POS-061 THE IMPACT OF NEGLECTING URINARY OUTPUT WHEN DEFINING AKI IN PREDICTION MODELLING: A BIG DATA ANALYSIS

Abstract

KDIGO criteria have improved standardization of AKI definition, but most published automated algorithms fail to include the urinary output (UO) criterium in their ground truth definition of AKI. UO is also associated with mortality, and ignoring this parameter in the AKI definition might thus not accurately reflect the risk associated with the label AKI. We classified KDIGO AKI stage≥2 using routinely collected ICU data to identify the impact of including UO or not in an automated AKI algorithm on estimates of AKI incidence and prognostic value. All patients admitted to an intensive care unit (ICU) in a tertiary care centre between 1/1/2013 and 31/12/2017 were prospectively included. We used different interpretations for the baseline of the serum creatinine (Screa) criterion, and/or the urinary output (UO) criterion (either a moving 12h time window or in each of its 12 constituent 1h blocks) to identify patients fulfilling KDIGO AKI stage≥2 diagnosis. As outcome, we assessed incidence of AKI and association with ICU mortality using state of the art statistics to avoid immortality bias. In 13,403 ICU admissions (62.2% male, age 60.8±16.8 years, SOFA at admission 7.0±4.1), incidence of KDIGO AKI stage≥2 varied from 4.3% to 19.3% for the different baseline definitions of the Screa criterion, and 14.1 to 34.3% for the two interpretations of the UO criteria (fig dashed columns). Cause-specific hazard ratios of ICU mortality for patients with versus without AKI diagnosis also ranged from 1.76 (95% CI: 1.47–2.10) to 2.38 (95% CI: 2.09–2.71) for the Screa criteria and from 3.00 (95%CI: 2.55-3.53) to 3.21 (95%CI: 2.80-3.69) for the two interpretations of the UO criteria (Fig plain columns). Neglecting urinary output when defining AKI according to KDIGO negatively impacts the sensitivity and prognostic value of the label AKI. Despite its higher incidence, AKI based on UO had a higher ICU based mortality risk than AKI based on any of the Screa criteria of KDIGO. Algorithms to detect or predict AKI should therefore include urinary output in their ground truth definition of AKI.