External Validation of the Simple Postoperative Acute Kidney Injury Risk Index in Patients Admitted to the Intensive Care Unit After Noncardiac Surgery.

Abstract

Background: The Simple Postoperative AKI Risk (SPARK) index is a novel model for predicting risk of postoperative acute kidney injury (PO-AKI) among patients after noncardiac surgery. However, the performance of the index has been inconsistent partly due to heterogeneity in case mix and effects of the involved clinical features. To clarify potential reasons for poor performance, we tested the SPARK index in a cohort of high-risk patients requiring intensive care unit (ICU) care after noncardiac surgery and examined whether model modification by refitting coefficients of clinical features could optimize model performance.

Methods: This was a single-center prospective cohort study. Preoperative variables of the SPARK index were extracted from electronic medical records. PO-AKI was defined by an increase in sCr ≥26.5 mmol/L within 48 hours or 150% compared with the preoperative baseline value within 7 days after surgery, whereas critical AKI was defined as AKI stage 2 or greater and/or any AKI connected to postoperative death or requiring renal replacement therapy during the hospital stay. Discrimination was evaluated by the area under the receiver operating characteristic curve (AUC), and calibration was evaluated by the Hosmer-Lemeshow χ 2 test and calibration plot. Model modification was performed by rebuilding the model with the original variables of the SPARK index through proportional odds logistic regression among participants in the earlier study period and was validated in the later one.

Results: A total of 973 patients were enrolled, among whom 79 (8.1%) PO-AKI cases and 14 (1.4%) critical AKI cases occurred. Our study participants demonstrated a higher SPARK risk score than the SPARK discovery cohort (eg, 8.02% vs 1.20% allocated in the highest risk group), and the incidence of both outcomes increased through the classes of the score (incidence proportion of PO-AKI increased from 2.56% in the lowest risk group to 25.64% in the highest risk group). The AUCs for PO-AKI and critical AKI were 0.703 (95% confidence interval [CI], 0.641-0.765) and 0.699 (95% CI, 0.550-0.848), respectively. The sensitivity, specificity, negative predictive value and positive predictive value were 68.35%, 57.49%, 95.36%, and 12.44%, respectively, when using 10% of predicted probability of PO-AKI as threshold. Calibration plots suggested acceptable consistency between the predicted and actual risk. After model modification, external validation demonstrated a significantly improved AUC for PO-AKI.

Conclusions: The SPARK index showed fair discrimination and calibration among patients admitted to the ICU after noncardiac surgery. Modification of the model improved the performance of the model in terms of predicting PO-AKI.