(693) Predicting Major Adverse Kidney Events in Pediatric Patients on Continuous Renal Replacement Therapy

Introduction: More than half of children who survive continuous renal replacement therapy (CRRT) have persistent renal dysfunction. Large-scale, real-world data paired with machine learning (ML) methods is a promising approach to develop practical prediction models to aid clinicians in improving patient outcomes. We utilized ML algorithms to model prediction of major adverse kidney events at 30 days (MAKE-30) in critically ill pediatric patients.

Methods: Data from the first 72 hours of CRRT from patients < 26 years who were admitted to an intensive care unit at two, quaternary pediatric centers. The outcome was MAKE-30 (estimated glomerular filtration rate (eGFR) decrease by >25% from baseline, need for dialysis, and mortality at 30 days). Features with >10% missingness were dropped and patients with end stage renal disease were excluded. Complete case analysis was then used to omit any patients with missing values. The data were split into 75-25 training and testing. We evaluated multiple well-known machine learning algorithms to investigate the most optimal algorithm. Variable importance was evaluated using the Shapley Additive Explanations (SHAP) plot and the final model was chosen based on the test set AUC value.

Results: 388 patients were included in the analysis. The median (IQR) age was 96 (18.9-180) months with a BSA of 0.92 (0.51-1.50) m2. The most common indication for ICU admission and co-morbid condition was respiratory 88 (23%) and hematological/oncological 80 (21%). The median duration of CRRT was 10 (4-21) days. MAKE-30 was observed in 292 (75%) patients. Our final, best-performing model was the random forest model. The five most important features within this model were urine output (UOP) on day 2 of CRRT in ml/kg/hr, UOP on day 1 of CRRT in ml, UOP day 1 of CRRT in ml/kg, AKI as the indication for CRRT, and location of the dialysis catheter. Our model had an AUC of 0.8995 with the final test set.

Conclusions: Despite the modest cohort size and leveraging clinically collected data, ML algorithms can be used to predict MAKE-30 with excellent discrimination in critically ill pediatric patients requiring CRRT. Addition of more granular clinical data to improve the prediction of patient-centered outcomes within this critically ill population could potentially uncover modifiable risk factors.