(1324) Development and Validation of a Prediction Model for Stages of Acute Kidney Injury in ICU Patients

Introduction: Acute kidney injury (AKI) is a global concern with high incidence and poor prognosis among critically ill patients. As AKI is frequently only detected well after its onset, early risk stratification is crucial. We aimed to develop and internally validate the first clinical prediction model for different stages of AKI occurring within seven days after ICU admission.

Methods: We used data from the Simple Intensive Care Studies II (SICS-II), a prospective cohort study including 1010 critically ill adults at the University Medical Center Groningen, the Netherlands. The prognostic outcome was the highest KDIGO-based stage of AKI occurring within the first seven days of ICU stay. Candidate predictors included 59 readily available variables in critical care. Least absolute shrinkage and selection operator (LASSO) and proportional odds logistic regression were used for variable selection and model estimation, respectively. Receiver operating characteristic (ROC) curve, calibration plot, and decision curve analysis were applied to evaluate model performance and clinical usefulness. We internally validated the model using bootstrapping.

Results: Of the SICS-II cohort, 976 patients were eligible for our analyses (median [IQR] 64 [52-72] years, 38% female). Within seven days after ICU admission, 283 (29%), 228 (23%), and 135 (14%) patients progressed to their highest severity of AKI at stages 1, 2, and 3, respectively. We derived a 15-variable model for predicting this maximum ordinal outcome with an area under the ROC curve of 0.763 (95% confidence interval, 0.740-0.784) in the internal validation. The model showed good calibration and improved net benefit in decision curve analysis over a range of clinically plausible thresholds.

Conclusions: Using readily available predictors in the ICU setting, we could develop a prediction model for different stages of AKI with good performance and promising clinical usefulness in decision curve analysis. Our findings serve as an initial step to applying a valid and timely prediction model of multiple AKI stages, possibly helping to limit morbidity and improve patient outcomes.