs ranging from 0.82 (0.82 - 0.82) to 0.84 (0.82 - 0.86) and a maximum Brier score of 0.014 (0.012 - 0.015). CONCLUSIONS The Klinrisk machine learning model was accurate in predicting CKD progression in 4.8 million US adults across commercial, Medicare, and Medicaid populations.