O3015: Predicting Acute Kidney Injury in Pediatric Cardiac Surgery Patients Using Supervised Machine Learning

Background: Acute kidney injury is highly prevalent after cardiac surgery (CSA-AKI) in children of all ages, specifically for newborns and young infants, with the reported incidence being between 15% to 60%. A key challenge in critical care medicine has been accurately predicting which patients will develop CSA-AKI in the subacute postoperative period. The aim of this project was to build a supervised machine learning (ML) model to predict CSA-AKI at postoperative day (POD) 2 in patients < 21-years-old using only pre-, intra-, and immediately post-op data.

Methods: Patients younger than 21 years old who underwent cardiac surgery at the West Virginia University Children's Hospital between September 1, 2007, and June 31, 2013, were included in this single-center, retrospective study. The ML model, Light GBM (gradient boosted machine), was trained on 24 input features consisting of demographic, preoperative, intraoperative and immediately post-operative clinical and laboratory data. The model was trained to predict whether patients would develop moderate to severe AKI categorized by Kidney Disease: Improving Global Outcomes (KDIGO) stage 2 or 3 at POD2. Input feature importance was assessed by SHapley Additive exPlanations (SHAP) values. ML model performance was assessed for accuracy, area under the receiver operating curve (ROC AUC), precision, recall, area under the precision recall curve (PR AUC), F1-score, and brier score.

Results: Data from 399 patients were analyzed (6.0 [2.0,27.0] months, 44% female). Overall, 13.5% of patients experienced AKI of KDIGO stage 2 or 3 at POD2. The supervised ML model achieved an accuracy of 0.90, ROC AUC of 0.82, precision of 1.00, recall of 0.38, PR AUC of 0.60, F1-score of 0.56, and brier score of 0.10. The top ten most important features assessed by SHAP analyses in this model, in the order of importance were preoperative serum creatinine, surgery duration, POD0 serum pH, POD0 lactate, POD0 hematocrit, urine output in first 8 hours post-surgery, POD0 central venous pressure, cardiac bypass minutes, pre-operative weight, age, POD0 serum creatinine (Figure 1).

Conclusion: A supervised ML model trained only on demographic, preoperative, intraoperative and immediate postoperative clinical and laboratory data was able to achieve promising performance in predicting moderate to severe CSA-AKI at POD2. Next steps include increasing the training data size and further validating the model on external datasets. As the ability to predict CSA-AKI development increases, clinicians will be afforded an increased array of tools for aiding critical care clinical decision making for pediatric cardiac surgery patients.