Derivation of a prediction model for emergency department acute kidney injury

Abstract

BACKGROUND AND OBJECTIVE: Quality management of Acute Kidney Injury (AKI) is dependent on early detection, which is currently deemed to be suboptimal. The aim of this study was to identify combinations of variables associated with AKI and to derive a prediction tool for detecting patients attending the emergency department (ED) or hospital with AKI (ED-AKI). DESIGN, SETTING, PARTICIPANTS AND MEASUREMENTS: This retrospective observational study was conducted in the ED of a tertiary university hospital in Wales. Between April and August 2016 20,421 adult patients attended the ED of a University Hospital in Wales and had a serum creatinine measurement. Using an electronic AKI reporting system, 548 incident adult ED-AKI patients were identified and compared to a randomly selected cohort of adult non-AKI ED patients (n = 571). A prediction model for AKI was derived and subsequently internally validated using bootstrapping. The primary outcome measure was the number of patients with ED-AKI. RESULTS: In 1119 subjects, 27 variables were evaluated. Four ED-AKI models were generated with C-statistics ranging from 0.800 to 0.765. The simplest and most practical multivariate model (model 3) included eight variables that could all be assessed at ED arrival. A 31-point score was derived where 0 is minimal risk of ED-AKI. The model discrimination was adequate (C-statistic 0.793) and calibration was good (Hosmer & Lomeshow test 27.4). ED-AKI could be ruled out with a score of <2.5 (sensitivity 95%). Internal validation using bootstrapping yielded an optimal Youden index of 0.49 with sensitivity of 80% and specificity of 68%. CONCLUSION: A risk-stratification model for ED-AKI has been derived and internally validated. The discrimination of this model is objective and adequate. It requires refinement and external validation in more generalisable settings.