(689) Prediction of Short-Term Creatinine Change in Critically Ill Children

Introduction: Acute kidney injury (AKI) affects 25-30% of children in pediatric intensive care units (PICUs). AKI is often multifactorial with disease-related and iatrogenic processes. Previous efforts to predict AKI have used empirically selected variables and had limited success. No studies have attempted to predict short-term creatinine (Cr) change. We aimed to predict 12-hour change in Cr in PICU children.

Methods: Patient encounters with the following criteria were extracted from Real-World Data (RWD), a large multi-center database: age 1 month-18 years, PICU or ICU location, PICU stay 6 hours or more, and at least 2 Cr measurements and 1 urine output (UOP) record. ICU level of care was confirmed with at least 2 measurements of heart rate, respiratory rate, blood pressure, and oxygen saturation, and at least 1 temperature measurement. Patients with pre-existing chronic kidney disease, renal anomalies, renal transplant, or dialysis dependence were excluded. Data were extracted in the following categories: demographics, admission data, vital signs, laboratory tests, and medication administration. A lasso regression model was used to predict 12-hour change in Cr based on most recent available data. The model was developed on a randomly selected 80% of the cohort, then tested on remaining 20%.

Results: 39,932 encounters met inclusion criteria. 53.8% (n=21,500) were male. Median age was 8.0 years (IQR 1.9-14.6 years). Mortality was 2.3% (n=813). 21.0% (n=8,393) had AKI by Cr or UOP. Among all 12-hour intervals, 3.1% had a Cr increase of 50% or more, equating to 11.6% of patients (n=4640). Model-predicted Cr values had a mean absolute error (predicted to actual) of 16.0% (95% CI 0.0-63.4%). The variables with the largest contribution to prediction of Cr change were heart rate, blood pressure, invasive mechanical ventilation (tidal volume and peak inspiratory pressure), use of vasoactives, and laboratory values including C-reactive protein, pH, INR, and vancomycin level.

Conclusions: Creatinine change in critically ill children can be predicted using routinely collected clinical, laboratory, medication administration, and medication level variables. Predicting future Cr could aid in early identification of renal dysfunction and may allow clinicians to alter treatments and avoid nephrotoxic therapies.