TO020POTENTIAL INTERPRETATIONS OF CRITERIA FOR AKI BY AUTOMATED DECISION SUPPORT ALGORITHMS: IMPACT ON AKI INCIDENCE

Background and Aims The reported incidence of Acute Kidney Injury (AKI) at the intensive care unit (ICU) is variable. Although the Kidney Disease Improving Global Outcome (K-DIGO) improved harmonisation of this definition, there is remaining variability in the actual implementation of this AKI definition, with variable interpretation of the urinary output (UO) criterion, and of the baseline serum creatinine (Screa) criterion. This hampers progress of our understanding of the clinical concept AKI and leads to confusion and unclarity when interpreting models to predict AKI or associated outcomes. With the advent of big data and artificial intelligence based decision algorithms, this problem will only become more of interest, as the user will not know what exactly the construct AKI in the application used means and represents. Therefore, we intended to explore the impact of different interpretations of the Screa and the UO criterium as presented in the K-DIGO definition on the incidence of AKI stage 2. Method We included all patients of an electronic health data system applied in a tertiary ICU between 2013 and 2017. Sequential Organ Failure Assessment (SOFA) score was calculated, and gender, age, weight and mortality at ICU and in hospital were extracted. All serum creatinine (sCrea) values during ICU stay and hospitalisation were extracted, as were UO data, with their time stamps. In addition, all available Screa data up to 1 year before ICU admission were retrieved from a dataset external to the ICU. AKI was defined according to KDIGO stage 2, using different possible interpretations of the Screa and/or the UO criterion. For the evolution of Screa as compared to a baseline value, we sued either a value directly available to ICU staff (def 1), a presumed eGFR of 75ml/min (def 2), the first available value after admission to ICU (def 3), the lowest value during the current hospitalisation before ICU admission (def 4), the lowest value before the hospitalisation episode as found in an external dataset (def 5). For the UO criterion, we also applied two criteria in line with K-DIGO stage 2: a UO below 6ml/kg during a 12 hour block (def 6) or a UO below 0.5ml/kg/hour during each of 12 consecutive one hour intervals (def 7). Def 8 identified patients who did not comply with any of the definitions (1-7), so who had no AKI according to any definition. Definition 9 and 10 identified patients who complied with at least one out of the Screa criteria 1-5 (def 9) or out of the UO criteria (def 10). Definition 11 identified patients who complied both with at least one Screa and one UO criterium. Results Our dataset included 16433 ICU admissions (34.7% female, age 60.7±16.4 years). Overall, 8.1% of patients died at ICU, and another 5.2% during their hospitalisation. The SOFA score at admission was 6.9±4.1. The incidence of AKI according to the stage 2 definition of K-DIGO varied according to the interpretation of the diagnostic criteria from 4.3% when baseline creatinine was defined as the first ICU value, to 35.3% when the UO criterium was interpreted as a UO below 6ml/kg over a 12 hour block (fig). Only half of patients (53.7%) did not comply with any of the definitions (def 8), 10.9% and 19.7% complied with one of the Screa (def 9) OR one of the UO criteria (def 10) respectively, and 15.7% complied with both (def 11). There was substantial reclassification across the different definitions. Conclusion Unclarity on the actual interpretation of the Screa and UO criteria used in the K-DIGO definition of AKI leads to substantial differences in incidence of AKI, and also with substantial reclassification according to different definitions. This is especially concerning in an era of big data and automated decision support, as clinicians might not know which construct of AKI is actually being represented.